feat(voice): first-class voice input and local TTS across web, desktop, and mobile (#2018)

Complete rebuild of voice input on a server-authoritative streaming
architecture, replacing the legacy Web Speech / whole-blob / WASM engines
and the dead voice-agent layer (~4k lines removed).

Speech-to-text (dictation):
- Client streams 16 kHz mono PCM16 chunks over /api/dictation/ws with
  seq/ack ordering; buffered audio is retained and replayed on reconnect
- Server transcribes and streams live partial transcripts back;
  segments auto-commit every ~15s with silence suppression and adaptive
  finalization timeouts
- Local provider (default, zero config): sherpa-onnx models in a forked
  worker process — auto-download with progress, staged extraction with
  verification, corrupt-model auto-recovery, idle shutdown after 5 min
- Model catalog with settings picker (accuracy/speed ratings, sizes,
  download/delete): Parakeet TDT v2 (English) and v3 (25 European
  languages, auto-detected), Whisper base and tiny (multilingual, light)
- OpenAI-compatible provider for any Whisper endpoint
- Composer overlay with live transcript, volume meter, timer, and
  cancel / insert / insert-and-send actions; failed transcriptions keep
  their audio for retry or accepting the partial text as-is
- Configurable keyboard shortcut (default mod+alt+v) toggles dictation;
  Enter confirms and Escape cancels while recording
- Overlay is pixel-aligned with the composer (measured footer height,
  matching paddings/typography/gaps) — no layout shift when toggling

Text-to-speech:
- Local Kokoro provider (English, 11 voices) synthesized in the same
  worker via /api/dictation/tts/speak, managed by the shared model
  pipeline; sentence-pipelined playback keeps time-to-first-audio at
  ~1 sentence regardless of message length, and stop cancels in-flight
  synthesis
- Sanitizer keeps inline-code content (strips backticks only), reads
  interword slashes aloud, and removes only absolute file paths

Settings:
- Voice page unified: a single read-aloud toggle owns all playback
  options (the confusing "Enable Voice Mode" is gone); a new "Enable
  voice input" toggle (default on, persisted to settings.json) hides
  the composer mic entirely when disabled

Mobile and transport:
- iOS/Android microphone permissions added (dictation was previously
  impossible on mobile)
- Fixed Android WebSocket upgrades: the Capacitor WebView origin
  (https://localhost) was missing from the packaged-client allowlist,
  403-ing every WS connection — root cause of the old mobile SSE lock,
  which is now removed for all transports

Security and conventions:
- All HTTP routes sit behind the global /api auth gate; the WS upgrade
  explicitly validates the UI session and origin, with oc_url_token
  narrowly allowlisted and covered by tests; the dictation socket mints
  a fresh URL token before connecting
- Routes register before the generic OpenCode proxy; the client goes
  through runtimeFetch/getRuntimeUrlResolver, and runtime switches
  reset the dictation socket
- VS Code deliberately reports dictation as unavailable (no server
  process in that runtime)

CI: workflow Node bumped 20 -> 22 to match the repo engines and fix
better-sqlite3 installs broken by node-gyp@latest on Node 20.

New dependency: sherpa-onnx-node (prebuilt N-API; macOS/Linux x64+arm64,
Windows x64 — Windows-on-ARM falls back to the OpenAI-compatible provider)

.github/workflows/build-macos-arm64-dmg.yml

@@ -32,7 +32,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: "20"
+          node-version: "22"
 
       - name: Install dependencies
         run: bun install --frozen-lockfile

.github/workflows/oc-review.yml

@@ -18,7 +18,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Install dependencies
         run: bun install --frozen-lockfile

.github/workflows/release-desktop-smoke.yml

@@ -65,7 +65,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Install dependencies
         run: bun install --frozen-lockfile
@@ -206,7 +206,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Install dependencies
         run: bun install --frozen-lockfile

.github/workflows/release.yml

@@ -90,7 +90,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
           registry-url: 'https://registry.npmjs.org'
 
       - name: Install dependencies
@@ -141,7 +141,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Install dependencies
         run: bun install --frozen-lockfile
@@ -288,7 +288,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Install dependencies
         run: bun install --frozen-lockfile
@@ -365,7 +365,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Download per-arch latest-mac.yml
         uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0

.github/workflows/vscode-extension.yml

@@ -25,7 +25,7 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: '20'
+          node-version: '22'
 
       - name: Install dependencies
         run: bun install --frozen-lockfile

.gitignore

@@ -67,3 +67,4 @@ data/
 workspaces/
 *.pid
 .worktrees/
+test-results/

bun.lock

@@ -283,6 +283,7 @@
         "openai": "^4.79.0",
         "qrcode-terminal": "^0.12.0",
         "reflect-metadata": "^0.2.2",
+        "sherpa-onnx-node": "1.12.28",
         "simple-git": "^3.28.0",
         "web-push": "^3.6.7",
         "ws": "^8.18.3",
@@ -2980,6 +2981,20 @@
 
     "shell-quote": ["shell-quote@1.8.3", "", {}, "sha512-ObmnIF4hXNg1BqhnHmgbDETF8dLPCggZWBjkQfhZpbszZnYur5DUljTcCHii5LC3J5E0yeO/1LIMyH+UvHQgyw=="],
 
+    "sherpa-onnx-darwin-arm64": ["sherpa-onnx-darwin-arm64@1.13.3", "", { "os": "darwin", "cpu": "arm64" }, "sha512-9x86Cbf+BDFONdtCPM3cnjvtAW0ER8tMaHK5pVfz+SHPt8GeuwRXaiR/BzcByFBUyxCgmceO09/WMZOCi44P/g=="],
+
+    "sherpa-onnx-darwin-x64": ["sherpa-onnx-darwin-x64@1.13.3", "", { "os": "darwin", "cpu": "x64" }, "sha512-TVQ35g7JIpDPB1lUDdcog+JtI0cI45ZzOnvHXm0DtWs/dgxnJXtWMY3uLRtBbLnysV9j5ljffwZ1IX9VDHsCzQ=="],
+
+    "sherpa-onnx-linux-arm64": ["sherpa-onnx-linux-arm64@1.13.3", "", { "os": "linux", "cpu": "arm64" }, "sha512-uDtZkkoP6QQ/3DHOscCpEZ2WpaiHUQsDpbyYaHURrJ7DbsjqGnS6G8l+R589Ro5Bf282QElzBy3okwxXbt3Kxw=="],
+
+    "sherpa-onnx-linux-x64": ["sherpa-onnx-linux-x64@1.13.3", "", { "os": "linux", "cpu": "x64" }, "sha512-OFVK0GYwKwKNsjxbPmfcLQm/dfA0IwAoiIQJ96s+eFYcDqhlapcY06ocdb7SNluGBcM7xgU5jEW2QXBkMIOEvQ=="],
+
+    "sherpa-onnx-node": ["sherpa-onnx-node@1.12.28", "", { "optionalDependencies": { "sherpa-onnx-darwin-arm64": "^1.12.28", "sherpa-onnx-darwin-x64": "^1.12.28", "sherpa-onnx-linux-arm64": "^1.12.28", "sherpa-onnx-linux-x64": "^1.12.28", "sherpa-onnx-win-ia32": "^1.12.28", "sherpa-onnx-win-x64": "^1.12.28" } }, "sha512-EHSB3EG6hKyXaTNh6GU/bwh6i3dncCH6ZCU2mScNzkxRbVStZ7QmNj0Oo4E9XrGGo9jX9pKg9MPHEPyjdK+ApA=="],
+
+    "sherpa-onnx-win-ia32": ["sherpa-onnx-win-ia32@1.13.3", "", { "os": "win32", "cpu": "ia32" }, "sha512-VDZh1M7Ccx/bkP3WwBCFoJzwAwq+b5nR1KRkYRz5p1w5bfhzfa3ACBGr7vpUt5AGUge4qSLe0MSKXyKtSmy1uA=="],
+
+    "sherpa-onnx-win-x64": ["sherpa-onnx-win-x64@1.13.3", "", { "os": "win32", "cpu": "x64" }, "sha512-ZQzcSmFvZK4jzmtWckqxocDUuEjYnBV2MHrDD21HPTeUMfGdE9yfvuSPpesIVfdzKbQzIQY42RAcfZEGWu0FbQ=="],
+
     "shiki": ["shiki@3.23.0", "", { "dependencies": { "@shikijs/core": "3.23.0", "@shikijs/engine-javascript": "3.23.0", "@shikijs/engine-oniguruma": "3.23.0", "@shikijs/langs": "3.23.0", "@shikijs/themes": "3.23.0", "@shikijs/types": "3.23.0", "@shikijs/vscode-textmate": "^10.0.2", "@types/hast": "^3.0.4" } }, "sha512-55Dj73uq9ZXL5zyeRPzHQsK7Nbyt6Y10k5s7OjuFZGMhpp4r/rsLBH0o/0fstIzX1Lep9VxefWljK/SKCzygIA=="],
 
     "side-channel": ["side-channel@1.1.0", "", { "dependencies": { "es-errors": "^1.3.0", "object-inspect": "^1.13.3", "side-channel-list": "^1.0.0", "side-channel-map": "^1.0.1", "side-channel-weakmap": "^1.0.2" } }, "sha512-ZX99e6tRweoUXqR+VBrslhda51Nh5MTQwou5tnUDgbtyM0dBgmhEDtWGP/xbKn6hqfPRHujUNwz5fy/wbbhnpw=="],

packages/mobile/android/app/src/main/AndroidManifest.xml

@@ -53,4 +53,10 @@
     <!-- Android 13+ requires this declared for FCM push notifications to be shown; the runtime
          prompt is requested by @capacitor/push-notifications on register. -->
     <uses-permission android:name="android.permission.POST_NOTIFICATIONS" />
+    <!-- Microphone: voice dictation via WebView getUserMedia. Capacitor's
+         BridgeWebChromeClient forwards the WebView AUDIO_CAPTURE permission
+         request to the Android runtime prompt when this is declared. -->
+    <uses-permission android:name="android.permission.RECORD_AUDIO" />
+    <uses-permission android:name="android.permission.MODIFY_AUDIO_SETTINGS" />
+    <uses-feature android:name="android.hardware.microphone" android:required="false" />
 </manifest>

packages/mobile/ios/App/App/Info.plist

@@ -35,6 +35,8 @@
 	<string>OpenChamber connects to OpenChamber servers on your local network.</string>
 	<key>NSCameraUsageDescription</key>
 	<string>OpenChamber uses the camera to scan a server's pairing QR code.</string>
+	<key>NSMicrophoneUsageDescription</key>
+	<string>OpenChamber uses the microphone for voice dictation in the chat composer.</string>
 	<key>CFBundleURLTypes</key>
 	<array>
 		<dict>

packages/ui/src/App.tsx

@@ -44,7 +44,6 @@ import { ConfigUpdateOverlay } from '@/components/ui/ConfigUpdateOverlay';
 import { AboutDialog } from '@/components/ui/AboutDialog';
 import { RuntimeAPIProvider } from '@/contexts/RuntimeAPIProvider';
 import { registerRuntimeAPIs } from '@/contexts/runtimeAPIRegistry';
-import { VoiceProvider } from '@/components/voice';
 import { useUIStore } from '@/stores/useUIStore';
 import { useGitHubAuthStore } from '@/stores/useGitHubAuthStore';
 import { useFeatureFlagsStore } from '@/stores/useFeatureFlagsStore';
@@ -928,8 +927,8 @@ function App({ apis }: AppProps) {
   }
 
   // Always mount the full provider tree to avoid remounts when isInitialized
-  // flips from false → true. FireworksProvider and VoiceProvider are lightweight
+  // flips from false → true. FireworksProvider is a lightweight shell; its
-  // shells; their heavy children are only activated when actually needed.
+  // heavy children are only activated when actually needed.
   const isBootShell = !isInitialized && !isDesktopRuntime;
 
   return (
@@ -937,7 +936,6 @@ function App({ apis }: AppProps) {
       <SyncProvider key={runtimeEndpointEpoch} sdk={opencodeClient.getSdkClient()} directory={currentDirectory || ''}>
         <RuntimeAPIProvider apis={apis}>
           <FireworksProvider>
-            <VoiceProvider>
               <TooltipProvider delayDuration={300} skipDelayDuration={150}>
                 <div className={isDesktopRuntime ? 'h-full text-foreground bg-transparent' : 'h-full text-foreground bg-background'}>
                   <SyncAppEffects embeddedBackgroundWorkEnabled={embeddedBackgroundWorkEnabled} />
@@ -955,7 +953,6 @@ function App({ apis }: AppProps) {
                   )}
                 </div>
               </TooltipProvider>
-            </VoiceProvider>
           </FireworksProvider>
         </RuntimeAPIProvider>
       </SyncProvider>

packages/ui/src/components/chat/ChatInput.tsx

@@ -1,6 +1,6 @@
 import React from 'react';
 import { Textarea } from '@/components/ui/textarea';
-import { BrowserVoiceButton } from '@/components/voice';
+import { ComposerDictation } from '@/components/dictation/ComposerDictation';
 // sessionStore removed — currentSessionId comes from useSessionUIStore
 import { useConfigStore } from '@/stores/useConfigStore';
 import { useUIStore } from '@/stores/useUIStore';
@@ -381,7 +381,7 @@ const getProjectIconColor = (projectColor?: string | null): string | undefined =
 };
 
 const MemoModelControls = React.memo(ModelControls);
-const MemoBrowserVoiceButton = React.memo(BrowserVoiceButton);
+const MemoComposerDictation = React.memo(ComposerDictation);
 const MemoMobileAgentButton = React.memo(MobileAgentButton);
 const MemoMobileModelButton = React.memo(MobileModelButton);
 const MemoStatusRow = React.memo(StatusRow);
@@ -2318,6 +2318,33 @@ const ChatInputComponent: React.FC<ChatInputProps> = ({ onOpenSettings, scrollTo
         void handleSubmitRef.current();
     }, []);
 
+    // Dictation: insert the transcript inline; optionally submit immediately.
+    // getCurrentInputSnapshot reads textareaRef.current.value first, so setting
+    // it synchronously lets handleSubmit pick up the text in the same tick.
+    const handleDictationInsert = React.useCallback((text: string) => {
+        setMessage((prev) => {
+            const next = appendInlineText(prev, text);
+            const textarea = textareaRef.current;
+            if (textarea) {
+                textarea.value = next;
+            }
+            return next;
+        });
+        setTimeout(() => {
+            textareaRef.current?.focus();
+        }, 0);
+    }, []);
+
+    const handleDictationInsertAndSend = React.useCallback((text: string) => {
+        const textarea = textareaRef.current;
+        const next = appendInlineText(textarea?.value ?? messageRef.current, text);
+        if (textarea) {
+            textarea.value = next;
+        }
+        setMessage(next);
+        void handleSubmitRef.current();
+    }, []);
+
     // Preset chips rendered outside this component (e.g. under the welcome
     // message on narrow surfaces) request a submit via the input store; consume
     // it here so it routes through the same command-aware submit path.
@@ -4420,6 +4447,9 @@ const ChatInputComponent: React.FC<ChatInputProps> = ({ onOpenSettings, scrollTo
                                 : undefined}
                         />
                     )}
+                    {/* Positioning context for the dictation overlay: covers the
+                        text area + footer exactly, excluding MobileSessionStatusBar. */}
+                    <div className={cn('relative flex flex-col', isDesktopExpanded && 'flex-1 min-h-0')}>
                     <div className={cn("overflow-hidden", isDesktopExpanded && 'flex flex-1 min-h-0 flex-col')}>
                         <div className="flex items-center gap-1 px-3 pt-1 flex-wrap relative z-10">
                             <AttachedVSCodeFileChips onShowPopup={handleShowAttachmentPreview} />
@@ -4559,7 +4589,16 @@ const ChatInputComponent: React.FC<ChatInputProps> = ({ onOpenSettings, scrollTo
                                             />
                                         </div>
                                         <div className="flex items-center gap-x-1 flex-shrink-0">
-                                            <MemoBrowserVoiceButton />
+                                            <MemoComposerDictation
+                                                radius={chatInputRadius}
+                                                isMobile={isMobile}
+                                                footerIconButtonClass={footerIconButtonClass}
+                                                footerPaddingClass={footerPaddingClass}
+                                                iconSizeClass={iconSizeClass}
+                                                sendIconSizeClass={sendIconSizeClass}
+                                                onInsert={handleDictationInsert}
+                                                onInsertAndSend={handleDictationInsertAndSend}
+                                            />
                                             <ComposerActionButtons
                                                 isMobile={isMobile}
                                                 footerIconButtonClass={footerIconButtonClass}
@@ -4614,7 +4653,16 @@ const ChatInputComponent: React.FC<ChatInputProps> = ({ onOpenSettings, scrollTo
                                 </div>
                                 <div className={cn('flex items-center flex-1 justify-end', footerGapClass, 'md:gap-x-3')}>
                                     <MemoModelControls className={cn('flex-1 min-w-0 justify-end')} />
-                                    <MemoBrowserVoiceButton />
+                                    <MemoComposerDictation
+                                        radius={chatInputRadius}
+                                        isMobile={isMobile}
+                                        footerIconButtonClass={footerIconButtonClass}
+                                        footerPaddingClass={footerPaddingClass}
+                                        iconSizeClass={iconSizeClass}
+                                        sendIconSizeClass={sendIconSizeClass}
+                                        onInsert={handleDictationInsert}
+                                        onInsertAndSend={handleDictationInsertAndSend}
+                                    />
                                     <ComposerActionButtons
                                         isMobile={isMobile}
                                         footerIconButtonClass={footerIconButtonClass}
@@ -4633,6 +4681,7 @@ const ChatInputComponent: React.FC<ChatInputProps> = ({ onOpenSettings, scrollTo
                             </>
                         )}
                     </div>
+                    </div>
 
                     {/* Mobile session panel: slide-up overlay toggled by MobileSessionPanelTrigger. */}
                     {isMobile && <MobileSessionStatusBar />}

packages/ui/src/components/dictation/ComposerDictation.tsx

@@ -0,0 +1,439 @@
+/**
+ * Composer dictation controls: a mic button for the composer footer plus a
+ * full-composer overlay while dictation is active (recording, transcribing,
+ * or failed). The overlay mirrors the composer's own layout — the transcript
+ * area uses the same paddings/typography as the textarea and the action row
+ * reuses the footer icon-button styling — so toggling dictation causes no
+ * vertical shift.
+ */
+
+import React from 'react';
+
+import { Icon } from '@/components/icon/Icon';
+import { useI18n } from '@/lib/i18n';
+import { useThemeSystem } from '@/contexts/useThemeSystem';
+import { cn } from '@/lib/utils';
+import { runtimeFetch } from '@/lib/runtime-fetch';
+import { useDictation } from '@/hooks/useDictation';
+import { isDictationCaptureSupported } from '@/lib/dictation/use-dictation-audio-source';
+import { isVSCodeRuntime } from '@/lib/desktop';
+import { useConfigStore } from '@/stores/useConfigStore';
+import { useUIStore } from '@/stores/useUIStore';
+import { formatShortcutForDisplay, getEffectiveShortcutCombo } from '@/lib/shortcuts';
+
+interface ComposerDictationProps {
+    radius?: number | string;
+    isMobile: boolean;
+    footerIconButtonClass: string;
+    footerPaddingClass: string;
+    iconSizeClass: string;
+    sendIconSizeClass: string;
+    disabled?: boolean;
+    onInsert: (text: string) => void;
+    onInsertAndSend: (text: string) => void;
+}
+
+const formatDuration = (seconds: number): string => {
+    const mins = Math.floor(seconds / 60);
+    const secs = seconds % 60;
+    return `${mins}:${String(secs).padStart(2, '0')}`;
+};
+
+const VolumeMeter: React.FC<{ volume: number }> = ({ volume }) => {
+    const { currentTheme } = useThemeSystem();
+    return (
+        <div
+            className="h-1.5 w-16 flex-shrink-0 overflow-hidden rounded-full"
+            style={{ backgroundColor: currentTheme.colors.interactive.border }}
+            aria-hidden="true"
+        >
+            <div
+                className="h-full rounded-full transition-[width] duration-75"
+                style={{
+                    width: `${Math.round(Math.min(1, volume) * 100)}%`,
+                    backgroundColor: currentTheme.colors.primary.base,
+                }}
+            />
+        </div>
+    );
+};
+
+/**
+ * Polls the dictation status route while the local model is downloading and
+ * returns the download percent (null while unknown / not downloading).
+ */
+const useModelDownloadProgress = (active: boolean): number | null => {
+    const sttLocalModel = useConfigStore((state) => state.sttLocalModel);
+    const [percent, setPercent] = React.useState<number | null>(null);
+
+    React.useEffect(() => {
+        if (!active) {
+            setPercent(null);
+            return;
+        }
+        let cancelled = false;
+        const poll = async () => {
+            try {
+                const response = await runtimeFetch('/api/dictation/status', {
+                    query: { provider: 'local', localModel: sttLocalModel },
+                });
+                if (!response.ok || cancelled) {
+                    return;
+                }
+                const data = await response.json();
+                const model = Array.isArray(data?.models)
+                    ? data.models.find((m: { id: string }) => m.id === sttLocalModel)
+                    : null;
+                if (!cancelled) {
+                    setPercent(typeof model?.downloadProgress === 'number' ? model.downloadProgress : null);
+                }
+            } catch {
+                // Display-only; keep the previous value.
+            }
+        };
+        void poll();
+        const interval = setInterval(() => {
+            void poll();
+        }, 2000);
+        return () => {
+            cancelled = true;
+            clearInterval(interval);
+        };
+    }, [active, sttLocalModel]);
+
+    return active ? percent : null;
+};
+
+export const ComposerDictation: React.FC<ComposerDictationProps> = ({
+    radius,
+    isMobile,
+    footerIconButtonClass,
+    footerPaddingClass,
+    iconSizeClass,
+    sendIconSizeClass,
+    disabled,
+    onInsert,
+    onInsertAndSend,
+}) => {
+    const { t } = useI18n();
+    const { currentTheme } = useThemeSystem();
+    const dictationEnabled = useConfigStore((state) => state.dictationEnabled);
+    const shortcutOverrides = useUIStore((state) => state.shortcutOverrides);
+    const dictationShortcut = formatShortcutForDisplay(getEffectiveShortcutCombo('toggle_dictation', shortcutOverrides));
+    // The dictation server (WebSocket + STT worker) lives in the OpenChamber
+    // web server; the VS Code bridge has no server process for it.
+    const [supported] = React.useState(() => !isVSCodeRuntime() && isDictationCaptureSupported());
+
+    const pendingActionRef = React.useRef<'insert' | 'send' | null>(null);
+    const onInsertRef = React.useRef(onInsert);
+    const onInsertAndSendRef = React.useRef(onInsertAndSend);
+    React.useEffect(() => {
+        onInsertRef.current = onInsert;
+        onInsertAndSendRef.current = onInsertAndSend;
+    }, [onInsert, onInsertAndSend]);
+
+    const dictation = useDictation({
+        onTranscript: (text) => {
+            const action = pendingActionRef.current;
+            pendingActionRef.current = null;
+            if (action === 'send') {
+                onInsertAndSendRef.current(text);
+            } else {
+                onInsertRef.current(text);
+            }
+        },
+    });
+
+    const {
+        status,
+        partialTranscript,
+        volume,
+        duration,
+        error,
+        errorReason,
+        startDictation,
+        confirmDictation,
+        cancelDictation,
+        retryFailedDictation,
+        acceptPartialTranscript,
+        discardFailedDictation,
+    } = dictation;
+
+    const isModelDownloading = status === 'recording' && errorReason === 'model_download_in_progress';
+    const downloadPercent = useModelDownloadProgress(isModelDownloading);
+
+    const statusRef = React.useRef(status);
+    React.useEffect(() => {
+        statusRef.current = status;
+    }, [status]);
+
+    // Keyboard shortcut (toggle_dictation): idle -> start recording,
+    // recording -> confirm and insert. Dispatched by useKeyboardShortcuts.
+    React.useEffect(() => {
+        const onToggle = () => {
+            if (statusRef.current === 'idle') {
+                void startDictation();
+            } else if (statusRef.current === 'recording') {
+                pendingActionRef.current = 'insert';
+                void confirmDictation();
+            }
+        };
+        window.addEventListener('openchamber:dictation-toggle', onToggle);
+        return () => window.removeEventListener('openchamber:dictation-toggle', onToggle);
+    }, [startDictation, confirmDictation]);
+
+    // While recording: Enter confirms (insert), Escape cancels. Capture-phase
+    // so the composer's own Enter-to-send never fires underneath the overlay.
+    React.useEffect(() => {
+        if (status !== 'recording') {
+            return;
+        }
+        const onKeyDown = (event: KeyboardEvent) => {
+            if (event.isComposing) {
+                return;
+            }
+            if (event.key === 'Enter' && !event.shiftKey && !event.metaKey && !event.ctrlKey && !event.altKey) {
+                event.preventDefault();
+                event.stopPropagation();
+                pendingActionRef.current = 'insert';
+                void confirmDictation();
+                return;
+            }
+            if (event.key === 'Escape') {
+                event.preventDefault();
+                event.stopPropagation();
+                void cancelDictation();
+            }
+        };
+        window.addEventListener('keydown', onKeyDown, { capture: true });
+        return () => window.removeEventListener('keydown', onKeyDown, { capture: true });
+    }, [status, confirmDictation, cancelDictation]);
+
+    // Pixel-parity with the composer: the real footer row is taller than our
+    // buttons (its height comes from the tallest control, e.g. the model
+    // picker), so measure it — it stays mounted underneath the overlay — and
+    // give our action row the same height so the icons line up exactly.
+    const overlayRef = React.useRef<HTMLDivElement | null>(null);
+    const [footerHeight, setFooterHeight] = React.useState<number | null>(null);
+    const isActiveStatus = status !== 'idle';
+    React.useLayoutEffect(() => {
+        if (!isActiveStatus) {
+            return;
+        }
+        // The overlay is rendered inside the composer footer itself, so the
+        // real footer is an ancestor, not a sibling.
+        const realFooter = overlayRef.current?.closest<HTMLElement>('[data-chat-input-footer="true"]');
+        if (realFooter && realFooter.offsetHeight > 0) {
+            setFooterHeight(realFooter.offsetHeight);
+        }
+    }, [isActiveStatus]);
+
+    if (!supported || !dictationEnabled) {
+        return null;
+    }
+
+    const isActive = status !== 'idle';
+
+    const confirmWith = (action: 'insert' | 'send') => {
+        pendingActionRef.current = action;
+        void confirmDictation();
+    };
+
+    const retry = () => {
+        pendingActionRef.current = 'insert';
+        void retryFailedDictation();
+    };
+
+    const placeholderText = (() => {
+        if (status === 'failed') {
+            return '';
+        }
+        if (status === 'uploading') {
+            return t('chat.dictation.processing');
+        }
+        if (isModelDownloading) {
+            return downloadPercent !== null
+                ? t('chat.dictation.downloadingModelProgress', { percent: String(downloadPercent) })
+                : t('chat.dictation.downloadingModel');
+        }
+        return t('chat.dictation.listening');
+    })();
+
+    return (
+        <>
+            <button
+                type="button"
+                className={footerIconButtonClass}
+                onClick={() => {
+                    void startDictation();
+                }}
+                disabled={disabled || isActive}
+                title={dictationShortcut ? `${t('chat.dictation.start')} (${dictationShortcut})` : t('chat.dictation.start')}
+                aria-label={t('chat.dictation.start')}
+            >
+                <Icon name="mic" className={cn(iconSizeClass, 'text-current')} />
+            </button>
+            {isActive ? (
+                <div
+                    ref={overlayRef}
+                    // overflow-x/y split on purpose: mobile.css rewrites the
+                    // shorthand `.overflow-hidden` to overflow-y:auto on touch
+                    // devices, which painted a phantom scrollbar on Android.
+                    className="absolute inset-0 z-50 flex flex-col overflow-x-hidden overflow-y-hidden"
+                    style={{
+                        borderRadius: radius,
+                        // Must match the composer box background exactly so the
+                        // overlay reads as the same surface, not a layer on top.
+                        backgroundColor: currentTheme.colors.surface.subtle,
+                    }}
+                    role="dialog"
+                    aria-label={t('chat.dictation.overlayAria')}
+                >
+                    <div
+                        className={cn(
+                            // Text paddings match the composer textarea, plus the
+                            // 4px (pt-1) attachment-chips row that always renders
+                            // above it: desktop 16+4px, mobile 10+4px from the top.
+                            // min-h-0 (not a fixed min height): the mobile composer
+                            // is shorter than 52px of text area + footer, and a
+                            // fixed min pushed the action row 4px below the real
+                            // footer. The area must shrink to whatever space the
+                            // underlying composer actually has.
+                            'flex-1 min-h-0 overflow-y-auto px-3',
+                            isMobile ? 'pt-3.5 pb-2.5' : 'pt-5 pb-2',
+                        )}
+                    >
+                        {partialTranscript ? (
+                            <p className="typography-markdown md:typography-ui-label whitespace-pre-wrap" style={{ color: currentTheme.colors.surface.foreground }}>
+                                {partialTranscript}
+                            </p>
+                        ) : (
+                            <p className="typography-markdown md:typography-ui-label" style={{ color: currentTheme.colors.surface.mutedForeground }}>
+                                {placeholderText}
+                            </p>
+                        )}
+                        {status === 'failed' ? (
+                            <p className="typography-meta mt-1" style={{ color: currentTheme.colors.status.error }}>
+                                {error || t('chat.dictation.failed')}
+                            </p>
+                        ) : null}
+                        {status === 'recording' && error && !isModelDownloading ? (
+                            <p className="typography-meta mt-1" style={{ color: currentTheme.colors.status.warning }}>
+                                {error}
+                            </p>
+                        ) : null}
+                    </div>
+                    <div
+                        className={cn('flex flex-shrink-0 items-center gap-x-3', footerPaddingClass)}
+                        style={footerHeight ? { height: footerHeight } : undefined}
+                    >
+                        {status === 'recording' ? (
+                            <>
+                                <span className="relative ml-1 flex h-2 w-2 flex-shrink-0" aria-hidden="true">
+                                    <span
+                                        className="absolute inline-flex h-full w-full animate-ping rounded-full opacity-60"
+                                        style={{ backgroundColor: currentTheme.colors.status.error }}
+                                    />
+                                    <span
+                                        className="relative inline-flex h-2 w-2 rounded-full"
+                                        style={{ backgroundColor: currentTheme.colors.status.error }}
+                                    />
+                                </span>
+                                <VolumeMeter volume={volume} />
+                                <span className="typography-meta tabular-nums" style={{ color: currentTheme.colors.surface.mutedForeground }}>
+                                    {formatDuration(duration)}
+                                </span>
+                            </>
+                        ) : status === 'uploading' ? (
+                            <Icon name="loader-4" className="ml-1 h-4 w-4 animate-spin" style={{ color: currentTheme.colors.surface.mutedForeground }} />
+                        ) : null}
+                        {/* Same inter-control gap as the composer's right cluster:
+                            gap-x-1 on mobile, md:gap-x-3 on desktop. */}
+                        <div className={cn('ml-auto flex items-center', isMobile ? 'gap-x-1' : 'gap-x-1.5 md:gap-x-3')}>
+                            {status === 'recording' ? (
+                                <>
+                                    <button
+                                        type="button"
+                                        className={cn(footerIconButtonClass, 'text-muted-foreground hover:text-foreground')}
+                                        onClick={() => {
+                                            void cancelDictation();
+                                        }}
+                                        title={t('chat.dictation.cancel')}
+                                        aria-label={t('chat.dictation.cancel')}
+                                    >
+                                        <Icon name="close" className={iconSizeClass} />
+                                    </button>
+                                    <button
+                                        type="button"
+                                        className={footerIconButtonClass}
+                                        onClick={() => confirmWith('insert')}
+                                        title={t('chat.dictation.insert')}
+                                        aria-label={t('chat.dictation.insert')}
+                                    >
+                                        <Icon name="check" className={iconSizeClass} />
+                                    </button>
+                                    <button
+                                        type="button"
+                                        className={cn(footerIconButtonClass, 'text-primary hover:text-primary')}
+                                        onClick={() => confirmWith('send')}
+                                        title={t('chat.dictation.insertAndSend')}
+                                        aria-label={t('chat.dictation.insertAndSend')}
+                                    >
+                                        <Icon name="send-plane-2" className={sendIconSizeClass} />
+                                    </button>
+                                </>
+                            ) : status === 'uploading' ? (
+                                <button
+                                    type="button"
+                                    className={cn(footerIconButtonClass, 'text-muted-foreground hover:text-foreground')}
+                                    onClick={() => {
+                                        void cancelDictation();
+                                    }}
+                                    title={t('chat.dictation.cancel')}
+                                    aria-label={t('chat.dictation.cancel')}
+                                >
+                                    <Icon name="close" className={iconSizeClass} />
+                                </button>
+                            ) : (
+                                <>
+                                    <button
+                                        type="button"
+                                        className={cn(footerIconButtonClass, 'text-muted-foreground hover:text-foreground')}
+                                        onClick={discardFailedDictation}
+                                        title={t('chat.dictation.discard')}
+                                        aria-label={t('chat.dictation.discard')}
+                                    >
+                                        <Icon name="close" className={iconSizeClass} />
+                                    </button>
+                                    <button
+                                        type="button"
+                                        className={footerIconButtonClass}
+                                        onClick={retry}
+                                        title={t('chat.dictation.retry')}
+                                        aria-label={t('chat.dictation.retry')}
+                                    >
+                                        <Icon name="refresh" className={iconSizeClass} />
+                                    </button>
+                                    {partialTranscript.trim() ? (
+                                        <button
+                                            type="button"
+                                            className={footerIconButtonClass}
+                                            onClick={() => {
+                                                pendingActionRef.current = 'insert';
+                                                acceptPartialTranscript();
+                                            }}
+                                            title={t('chat.dictation.insert')}
+                                            aria-label={t('chat.dictation.insert')}
+                                        >
+                                            <Icon name="check" className={iconSizeClass} />
+                                        </button>
+                                    ) : null}
+                                </>
+                            )}
+                        </div>
+                    </div>
+                </div>
+            ) : null}
+        </>
+    );
+};

packages/ui/src/components/sections/openchamber/OpenChamberVisualSettings.tsx

@@ -7,7 +7,6 @@ import type { ThemeMode } from '@/types/theme';
 import { useUIStore } from '@/stores/useUIStore';
 import { useMessageQueueStore, type FollowUpBehavior } from '@/stores/messageQueueStore';
 import { cn } from '@/lib/utils';
-import { isCapacitorApp } from '@/lib/platform';
 import { Button } from '@/components/ui/button';
 import { Checkbox } from '@/components/ui/checkbox';
 import { NumberInput } from '@/components/ui/number-input';
@@ -322,10 +321,7 @@ export const OpenChamberVisualSettings: React.FC<OpenChamberVisualSettingsProps>
     const setShowSplitAssistantMessageActions = useUIStore(state => state.setShowSplitAssistantMessageActions);
     const messageStreamTransport = useConfigStore((state) => state.settingsMessageStreamTransport);
     const setMessageStreamTransport = useConfigStore((state) => state.setSettingsMessageStreamTransport);
-    // Capacitor apps are locked to SSE (native WebSocket streaming is unreliable on mobile);
+    const effectiveMessageStreamTransport = messageStreamTransport;
-    // sync-context forces it too. Show SSE selected and disable the other options here.
-    const isCapacitorAppRuntime = React.useMemo(() => isCapacitorApp(), []);
-    const effectiveMessageStreamTransport = isCapacitorAppRuntime ? 'sse' : messageStreamTransport;
     const settingsDefaultFileViewerPreview = useConfigStore((state) => state.settingsDefaultFileViewerPreview);
     const setSettingsDefaultFileViewerPreview = useConfigStore((state) => state.setSettingsDefaultFileViewerPreview);
     const isSettingsDialogOpen = useUIStore(state => state.isSettingsDialogOpen);
@@ -1531,7 +1527,6 @@ export const OpenChamberVisualSettings: React.FC<OpenChamberVisualSettingsProps>
                                                             variant="chip"
                                                             size="xs"
                                                             aria-pressed={effectiveMessageStreamTransport === option.id}
-                                                            disabled={isCapacitorAppRuntime && option.id !== 'sse'}
                                                             className="!font-normal"
                                                             onClick={() => handleMessageStreamTransportChange(option.id)}
                                                         >

packages/ui/src/components/sections/openchamber/VoiceSettings.tsx

@@ -1,5 +1,4 @@
 import React, { useEffect, useState, useCallback, useMemo } from 'react';
-import { useBrowserVoice } from '@/hooks/useBrowserVoice';
 import { useConfigStore } from '@/stores/useConfigStore';
 import { useDeviceInfo } from '@/lib/device';
 
@@ -11,103 +10,362 @@ import {
     SelectValue,
 } from '@/components/ui/select';
 import { Checkbox } from '@/components/ui/checkbox';
+import { Radio } from '@/components/ui/radio';
 import { Button } from '@/components/ui/button';
 import { NumberInput } from '@/components/ui/number-input';
 import { Tooltip, TooltipContent, TooltipTrigger } from '@/components/ui/tooltip';
 import { Icon } from "@/components/icon/Icon";
 import { browserVoiceService } from '@/lib/voice/browserVoiceService';
-import { audioStreamService } from '@/lib/voice/audioStreamService';
-import { wasmSttService, WASM_MODELS } from '@/lib/voice/wasmSttService';
-import type { WasmModelStatus } from '@/lib/voice/wasmSttService';
 import { cn } from '@/lib/utils';
 import { runtimeFetch } from '@/lib/runtime-fetch';
 import { useI18n } from '@/lib/i18n';
+import { useLocalTTS } from '@/hooks/useLocalTTS';
 import { disposePreviewAudio } from './voicePreviewAudio';
-const LANGUAGE_OPTIONS = [
-    { value: 'en-US', label: 'English' },
-    { value: 'es-ES', label: 'Español' },
-    { value: 'fr-FR', label: 'Français' },
-    { value: 'de-DE', label: 'Deutsch' },
-    { value: 'ja-JP', label: '日本語' },
-    { value: 'zh-CN', label: '中文' },
-    { value: 'pt-BR', label: 'Português' },
-    { value: 'it-IT', label: 'Italiano' },
-    { value: 'ko-KR', label: '한국어' },
-    { value: 'uk-UA', label: 'Українська' },
-];
 
-const WasmModelStatusIndicator = ({ modelId }: { modelId: string }) => {
+const LOCAL_STT_MODELS = [
+    {
+        id: 'parakeet-tdt-0.6b-v2-int8',
+        labelKey: 'settings.voice.page.stt.model.parakeetV2',
+        badgeKey: 'settings.voice.page.stt.badge.bestForEnglish',
+        accuracy: 5,
+        speed: 5,
+        size: '460 MB',
+    },
+    {
+        id: 'parakeet-tdt-0.6b-v3-int8',
+        labelKey: 'settings.voice.page.stt.model.parakeetV3',
+        badgeKey: 'settings.voice.page.stt.badge.bestForMultilingual',
+        accuracy: 5,
+        speed: 4,
+        size: '465 MB',
+    },
+    {
+        id: 'whisper-base-int8',
+        labelKey: 'settings.voice.page.stt.model.whisperBase',
+        badgeKey: null,
+        accuracy: 3,
+        speed: 3,
+        size: '200 MB',
+    },
+    {
+        id: 'whisper-tiny-int8',
+        labelKey: 'settings.voice.page.stt.model.whisperTiny',
+        badgeKey: null,
+        accuracy: 2,
+        speed: 4,
+        size: '115 MB',
+    },
+] as const;
+
+interface DictationModelState {
+    id: string;
+    installed: boolean;
+    downloading: boolean;
+    downloadProgress: number | null;
+    downloadError: string | null;
+}
+
+const RatingBar = ({ value, label }: { value: number; label: string }) => (
+    <span className="flex items-center gap-1.5" title={`${label}: ${value}/5`}>
+        <span
+            className="h-1.5 w-12 flex-shrink-0 overflow-hidden rounded-full bg-[var(--interactive-border)]"
+            aria-hidden="true"
+        >
+            <span
+                className="block h-full rounded-full bg-[var(--primary-base)]"
+                style={{ width: `${(value / 5) * 100}%` }}
+            />
+        </span>
+        <span className="typography-ui-compact text-muted-foreground">{label}</span>
+    </span>
+);
+
+const LocalModelPicker = ({
+    selectedModelId,
+    onSelect,
+}: {
+    selectedModelId: string;
+    onSelect: (modelId: string) => void;
+}) => {
     const { t } = useI18n();
-    const [status, setStatus] = useState<WasmModelStatus>(wasmSttService.getModelStatus());
+    const tUnsafe = useCallback((key: string) => t(key as Parameters<typeof t>[0]), [t]);
-    const [loading, setLoading] = useState(false);
+    const [models, setModels] = useState<Map<string, DictationModelState>>(new Map());
+    const [requestingId, setRequestingId] = useState<string | null>(null);
 
-    useEffect(() => {
+    const refresh = useCallback(async () => {
-        const handler = (s: WasmModelStatus) => setStatus(s);
+        try {
-        wasmSttService.onModelStatusChange = handler;
+            const response = await runtimeFetch('/api/dictation/status', {
-        return () => { wasmSttService.onModelStatusChange = null; };
+                query: { provider: 'local' },
+            });
+            if (!response.ok) {
+                return;
+            }
+            const data = await response.json();
+            if (Array.isArray(data?.models)) {
+                setModels(new Map(data.models.map((m: DictationModelState) => [m.id, m])));
+            }
+        } catch {
+            // Display-only status; keep the previous state on fetch failure.
+        }
     }, []);
 
-    const currentModelId = wasmSttService.getCurrentModelId();
-    const isLoadingOrDownloading = status.state === 'downloading' || status.state === 'loading';
-
-    // Reset local loading state when model finishes loading or errors
     useEffect(() => {
-        if (status.state !== 'downloading' && status.state !== 'loading') {
+        void refresh();
-            setLoading(false);
+    }, [refresh]);
-        }
-    }, [status.state]);
 
-    const handleDownload = async () => {
+    const anyDownloading = Array.from(models.values()).some((m) => m.downloading);
-        setLoading(true);
+    useEffect(() => {
+        if (!anyDownloading) {
+            return;
+        }
+        const interval = setInterval(() => {
+            void refresh();
+        }, 2000);
+        return () => clearInterval(interval);
+    }, [anyDownloading, refresh]);
+
+    const handleDownload = async (modelId: string) => {
+        setRequestingId(modelId);
         try {
-            await wasmSttService.loadModel(modelId);
+            await runtimeFetch(`/api/dictation/models/${encodeURIComponent(modelId)}/download`, {
+                method: 'POST',
+            });
+            await refresh();
         } catch {
-            // Error is shown via status indicator
+            // Status refresh reports errors.
+        } finally {
+            setRequestingId(null);
         }
     };
 
-    if (status.state === 'ready' && currentModelId === modelId) {
+    const handleDelete = async (modelId: string) => {
-        return (
+        setRequestingId(modelId);
-            <div className="flex items-center gap-2">
+        try {
-                <span className="typography-ui-compact text-green-600 dark:text-green-400">
+            await runtimeFetch(`/api/dictation/models/${encodeURIComponent(modelId)}`, {
-                    {t('settings.voice.page.stt.wasmLoaded')}
+                method: 'DELETE',
-                </span>
+            });
-            </div>
+            await refresh();
-        );
+        } catch {
+            // Status refresh reports errors.
+        } finally {
+            setRequestingId(null);
         }
+    };
 
-    if (isLoadingOrDownloading) {
-        const progress = status.state === 'downloading' ? Math.round(status.progress) : undefined;
     return (
-            <div className="flex items-center gap-2">
+        <div role="radiogroup" aria-label={t('settings.voice.page.field.model')} className="mt-1 space-y-0">
-                <span className="typography-ui-compact text-muted-foreground">
+            {LOCAL_STT_MODELS.map((entry) => {
-                    {status.state === 'downloading' ? t('settings.voice.page.stt.wasmDownloading') : t('settings.voice.page.stt.wasmLoading')}
+                const selected = selectedModelId === entry.id;
-                    {progress !== undefined ? ` (${progress}%)` : ''}
+                const state = models.get(entry.id) ?? null;
-                </span>
+                return (
+                    <div key={entry.id} className="flex w-full items-start gap-2 py-1.5">
+                        <div className="pt-0.5">
+                            <Radio
+                                checked={selected}
+                                onChange={() => onSelect(entry.id)}
+                                ariaLabel={tUnsafe(entry.labelKey)}
+                            />
                         </div>
-        );
+                        <div
-    }
+                            className="min-w-0 flex-1 cursor-pointer"
-
+                            role="button"
-    if (status.state === 'error') {
+                            tabIndex={0}
-        // Partial download (cached progress): show retry button.
+                            onClick={() => onSelect(entry.id)}
-        return (
+                            onKeyDown={(e) => { if (e.key === ' ' || e.key === 'Enter') { e.preventDefault(); onSelect(entry.id); } }}
-            <div className="flex items-center gap-2">
+                        >
-                <span className="typography-ui-compact text-destructive">{status.error}</span>
+                            <div className="flex flex-wrap items-center gap-x-2 gap-y-0.5">
-                <Button variant="chip" size="xs" disabled={loading} onClick={handleDownload}>
+                                <span className={cn('typography-ui-label font-normal', selected ? 'text-foreground' : 'text-foreground/50')}>
-                    {t('settings.voice.page.stt.wasmRetry')}
+                                    {tUnsafe(entry.labelKey)}
+                                </span>
+                                {entry.badgeKey ? (
+                                    <span
+                                        className="rounded px-1 text-[9px] font-medium uppercase leading-[14px] tracking-wide"
+                                        style={{
+                                            backgroundColor: 'var(--status-success-background)',
+                                            color: 'var(--status-success)',
+                                            border: '1px solid var(--status-success-border)',
+                                        }}
+                                    >
+                                        {tUnsafe(entry.badgeKey)}
+                                    </span>
+                                ) : null}
+                            </div>
+                            <div className="mt-0.5 flex flex-wrap items-center gap-x-4 gap-y-0.5">
+                                <RatingBar value={entry.accuracy} label={t('settings.voice.page.stt.meta.accuracy')} />
+                                <RatingBar value={entry.speed} label={t('settings.voice.page.stt.meta.speed')} />
+                                <span className="typography-ui-compact tabular-nums text-muted-foreground">{entry.size}</span>
+                            </div>
+                            {state?.downloadError ? (
+                                <p className="mt-0.5 typography-meta text-[var(--status-error)]">{state.downloadError}</p>
+                            ) : null}
+                        </div>
+                        <div className="flex w-14 flex-shrink-0 items-center justify-end gap-1 pt-0.5">
+                            {state?.installed ? (
+                                <>
+                                    <Button
+                                        variant="ghost"
+                                        size="xs"
+                                        className="h-6 w-6 p-0 text-muted-foreground hover:text-[var(--status-error)]"
+                                        disabled={requestingId === entry.id}
+                                        onClick={() => { void handleDelete(entry.id); }}
+                                        title={t('settings.voice.page.stt.modelDelete')}
+                                        aria-label={t('settings.voice.page.stt.modelDelete')}
+                                    >
+                                        <Icon name="delete-bin" className="h-4 w-4" />
                                     </Button>
+                                    <Icon
+                                        name="checkbox-circle"
+                                        className="h-4 w-4 flex-shrink-0 text-[var(--status-success)]"
+                                        aria-label={t('settings.voice.page.stt.modelInstalled')}
+                                    />
+                                </>
+                            ) : state?.downloading ? (
+                                <span className="flex items-center gap-1.5">
+                                    <Icon name="loader-4" className="h-3.5 w-3.5 animate-spin text-muted-foreground" />
+                                    <span className="typography-ui-compact tabular-nums text-muted-foreground">
+                                        {typeof state.downloadProgress === 'number' ? `${state.downloadProgress}%` : ''}
+                                    </span>
+                                </span>
+                            ) : (
+                                <Button
+                                    variant="ghost"
+                                    size="xs"
+                                    className="h-6 w-6 p-0"
+                                    disabled={requestingId === entry.id}
+                                    onClick={() => { void handleDownload(entry.id); }}
+                                    title={t('settings.voice.page.stt.modelDownload')}
+                                    aria-label={t('settings.voice.page.stt.modelDownload')}
+                                >
+                                    <Icon name="download" className="h-4 w-4" />
+                                </Button>
+                            )}
+                        </div>
                     </div>
                 );
+            })}
+        </div>
+    );
+};
+
+/** Kokoro en_v0_19 speaker ids in sherpa-onnx order. */
+const KOKORO_VOICE_OPTIONS = [
+    { id: 0, label: 'Alloy (af)' },
+    { id: 1, label: 'Bella (af)' },
+    { id: 2, label: 'Nicole (af)' },
+    { id: 3, label: 'Sarah (af)' },
+    { id: 4, label: 'Sky (af)' },
+    { id: 5, label: 'Adam (am)' },
+    { id: 6, label: 'Michael (am)' },
+    { id: 7, label: 'Emma (bf)' },
+    { id: 8, label: 'Isabella (bf)' },
+    { id: 9, label: 'George (bm)' },
+    { id: 10, label: 'Lewis (bm)' },
+];
+
+const LOCAL_TTS_MODEL_ID = 'kokoro-en-v0_19';
+
+const LocalTtsModelStatus = () => {
+    const { t } = useI18n();
+    const [model, setModel] = useState<DictationModelState | null>(null);
+    const [requesting, setRequesting] = useState(false);
+
+    const refresh = useCallback(async () => {
+        try {
+            const response = await runtimeFetch('/api/dictation/status', { query: { provider: 'local' } });
+            if (!response.ok) {
+                return;
+            }
+            const data = await response.json();
+            const entry = Array.isArray(data?.ttsModels)
+                ? data.ttsModels.find((m: DictationModelState) => m.id === LOCAL_TTS_MODEL_ID)
+                : null;
+            if (entry) {
+                setModel(entry);
+            }
+        } catch {
+            // Display-only status; keep the previous state on fetch failure.
+        }
+    }, []);
+
+    useEffect(() => {
+        void refresh();
+    }, [refresh]);
+
+    useEffect(() => {
+        if (!model?.downloading) {
+            return;
+        }
+        const interval = setInterval(() => {
+            void refresh();
+        }, 2000);
+        return () => clearInterval(interval);
+    }, [model?.downloading, refresh]);
+
+    const request = async (method: 'POST' | 'DELETE') => {
+        setRequesting(true);
+        try {
+            const path = method === 'POST'
+                ? `/api/dictation/models/${LOCAL_TTS_MODEL_ID}/download`
+                : `/api/dictation/models/${LOCAL_TTS_MODEL_ID}`;
+            await runtimeFetch(path, { method });
+            await refresh();
+        } catch {
+            // Status refresh reports errors.
+        } finally {
+            setRequesting(false);
+        }
+    };
+
+    if (!model) {
+        return null;
     }
 
     return (
-        <div className="flex items-center gap-2">
+        <div className="flex items-center gap-2 py-1.5">
-            <span className="typography-ui-compact text-muted-foreground">
+            <span className="typography-ui-label text-foreground">Kokoro</span>
-                {t('settings.voice.page.stt.wasmNotLoaded')}
+            <span className="typography-ui-compact tabular-nums text-muted-foreground">305 MB</span>
-            </span>
+            {model.installed ? (
-            <Button variant="chip" size="xs" disabled={loading} onClick={handleDownload}>
+                <>
-                {t('settings.voice.page.stt.wasmDownload')}
+                    <Icon
+                        name="checkbox-circle"
+                        className="h-4 w-4 text-[var(--status-success)]"
+                        aria-label={t('settings.voice.page.stt.modelInstalled')}
+                    />
+                    <Button
+                        variant="ghost"
+                        size="xs"
+                        className="h-6 w-6 p-0 text-muted-foreground hover:text-[var(--status-error)]"
+                        disabled={requesting}
+                        onClick={() => { void request('DELETE'); }}
+                        title={t('settings.voice.page.stt.modelDelete')}
+                        aria-label={t('settings.voice.page.stt.modelDelete')}
+                    >
+                        <Icon name="delete-bin" className="h-4 w-4" />
                     </Button>
+                </>
+            ) : model.downloading ? (
+                <span className="flex items-center gap-1.5">
+                    <Icon name="loader-4" className="h-3.5 w-3.5 animate-spin text-muted-foreground" />
+                    <span className="typography-ui-compact tabular-nums text-muted-foreground">
+                        {typeof model.downloadProgress === 'number' ? `${model.downloadProgress}%` : ''}
+                    </span>
+                </span>
+            ) : (
+                <Button
+                    variant="ghost"
+                    size="xs"
+                    className="h-6 w-6 p-0"
+                    disabled={requesting}
+                    onClick={() => { void request('POST'); }}
+                    title={t('settings.voice.page.stt.modelDownload')}
+                    aria-label={t('settings.voice.page.stt.modelDownload')}
+                >
+                    <Icon name="download" className="h-4 w-4" />
+                </Button>
+            )}
+            {model.downloadError ? (
+                <span className="typography-meta text-[var(--status-error)]">{model.downloadError}</span>
+            ) : null}
         </div>
     );
 };
@@ -131,11 +389,6 @@ const OPENAI_VOICE_OPTIONS = [
 export const VoiceSettings: React.FC = () => {
     const { t } = useI18n();
     const { isMobile } = useDeviceInfo();
-    const {
-        isSupported,
-        language,
-        setLanguage,
-    } = useBrowserVoice();
     const voiceProvider = useConfigStore((state) => state.voiceProvider);
     const setVoiceProvider = useConfigStore((state) => state.setVoiceProvider);
     const speechRate = useConfigStore((state) => state.speechRate);
@@ -146,6 +399,22 @@ export const VoiceSettings: React.FC = () => {
     const setSpeechVolume = useConfigStore((state) => state.setSpeechVolume);
     const sayVoice = useConfigStore((state) => state.sayVoice);
     const setSayVoice = useConfigStore((state) => state.setSayVoice);
+    const localTtsVoiceId = useConfigStore((state) => state.localTtsVoiceId);
+    const setLocalTtsVoiceId = useConfigStore((state) => state.setLocalTtsVoiceId);
+    const { speak: speakLocalTts, stop: stopLocalTts, isPlaying: isLocalTtsPlaying, error: localTtsError } = useLocalTTS();
+
+    const previewLocalVoice = useCallback(() => {
+        if (isLocalTtsPlaying) {
+            stopLocalTts();
+            return;
+        }
+        const voiceLabel = KOKORO_VOICE_OPTIONS.find((v) => v.id === localTtsVoiceId)?.label
+            ?? String(localTtsVoiceId);
+        void speakLocalTts(t('settings.voice.page.preview.voiceLine', { voiceName: voiceLabel }), {
+            speakerId: localTtsVoiceId,
+            speed: useConfigStore.getState().speechRate,
+        });
+    }, [isLocalTtsPlaying, localTtsVoiceId, speakLocalTts, stopLocalTts, t]);
     const browserVoice = useConfigStore((state) => state.browserVoice);
     const setBrowserVoice = useConfigStore((state) => state.setBrowserVoice);
     const openaiVoice = useConfigStore((state) => state.openaiVoice);
@@ -172,19 +441,13 @@ export const VoiceSettings: React.FC = () => {
     const setSttApiKey = useConfigStore((state) => state.setSttApiKey);
     const sttModel = useConfigStore((state) => state.sttModel);
     const setSttModel = useConfigStore((state) => state.setSttModel);
-    const wasmSttModel = useConfigStore((state) => state.wasmSttModel);
+    const sttLocalModel = useConfigStore((state) => state.sttLocalModel);
-    const setWasmSttModel = useConfigStore((state) => state.setWasmSttModel);
+    const setSttLocalModel = useConfigStore((state) => state.setSttLocalModel);
     const sttLanguage = useConfigStore((state) => state.sttLanguage);
     const setSttLanguage = useConfigStore((state) => state.setSttLanguage);
-    const sttSilenceThresholdDb = useConfigStore((state) => state.sttSilenceThresholdDb);
-    const setSttSilenceThresholdDb = useConfigStore((state) => state.setSttSilenceThresholdDb);
-    const sttSilenceHoldMs = useConfigStore((state) => state.sttSilenceHoldMs);
-    const setSttSilenceHoldMs = useConfigStore((state) => state.setSttSilenceHoldMs);
-    const sttTranscribeOnStop = useConfigStore((state) => state.sttTranscribeOnStop);
-    const setSttTranscribeOnStop = useConfigStore((state) => state.setSttTranscribeOnStop);
     const setShowMessageTTSButtons = useConfigStore((state) => state.setShowMessageTTSButtons);
-    const voiceModeEnabled = useConfigStore((state) => state.voiceModeEnabled);
+    const dictationEnabled = useConfigStore((state) => state.dictationEnabled);
-    const setVoiceModeEnabled = useConfigStore((state) => state.setVoiceModeEnabled);
+    const setDictationEnabled = useConfigStore((state) => state.setDictationEnabled);
 
     const [isSayAvailable, setIsSayAvailable] = useState(false);
     const [sayVoices, setSayVoices] = useState<Array<{ name: string; locale: string }>>([]);
@@ -274,7 +537,7 @@ export const VoiceSettings: React.FC = () => {
     }, [isBrowserPreviewPlaying]);
 
     useEffect(() => {
-        if (!voiceModeEnabled || (voiceProvider !== 'openai' && voiceProvider !== 'openai-compatible')) {
+        if (!showMessageTTSButtons || (voiceProvider !== 'openai' && voiceProvider !== 'openai-compatible')) {
             setIsOpenAIAvailable(openaiApiKey.trim().length > 0);
             return;
         }
@@ -292,10 +555,10 @@ export const VoiceSettings: React.FC = () => {
         };
 
         checkOpenAIAvailability();
-    }, [openaiApiKey, voiceModeEnabled, voiceProvider]);
+    }, [openaiApiKey, showMessageTTSButtons, voiceProvider]);
 
     useEffect(() => {
-        if (!voiceModeEnabled) {
+        if (!showMessageTTSButtons) {
             setIsSayAvailable(false);
             setSayVoices([]);
             return;
@@ -317,7 +580,7 @@ export const VoiceSettings: React.FC = () => {
             .catch(() => {
                 setIsSayAvailable(false);
             });
-    }, [voiceModeEnabled]);
+    }, [showMessageTTSButtons]);
 
     const previewVoice = useCallback(async () => {
         if (previewAudio) {
@@ -498,11 +761,11 @@ export const VoiceSettings: React.FC = () => {
     return (
         <div className="space-y-8">
 
-            {/* Voice Setup */}
+            {/* Playback (read messages aloud) */}
-            <div data-settings-item="voice.voice-setup" className="mb-8">
+            <div data-settings-item="voice.playback" className="mb-8">
                 <div className="mb-1 px-1">
                     <h3 className="typography-ui-header font-medium text-foreground">
-                        {t('settings.voice.page.section.voiceSetup')}
+                        {t('settings.voice.page.section.playbackAndSummary')}
                     </h3>
                 </div>
 
@@ -512,15 +775,15 @@ export const VoiceSettings: React.FC = () => {
                         className="group flex cursor-pointer items-center gap-2 py-1.5"
                         role="button"
                         tabIndex={0}
-                        aria-pressed={voiceModeEnabled}
+                        aria-pressed={showMessageTTSButtons}
-                        onClick={() => setVoiceModeEnabled(!voiceModeEnabled)}
+                        onClick={() => setShowMessageTTSButtons(!showMessageTTSButtons)}
-                        onKeyDown={(e) => { if (e.key === ' ' || e.key === 'Enter') { e.preventDefault(); setVoiceModeEnabled(!voiceModeEnabled); } }}
+                        onKeyDown={(e) => { if (e.key === ' ' || e.key === 'Enter') { e.preventDefault(); setShowMessageTTSButtons(!showMessageTTSButtons); } }}
                     >
-                        <Checkbox checked={voiceModeEnabled} onChange={setVoiceModeEnabled} ariaLabel={t('settings.voice.page.field.enableVoiceModeAria')} />
+                        <Checkbox checked={showMessageTTSButtons} onChange={setShowMessageTTSButtons} ariaLabel={t('settings.voice.page.field.messageReadAloudButtonAria')} />
-                        <span className="typography-ui-label text-foreground">{t('settings.voice.page.field.enableVoiceMode')}</span>
+                        <span className="typography-ui-label text-foreground">{t('settings.voice.page.field.messageReadAloudButton')}</span>
                     </div>
 
-                    {voiceModeEnabled && (
+                    {showMessageTTSButtons && (
                         <>
                             <div className="pb-1.5 pt-0.5">
                                 <div className="flex min-w-0 flex-col gap-1.5">
@@ -533,6 +796,7 @@ export const VoiceSettings: React.FC = () => {
                                             <TooltipContent sideOffset={8} className="max-w-xs">
                                                 <ul className="space-y-1">
                                                     <li><strong>{t('settings.voice.page.provider.browser')}</strong> {t('settings.voice.page.tooltip.browser')}</li>
+                                                    <li><strong>{t('settings.voice.page.provider.local')}</strong> {t('settings.voice.page.tooltip.localTts')}</li>
                                                     <li><strong>OpenAI:</strong> {t('settings.voice.page.tooltip.openai')}</li>
                                                     <li><strong>{t('settings.voice.page.provider.custom')}</strong> {t('settings.voice.page.tooltip.custom')}</li>
                                                     <li><strong>{t('settings.voice.page.provider.say')}</strong> {t('settings.voice.page.tooltip.say')}</li>
@@ -550,6 +814,15 @@ export const VoiceSettings: React.FC = () => {
                                         >
                                             {t('settings.voice.page.provider.browser')}
                                         </Button>
+                                        <Button
+                                            variant="chip"
+                                            size="xs"
+                                            aria-pressed={voiceProvider === 'local'}
+                                            onClick={() => setVoiceProvider('local')}
+                                            className="!font-normal"
+                                        >
+                                            {t('settings.voice.page.provider.local')}
+                                        </Button>
                                         <Button
                                             variant="chip"
                                             size="xs"
@@ -706,10 +979,39 @@ export const VoiceSettings: React.FC = () => {
                                 </div>
                             )}
 
+                            {/* Local (Kokoro) TTS model status */}
+                            {voiceProvider === 'local' && <LocalTtsModelStatus />}
+
                             {/* Voice Selection */}
                             <div className="flex items-center gap-8 py-1.5">
                                 <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.voice')}</span>
                                 <div className="flex items-center gap-2 w-fit">
+                                    {voiceProvider === 'local' && (
+                                        <>
+                                            <Select
+                                                value={String(localTtsVoiceId)}
+                                                onValueChange={(value) => setLocalTtsVoiceId(Number.parseInt(value, 10) || 0)}
+                                            >
+                                                <SelectTrigger className="w-fit">
+                                                    <SelectValue placeholder={t('settings.voice.page.field.selectVoicePlaceholder')}>
+                                                        {(value) => KOKORO_VOICE_OPTIONS.find((v) => String(v.id) === value)?.label ?? value}
+                                                    </SelectValue>
+                                                </SelectTrigger>
+                                                <SelectContent>
+                                                    {KOKORO_VOICE_OPTIONS.map((v) => (
+                                                        <SelectItem key={v.id} value={String(v.id)}>{v.label}</SelectItem>
+                                                    ))}
+                                                </SelectContent>
+                                            </Select>
+                                            <Button size="xs" variant="ghost" onClick={previewLocalVoice} title={t('settings.voice.page.actions.preview')}>
+                                                {isLocalTtsPlaying ? <Icon name="stop" className="w-3.5 h-3.5" /> : <Icon name="play" className="w-3.5 h-3.5" />}
+                                            </Button>
+                                            {localTtsError ? (
+                                                <span className="typography-meta text-[var(--status-error)]">{localTtsError}</span>
+                                            ) : null}
+                                        </>
+                                    )}
+
                                     {voiceProvider === 'openai' && isOpenAIAvailable && (
                                         <>
                                             <Select value={openaiVoice} onValueChange={setOpenaiVoice}>
@@ -775,7 +1077,7 @@ export const VoiceSettings: React.FC = () => {
                             <div className="flex items-center gap-8 py-1.5">
                                 <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.speechRate')}</span>
                                 <div className="flex items-center gap-2 w-fit">
-                                    {!isMobile && <input type="range" min={0.5} max={2} step={0.1} value={speechRate} onChange={(e) => setSpeechRate(Number(e.target.value))} disabled={!isSupported} className={sliderClass} />}
+                                    {!isMobile && <input type="range" min={0.5} max={2} step={0.1} value={speechRate} onChange={(e) => setSpeechRate(Number(e.target.value))} className={sliderClass} />}
                                     <NumberInput value={speechRate} onValueChange={setSpeechRate} min={0.5} max={2} step={0.1} className="w-16 tabular-nums" />
                                 </div>
                             </div>
@@ -784,7 +1086,7 @@ export const VoiceSettings: React.FC = () => {
                             <div className="flex items-center gap-8 py-1.5">
                                 <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.speechPitch')}</span>
                                 <div className="flex items-center gap-2 w-fit">
-                                    {!isMobile && <input type="range" min={0.5} max={2} step={0.1} value={speechPitch} onChange={(e) => setSpeechPitch(Number(e.target.value))} disabled={!isSupported} className={sliderClass} />}
+                                    {!isMobile && <input type="range" min={0.5} max={2} step={0.1} value={speechPitch} onChange={(e) => setSpeechPitch(Number(e.target.value))} className={sliderClass} />}
                                     <NumberInput value={speechPitch} onValueChange={setSpeechPitch} min={0.5} max={2} step={0.1} className="w-16 tabular-nums" />
                                 </div>
                             </div>
@@ -793,7 +1095,7 @@ export const VoiceSettings: React.FC = () => {
                             <div className="flex items-center gap-8 py-1.5">
                                 <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.speechVolume')}</span>
                                 <div className="flex items-center gap-2 w-fit">
-                                    {!isMobile && <input type="range" min={0} max={1} step={0.1} value={speechVolume} onChange={(e) => setSpeechVolume(Number(e.target.value))} disabled={!isSupported} className={sliderClass} />}
+                                    {!isMobile && <input type="range" min={0} max={1} step={0.1} value={speechVolume} onChange={(e) => setSpeechVolume(Number(e.target.value))} className={sliderClass} />}
                                     {isMobile ? (
                                         <NumberInput value={Math.round(speechVolume * 100)} onValueChange={(v) => setSpeechVolume(v / 100)} min={0} max={100} step={10} className="w-16 tabular-nums" />
                                     ) : (
@@ -804,20 +1106,34 @@ export const VoiceSettings: React.FC = () => {
                                 </div>
                             </div>
 
-                            {/* Language */}
+                            {/* TTS input mode */}
-                            <div className="flex items-center gap-8 py-1.5">
+                            <div className="pb-1.5 pt-0.5">
-                                <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.language')}</span>
+                                <div className="flex min-w-0 flex-col gap-1.5">
-                                <div className="flex items-center gap-2 w-fit">
+                                    <div className="flex items-center gap-1.5">
-                                    <Select value={language} onValueChange={setLanguage} disabled={!isSupported}>
+                                        <span className="typography-ui-label text-foreground">
-                                        <SelectTrigger className="w-fit">
+                                            {t('settings.voice.page.field.ttsInputMode')}
-                                            <SelectValue placeholder={t('settings.voice.page.field.selectLanguagePlaceholder')} />
+                                        </span>
-                                        </SelectTrigger>
+                                    </div>
-                                        <SelectContent>
+                                    <div className="flex flex-wrap items-center gap-1">
-                                            {LANGUAGE_OPTIONS.map((lang) => (
+                                        <Button
-                                                <SelectItem key={lang.value} value={lang.value}>{lang.label}</SelectItem>
+                                            variant="chip"
-                                            ))}
+                                            size="xs"
-                                        </SelectContent>
+                                            aria-pressed={ttsInputMode === 'sanitized'}
-                                    </Select>
+                                            onClick={() => setTtsInputMode('sanitized')}
+                                            className="!font-normal"
+                                        >
+                                            {t('settings.voice.page.field.ttsInputModeSanitized')}
+                                        </Button>
+                                        <Button
+                                            variant="chip"
+                                            size="xs"
+                                            aria-pressed={ttsInputMode === 'raw'}
+                                            onClick={() => setTtsInputMode('raw')}
+                                            className="!font-normal"
+                                        >
+                                            {t('settings.voice.page.field.ttsInputModeRaw')}
+                                        </Button>
+                                    </div>
                                 </div>
                             </div>
                         </>
@@ -826,7 +1142,6 @@ export const VoiceSettings: React.FC = () => {
             </div>
 
             {/* Speech Recognition */}
-            {voiceModeEnabled && (
             <div data-settings-item="voice.speech-recognition" className="mb-8">
                     <div className="mb-1 px-1">
                         <h3 className="typography-ui-header font-medium text-foreground">
@@ -835,6 +1150,19 @@ export const VoiceSettings: React.FC = () => {
                     </div>
 
                     <section className="px-2 pb-2 pt-0 space-y-0">
+                        <div
+                            className="group flex cursor-pointer items-center gap-2 py-1.5"
+                            role="button"
+                            tabIndex={0}
+                            aria-pressed={dictationEnabled}
+                            onClick={() => setDictationEnabled(!dictationEnabled)}
+                            onKeyDown={(e) => { if (e.key === ' ' || e.key === 'Enter') { e.preventDefault(); setDictationEnabled(!dictationEnabled); } }}
+                        >
+                            <Checkbox checked={dictationEnabled} onChange={setDictationEnabled} ariaLabel={t('settings.voice.page.field.enableVoiceInputAria')} />
+                            <span className="typography-ui-label text-foreground">{t('settings.voice.page.field.enableVoiceInput')}</span>
+                        </div>
+
+                        {dictationEnabled && (<>
                         <div className="pb-1.5 pt-0.5">
                             <div className="flex min-w-0 flex-col gap-1.5">
                                 <div className="flex items-center gap-1.5">
@@ -845,7 +1173,7 @@ export const VoiceSettings: React.FC = () => {
                                         </TooltipTrigger>
                                         <TooltipContent sideOffset={8} className="max-w-xs">
                                             <ul className="space-y-1">
-                                                <li><strong>{t('settings.voice.page.provider.browser')}</strong> {t('settings.voice.page.tooltip.sttBrowser')}</li>
+                                                <li><strong>{t('settings.voice.page.provider.local')}</strong> {t('settings.voice.page.tooltip.sttLocal')}</li>
                                                 <li><strong>{t('settings.voice.page.provider.server')}</strong> {t('settings.voice.page.tooltip.sttServer')}</li>
                                             </ul>
                                         </TooltipContent>
@@ -855,53 +1183,34 @@ export const VoiceSettings: React.FC = () => {
                                     <Button
                                         variant="chip"
                                         size="xs"
-                                        aria-pressed={sttProvider === 'browser'}
+                                        aria-pressed={sttProvider === 'local'}
-                                        onClick={() => setSttProvider('browser')}
+                                        onClick={() => setSttProvider('local')}
                                         className="!font-normal"
                                     >
-                                        {t('settings.voice.page.provider.browser')}
+                                        {t('settings.voice.page.provider.local')}
                                     </Button>
                                     <Button
                                         variant="chip"
                                         size="xs"
-                                        aria-pressed={sttProvider === 'server'}
+                                        aria-pressed={sttProvider === 'openai-compatible'}
-                                        onClick={() => setSttProvider('server')}
+                                        onClick={() => setSttProvider('openai-compatible')}
                                         className="!font-normal"
                                     >
                                         {t('settings.voice.page.provider.server')}
                                     </Button>
-                                    <Button
-                                        variant="chip"
-                                        size="xs"
-                                        aria-pressed={sttProvider === 'wasm'}
-                                        onClick={() => setSttProvider('wasm')}
-                                        className="!font-normal"
-                                    >
-                                        {t('settings.voice.page.provider.wasm')}
-                                    </Button>
                                 </div>
                             </div>
                         </div>
 
-                        {sttProvider === 'server' && (
+                        {sttProvider === 'local' && (
-                            <div className="py-1.5 space-y-2">
+                            <div className="py-1.5">
-                                <div
+                                <span className="typography-ui-label text-foreground">{t('settings.voice.page.field.model')}</span>
-                                    className="group flex cursor-pointer items-center gap-2 py-1.5"
+                                <LocalModelPicker selectedModelId={sttLocalModel} onSelect={setSttLocalModel} />
-                                    role="button"
-                                    tabIndex={0}
-                                    aria-pressed={sttTranscribeOnStop}
-                                    onClick={() => setSttTranscribeOnStop(!sttTranscribeOnStop)}
-                                    onKeyDown={(e) => { if (e.key === ' ' || e.key === 'Enter') { e.preventDefault(); setSttTranscribeOnStop(!sttTranscribeOnStop); } }}
-                                >
-                                    <Checkbox checked={sttTranscribeOnStop} onChange={setSttTranscribeOnStop} ariaLabel={t('settings.voice.page.field.transcribeOnStopAria')} />
-                                    <span className="typography-ui-label text-foreground">{t('settings.voice.page.field.transcribeOnStop')}</span>
                             </div>
-
-                                {!audioStreamService.isSupported() && (
-                                    <p className="typography-meta text-[var(--status-error)]">
-                                        {t('settings.voice.page.field.sttBrowserSupportError')}
-                                    </p>
                         )}
+
+                        {sttProvider === 'openai-compatible' && (
+                            <div className="py-1.5 space-y-2">
                                 <div>
                                     <span className={cn("typography-ui-label text-foreground", !sttServerUrl.trim() && "text-[var(--status-error)]")}>
                                         {t('settings.voice.page.field.serverUrl')}
@@ -979,128 +1288,13 @@ export const VoiceSettings: React.FC = () => {
                                         />
                                     </div>
                                 </div>
-                                <div className="flex items-center gap-8 py-0.5">
-                                    <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.silenceThreshold')}</span>
-                                    <div className="flex items-center gap-2 w-fit">
-                                        {!isMobile && <input type="range" min={-60} max={-20} step={1} value={sttSilenceThresholdDb} onChange={(e) => setSttSilenceThresholdDb(Number(e.target.value))} className={sliderClass} />}
-                                        <span className="typography-ui-label text-foreground tabular-nums min-w-[3.5rem] text-right">
-                                            {sttSilenceThresholdDb} dB
-                                        </span>
-                                    </div>
-                                </div>
-                                <div className="flex items-center gap-8 py-0.5">
-                                    <span className="typography-ui-label text-foreground sm:w-56 shrink-0">{t('settings.voice.page.field.silenceHold')}</span>
-                                    <div className="flex items-center gap-2 w-fit">
-                                        {!isMobile && <input type="range" min={500} max={3000} step={100} value={sttSilenceHoldMs} onChange={(e) => setSttSilenceHoldMs(Number(e.target.value))} className={sliderClass} />}
-                                        <NumberInput value={sttSilenceHoldMs} onValueChange={setSttSilenceHoldMs} min={500} max={3000} step={100} className="w-20 tabular-nums" />
-                                        <span className="typography-meta text-muted-foreground">{t('settings.voice.page.field.millisecondsUnit')}</span>
-                                    </div>
-                                </div>
                             </div>
                         )}
 
-                        {sttProvider === 'wasm' && (
+                        </>)}
-                            <div className="py-1.5 space-y-2">
-                                <div>
-                                    <span className="typography-ui-label text-muted-foreground">
-                                        {t('settings.voice.page.stt.wasmModel')}
-                                    </span>
-                                    <Select value={wasmSttModel} onValueChange={setWasmSttModel}>
-                                        <SelectTrigger className="mt-0.5">
-                                            <SelectValue placeholder={t('settings.voice.page.stt.wasmModel')} />
-                                        </SelectTrigger>
-                                        <SelectContent>
-                                            {WASM_MODELS.map((m) => (
-                                                <SelectItem key={m.id} value={m.id}>
-                                                    <div className="flex flex-col">
-                                                        <span>{m.name}</span>
-                                                        <span className="typography-ui-compact text-muted-foreground">
-                                                            {m.size} · {m.languages}
-                                                        </span>
-                                                    </div>
-                                                </SelectItem>
-                                            ))}
-                                        </SelectContent>
-                                    </Select>
-                                    <p className="typography-ui-compact text-muted-foreground mt-0.5">
-                                        {WASM_MODELS.find((m) => m.id === wasmSttModel)?.description}
-                                    </p>
-                                </div>
-                                {/* Model status indicator */}
-                                <WasmModelStatusIndicator modelId={wasmSttModel} />
-                            </div>
-                        )}
                     </section>
             </div>
-            )}
 
-            {/* Playback & Summarization */}
-            <div data-settings-item="voice.playback" className="mb-8">
-                <div className="mb-1 px-1">
-                    <h3 className="typography-ui-header font-medium text-foreground">
-                        {t('settings.voice.page.section.playbackAndSummary')}
-                    </h3>
-                </div>
-
-                <section className="px-2 pb-2 pt-0 space-y-0">
-                    <div
-                        className="group flex cursor-pointer items-center gap-2 py-1.5"
-                        role="button"
-                        tabIndex={0}
-                        aria-pressed={showMessageTTSButtons}
-                        onClick={() => setShowMessageTTSButtons(!showMessageTTSButtons)}
-                        onKeyDown={(e) => { if (e.key === ' ' || e.key === 'Enter') { e.preventDefault(); setShowMessageTTSButtons(!showMessageTTSButtons); } }}
-                    >
-                        <Checkbox checked={showMessageTTSButtons} onChange={setShowMessageTTSButtons} ariaLabel={t('settings.voice.page.field.messageReadAloudButtonAria')} />
-                        <span className="typography-ui-label text-foreground">{t('settings.voice.page.field.messageReadAloudButton')}</span>
-                    </div>
-
-                    <div className="pb-1.5 pt-0.5">
-                        <div className="flex min-w-0 flex-col gap-1.5">
-                            <div className="flex items-center gap-1.5">
-                                <span className="typography-ui-label text-foreground">
-                                    {t('settings.voice.page.field.ttsInputMode')}
-                                </span>
-                            </div>
-                            <div className="flex flex-wrap items-center gap-1">
-                                <Button
-                                    variant="chip"
-                                    size="xs"
-                                    aria-pressed={ttsInputMode === 'sanitized'}
-                                    onClick={() => setTtsInputMode('sanitized')}
-                                    className="!font-normal"
-                                >
-                                    {t('settings.voice.page.field.ttsInputModeSanitized')}
-                                </Button>
-                                <Button
-                                    variant="chip"
-                                    size="xs"
-                                    aria-pressed={ttsInputMode === 'raw'}
-                                    onClick={() => setTtsInputMode('raw')}
-                                    className="!font-normal"
-                                >
-                                    {t('settings.voice.page.field.ttsInputModeRaw')}
-                                </Button>
-                            </div>
-                        </div>
-                    </div>
-
-                </section>
-
-                {voiceModeEnabled && isSupported && (
-                    <div className="mt-2 px-2">
-                        <p className="typography-meta text-muted-foreground">
-                            {t('settings.voice.page.hint.shiftClickPrefix')}
-                            {' '}
-                            <kbd className="px-1 py-0.5 mx-0.5 rounded border border-[var(--interactive-border)] bg-background typography-mono text-[10px]">Shift</kbd>
-                            {' + '}
-                            <kbd className="px-1 py-0.5 mx-0.5 rounded border border-[var(--interactive-border)] bg-background typography-mono text-[10px]">Click</kbd>
-                            {' '}
-                            {t('settings.voice.page.hint.shiftClickSuffix')}
-                        </p>
-                    </div>
-                )}
-            </div>
 
         </div>
     );

packages/ui/src/components/views/SettingsView.tsx

@@ -1030,7 +1030,7 @@ export const SettingsView: React.FC<SettingsViewProps> = ({ onClose, forceMobile
                         : <Icon name={iconName!} className="h-4 w-4 shrink-0" />}
                       <span className="flex items-center gap-1.5 whitespace-nowrap overflow-hidden transition-opacity duration-150 opacity-100">
                         <span className="typography-ui-label font-normal truncate">{getPageTitle(page.slug)}</span>
-                        {(page.slug === 'voice' || page.slug === 'tunnel') && (
+                        {page.slug === 'tunnel' && (
                           <span className="shrink-0 typography-micro px-1 rounded leading-none pb-px text-[var(--status-warning)] bg-[var(--status-warning)]/10">
                             {t('settings.view.badge.beta')}
                           </span>

packages/ui/src/components/voice/BrowserVoiceButton.tsx

@@ -1,392 +0,0 @@
-/**
- * BrowserVoiceButton Component
- *
- * Voice toggle button for browser-based voice chat with language selection.
- * Shows visual state indicators for different voice modes.
- *
- * @example
- * ```tsx
- * <BrowserVoiceButton />
- * ```
- */
-
-import React, { useCallback, useEffect, useRef, useState } from 'react';
-import { useBrowserVoice } from '@/hooks/useBrowserVoice';
-import { useConfigStore } from '@/stores/useConfigStore';
-import { browserVoiceService } from '@/lib/voice/browserVoiceService';
-import { isVSCodeRuntime } from '@/lib/desktop';
-import { Button } from '@/components/ui/button';
-import {
-    Tooltip,
-    TooltipContent,
-    TooltipProvider,
-    TooltipTrigger,
-} from '@/components/ui/tooltip';
-import { VoiceStatusIndicator } from './VoiceStatusIndicator';
-import { toast } from '@/components/ui/toast';
-import { Icon } from "@/components/icon/Icon";
-import { useI18n } from '@/lib/i18n';
-
-// Status text for accessibility and labels
-const statusLabels: Record<string, string> = {
-    idle: 'Start Voice',
-    listening: 'Listening',
-    processing: 'Processing',
-    speaking: 'AI Speaking',
-    error: 'Voice Error',
-};
-
-// iOS Safari detection utility
-const isIOSSafari = (): boolean => {
-    if (typeof navigator === 'undefined') return false;
-    const userAgent = navigator.userAgent.toLowerCase();
-    const isIOS = /iphone|ipad|ipod/i.test(userAgent);
-    const isSafari = /safari/i.test(userAgent) && !/chrome|crios|crmo/i.test(userAgent);
-    return isIOS && isSafari;
-};
-
-const normalizeVoiceErrorMessage = (error: string): string => {
-    const isMediaDevicesError =
-        error.includes('getUserMedia') ||
-        error.includes('mediaDevices') ||
-        error.includes('Cannot read properties of undefined');
-
-    if (!isMediaDevicesError) {
-        return error;
-    }
-
-    if (typeof window !== 'undefined' && !window.isSecureContext) {
-        return 'Voice requires a secure connection (HTTPS) or localhost. Please use HTTPS or access via localhost.';
-    }
-
-    return 'Microphone access is unavailable in this runtime. On desktop, check System Settings -> Privacy & Security -> Microphone for OpenChamber.';
-};
-
-/**
- * Browser Voice Button with language selection
- */
-export function BrowserVoiceButton() {
-    const { t } = useI18n();
-    const voiceModeEnabled = useConfigStore((s) => s.voiceModeEnabled);
-    const sttProvider = useConfigStore((s) => s.sttProvider);
-    const sttTranscribeOnStop = useConfigStore((s) => s.sttTranscribeOnStop);
-    
-    const {
-        status,
-        isSupported,
-        error,
-
-        startVoice,
-        stopVoice,
-        finishVoiceInput,
-        conversationMode,
-        toggleConversationMode,
-        isMobile,
-    } = useBrowserVoice();
-    
-    const [isPressing, setIsPressing] = useState(false);
-    const isVSCode = isVSCodeRuntime();
-    const buttonSizeClass = isMobile ? 'h-8 w-8 min-h-[32px] min-w-[32px]' : (isVSCode ? 'h-5 w-5' : 'h-6 w-6');
-    const iconSizeClass = isMobile ? 'h-[18px] w-[18px]' : (isVSCode ? 'h-4 w-4' : 'h-[18px] w-[18px]');
-    const continuousIconSizeClass = 'size-[18px]';
-    const clearHoverBackgroundClass = 'bg-transparent hover:!bg-transparent active:!bg-transparent focus-visible:!bg-transparent';
-    
-    // Refs for touch handling
-    const touchHandledRef = useRef(false);
-    const isIOSSafariRef = useRef(false);
-    const longPressTimerRef = useRef<NodeJS.Timeout | null>(null);
-    const longPressTriggeredRef = useRef(false);
-    const lastToastedErrorRef = useRef<string | null>(null);
-    
-    // Initialize iOS detection on mount
-    useEffect(() => {
-        isIOSSafariRef.current = isIOSSafari();
-    }, []);
-
-    // NOTE: Do NOT pre-request microphone permission on mount.
-    // Permission is requested when the user explicitly taps the mic button.
-    // Pre-requesting causes an unwanted permission prompt on mobile page load.
-
-    // Determine active states
-    const isActive = status === 'listening' || status === 'speaking' || status === 'processing';
-    const isError = status === 'error';
-    const isIdle = status === 'idle';
-
-    const isSpeaking = status === 'speaking';
-    // WASM STT always needs finishVoiceInput to flush the recorder and transcribe.
-    // Server STT uses it when sttTranscribeOnStop is enabled.
-    const canTranscribeOnStop = sttProvider === 'wasm' || (sttProvider === 'server' && sttTranscribeOnStop);
-    const isListeningWithTranscribeOnStop = status === 'listening' && canTranscribeOnStop;
-
-    // Show toast notification when voice error occurs
-    useEffect(() => {
-        if (isError && error) {
-            if (lastToastedErrorRef.current === error) {
-                return;
-            }
-            lastToastedErrorRef.current = error;
-            const displayError = normalizeVoiceErrorMessage(error);
-            
-            toast.error(displayError, {
-                duration: 5000,
-            });
-        }
-
-        if (!isError) {
-            lastToastedErrorRef.current = null;
-        }
-    }, [isError, error]);
-
-    // Status text for accessibility
-    const statusText = isError
-        ? error || 'Voice Error'
-        : isListeningWithTranscribeOnStop
-          ? t('voice.action.finishAndTranscribe')
-        : conversationMode && status === 'idle'
-          ? 'Start Voice (Continuous mode on)'
-          : statusLabels[status] || 'Start Voice';
-
-    // Tooltip content based on state
-    const getTooltipContent = () => {
-        if (isError && error) {
-            return normalizeVoiceErrorMessage(error);
-        }
-        if (isListeningWithTranscribeOnStop) {
-            return t('voice.action.finishAndTranscribe');
-        }
-        if (isActive) {
-            return 'Stop voice conversation';
-        }
-        if (isMobile) {
-            return 'Start voice conversation';
-        }
-        return `Start voice conversation (Shift+Click for continuous mode) • Cmd/Ctrl+Shift+V to toggle`;
-    };
-
-    // Handle voice activation (used by both click and touch)
-    const activateVoice = useCallback(async () => {
-        if (isActive) {
-            if (status === 'listening' && canTranscribeOnStop) {
-                finishVoiceInput();
-                return;
-            }
-            stopVoice();
-        } else if (status !== 'error') {
-            // On mobile, we must NOT do any async operations before calling startVoice()
-            // because iOS Safari requires SpeechRecognition.start() to be called
-            // synchronously within the user gesture handler
-            if (isMobile) {
-                // Start voice immediately - no await before this!
-                // Audio unlock is now handled inside startVoice() for mobile
-                startVoice();
-            } else {
-                // Desktop can use async path
-                try {
-                    await startVoice();
-                } catch (err) {
-                    console.error('Failed to start voice:', err);
-                }
-            }
-        } else {
-            // Reset from error state
-            if (isMobile) {
-                startVoice();
-            } else {
-                try {
-                    await startVoice();
-                } catch (err) {
-                    console.error('Failed to start voice:', err);
-                }
-            }
-        }
-    }, [isActive, status, canTranscribeOnStop, finishVoiceInput, startVoice, stopVoice, isMobile]);
-
-    // Handle Shift+Click to toggle conversation mode
-    const handleClick = useCallback(async (e: React.MouseEvent) => {
-        // Prevent double-firing if touch already handled this
-        if (touchHandledRef.current) {
-            touchHandledRef.current = false;
-            return;
-        }
-
-        // Shift+Click toggles conversation mode
-        if (e.shiftKey) {
-            toggleConversationMode();
-            return;
-        }
-
-        await activateVoice();
-    }, [activateVoice, toggleConversationMode]);
-
-    // Handle touch start for mobile devices
-    const handleTouchStart = useCallback((e: React.TouchEvent) => {
-        // Prevent default to stop mouse event emulation
-        e.preventDefault();
-
-        // Mark that touch handled this interaction
-        touchHandledRef.current = true;
-        longPressTriggeredRef.current = false;
-
-        // Immediate visual feedback
-        setIsPressing(true);
-
-        // Set up long-press timer for toggling conversation mode (500ms)
-        longPressTimerRef.current = setTimeout(() => {
-            longPressTriggeredRef.current = true;
-            toggleConversationMode();
-            // Haptic feedback if available
-            if (navigator.vibrate) {
-                navigator.vibrate(50);
-            }
-            setIsPressing(false);
-        }, 500);
-    }, [toggleConversationMode]);
-
-    // Handle touch end
-    const handleTouchEnd = useCallback(() => {
-        // Clear long-press timer
-        if (longPressTimerRef.current) {
-            clearTimeout(longPressTimerRef.current);
-            longPressTimerRef.current = null;
-        }
-
-        // Only activate voice if long-press wasn't triggered
-        if (!longPressTriggeredRef.current) {
-            activateVoice();
-        }
-
-        setIsPressing(false);
-    }, [activateVoice]);
-
-    // Handle touch cancel
-    const handleTouchCancel = useCallback(() => {
-        if (longPressTimerRef.current) {
-            clearTimeout(longPressTimerRef.current);
-            longPressTimerRef.current = null;
-        }
-        setIsPressing(false);
-    }, []);
-
-    const handleToggleConversationMode = useCallback((event: React.MouseEvent<HTMLButtonElement>) => {
-        event.preventDefault();
-        event.stopPropagation();
-        toggleConversationMode();
-    }, [toggleConversationMode]);
-
-    // If voice mode is disabled, don't render anything
-    if (!voiceModeEnabled) {
-        return null;
-    }
-
-    // If not supported, show disabled button with tooltip
-    if (!isSupported) {
-        const supportDetails = browserVoiceService.getSupportDetails();
-        const tooltipMessage = !supportDetails.secureContext
-            ? 'Voice requires HTTPS or localhost. Please use a secure connection.'
-            : !supportDetails.recognition
-                ? 'Speech recognition not supported in this browser. Try Chrome, Edge, or Safari.'
-                : !supportDetails.synthesis
-                    ? 'Speech synthesis not supported in this browser.'
-                    : 'Voice not supported in this browser';
-
-        return (
-            <TooltipProvider>
-                <Tooltip>
-                    <TooltipTrigger asChild>
-                        <Button
-                            size="icon"
-                            variant="ghost"
-                            disabled
-                            aria-label={tooltipMessage}
-                            className={`${buttonSizeClass} p-0 ${clearHoverBackgroundClass}`}
-                        >
-                            <Icon name="mic-off" className={`${iconSizeClass} opacity-50`} />
-                        </Button>
-                    </TooltipTrigger>
-                    <TooltipContent side="bottom" align="center">
-                        <p className="max-w-[200px] text-center">{tooltipMessage}</p>
-                    </TooltipContent>
-                </Tooltip>
-            </TooltipProvider>
-        );
-    }
-
-    return (
-        <div className={`flex items-center ${isMobile ? 'gap-1' : 'gap-1.5'}`}>
-            {/* Status indicator with label - show when active, simplified on mobile */}
-            {isActive && !isMobile && (
-                <VoiceStatusIndicator
-                    status={status}
-                    showLabel
-                    size="sm"
-                    className="mr-1"
-                />
-            )}
-
-            {/* Voice button with tooltip */}
-            <TooltipProvider>
-                <Tooltip>
-                    <TooltipTrigger asChild>
-                        <Button
-                            size="icon"
-                            variant="ghost"
-                            onClick={handleClick}
-                            onTouchStart={handleTouchStart}
-                            onTouchEnd={handleTouchEnd}
-                            onTouchCancel={handleTouchCancel}
-                            aria-label={statusText}
-                            className={`
-                                relative
-                                ${buttonSizeClass}
-                                p-0
-                                ${clearHoverBackgroundClass}
-                                touch-manipulation
-                                ${isPressing ? 'scale-95 opacity-80' : ''}
-                                ${conversationMode && isIdle && isMobile ? 'ring-1 ring-primary/50' : ''}
-                            `}
-                            style={{
-                                WebkitTapHighlightColor: 'transparent',
-                                touchAction: 'manipulation',
-                            }}
-                        >
-                            {isActive ? (
-                                isSpeaking ? (
-                                    // Green speaker icon when AI is speaking
-                                    <Icon name="volume-up" className={`${iconSizeClass} text-green-400 animate-pulse`} />
-                                ) : (
-                                    // Red stop icon for listening/processing (both mobile and desktop)
-                                    <Icon name="stop-circle" className={`${iconSizeClass} text-[var(--status-error)]`} />
-                                )
-                            ) : (
-                                <VoiceStatusIndicator
-                                    status={isError ? 'idle' : status}
-                                    size={isMobile || isVSCode ? 'sm' : 'md'}
-                                />
-                            )}
-                        </Button>
-                    </TooltipTrigger>
-                    <TooltipContent side="bottom" align="center">
-                        <p className="max-w-[200px] text-center">{getTooltipContent()}</p>
-                    </TooltipContent>
-                </Tooltip>
-            </TooltipProvider>
-
-            {/* Conversation mode toggle button */}
-            {(status === 'idle' || status === 'error') && (
-                <Button
-                    size="icon"
-                    variant="ghost"
-                    onPointerDownCapture={(event) => event.stopPropagation()}
-                    onClick={handleToggleConversationMode}
-                    aria-label={conversationMode ? 'Continuous mode on' : 'Continuous mode off'}
-                    title={conversationMode ? 'Continuous mode on' : 'Continuous mode off'}
-                    className={
-                        `${buttonSizeClass} p-0 ${clearHoverBackgroundClass} ${conversationMode ? 'text-[var(--status-info)] hover:text-[var(--status-info)]' : 'text-muted-foreground hover:text-foreground'}`
-                    }
-                >
-                    <Icon name="voice-recognition" className={continuousIconSizeClass} />
-                </Button>
-            )}
-        </div>
-    );
-}

packages/ui/src/components/voice/VoiceProvider.tsx

@@ -1,30 +0,0 @@
-import React from 'react';
-import { useVoiceContext } from '@/hooks/useVoiceContext';
-import { useConfigStore } from '@/stores/useConfigStore';
-
-const VoiceContextBridge = React.memo(function VoiceContextBridge() {
-    useVoiceContext();
-    return null;
-});
-
-/**
- * Provider component that initializes voice context sync.
- * Wrap the app with this to enable voice session awareness.
- * 
- * @example
- * ```tsx
- * <VoiceProvider>
- *   <App />
- * </VoiceProvider>
- * ```
- */
-export function VoiceProvider({ children }: { children: React.ReactNode }) {
-    const voiceModeEnabled = useConfigStore((state) => state.voiceModeEnabled);
-
-    return (
-        <>
-            {voiceModeEnabled ? <VoiceContextBridge /> : null}
-            {children}
-        </>
-    );
-}

packages/ui/src/components/voice/VoiceStatusIndicator.tsx

@@ -1,139 +0,0 @@
-/**
- * VoiceStatusIndicator Component
- *
- * Reusable visual indicator for voice mode states with icons, animations,
- * and optional status text labels.
- *
- * @example
- * ```tsx
- * // Basic usage - icon only
- * <VoiceStatusIndicator status="listening" />
- *
- * // With label
- * <VoiceStatusIndicator status="listening" showLabel />
- *
- * // Different size
- * <VoiceStatusIndicator status="processing" size="lg" />
- * ```
- */
-
-import React from 'react';
-import type { BrowserVoiceStatus } from '@/hooks/useBrowserVoice';
-import { useI18n } from '@/lib/i18n';
-import { Icon } from "@/components/icon/Icon";
-import type { IconName } from "@/components/icon/icons";
-
-export interface VoiceStatusIndicatorProps {
-    /** Current voice status */
-    status: BrowserVoiceStatus;
-    /** Show text label next to icon */
-    showLabel?: boolean;
-    /** Size of the indicator */
-    size?: 'sm' | 'md' | 'lg';
-    /** Optional className for styling */
-    className?: string;
-    /** Whether conversation mode is active (shows indicator dot when idle) */
-    conversationMode?: boolean;
-}
-
-const sizeClasses = {
-    sm: {
-        icon: 'w-4 h-4',
-        container: 'gap-1.5',
-    },
-    md: {
-        icon: 'w-5 h-5',
-        container: 'gap-2',
-    },
-    lg: {
-        icon: 'w-6 h-6',
-        container: 'gap-2.5',
-    },
-};
-
-const statusConfig: Record<
-    BrowserVoiceStatus,
-    {
-        icon: IconName;
-        color: string;
-        labelKey:
-          | 'voice.status.idle'
-          | 'voice.status.listening'
-          | 'voice.status.processing'
-          | 'voice.status.speaking'
-          | 'voice.status.error';
-        animation?: string;
-    }
-> = {
-    idle: {
-        icon: "mic-off",
-        color: 'text-muted-foreground',
-        labelKey: 'voice.status.idle',
-    },
-    listening: {
-        icon: "mic",
-        color: 'text-primary',
-        labelKey: 'voice.status.listening',
-        animation: 'animate-pulse',
-    },
-    processing: {
-        icon: "loader-4",
-        color: 'text-primary',
-        labelKey: 'voice.status.processing',
-        animation: 'animate-spin',
-    },
-    speaking: {
-        icon: "volume-up",
-        color: 'text-green-500',
-        labelKey: 'voice.status.speaking',
-    },
-    error: {
-        icon: "alert",
-        color: 'text-destructive',
-        labelKey: 'voice.status.error',
-    },
-};
-
-/**
- * VoiceStatusIndicator - Visual indicator for voice mode states
- */
-export function VoiceStatusIndicator({
-    status,
-    showLabel = false,
-    size = 'md',
-    className = '',
-    conversationMode = false,
-}: VoiceStatusIndicatorProps) {
-    const { t } = useI18n();
-    const config = statusConfig[status];
-    const statusIconName = config.icon;
-    const sizeClass = sizeClasses[size];
-    const containerClass = showLabel ? sizeClass.container : '';
-
-    return (
-        <div className={`flex items-center ${containerClass} ${className}`}>
-            <div className="relative">
-                <Icon name={statusIconName}
-                    className={`
-                        ${sizeClass.icon}
-                        ${config.color}
-                        ${config.animation || ''}
-                    `}
-                    aria-hidden="true"
-                />
-                {/* Conversation mode indicator dot - only when idle and conversation mode is on */}
-                {conversationMode && status === 'idle' && (
-                    <span
-                        className="absolute -top-0.5 -right-0.5 w-2 h-2 bg-green-500 rounded-full"
-                        aria-label={t('voice.status.conversationModeActiveAria')}
-                    />
-                )}
-            </div>
-            {showLabel && (
-                <span className={`typography-meta ${config.color}`}>
-                    {t(config.labelKey)}
-                </span>
-            )}
-        </div>
-    );
-}

packages/ui/src/components/voice/index.ts

@@ -1,2 +0,0 @@
-export { VoiceProvider } from './VoiceProvider';
-export { BrowserVoiceButton } from './BrowserVoiceButton';

packages/ui/src/hooks/useDictation.ts

@@ -0,0 +1,408 @@
+/**
+ * Streaming dictation state machine.
+ *
+ * Status flow: idle -> recording -> uploading -> idle | failed.
+ * While recording, mic PCM chunks stream to the server, which sends back live
+ * partial transcripts. Confirm finalizes and resolves the full text; failed
+ * dictations retain their audio segments so retry can replay them.
+ */
+
+import { useCallback, useEffect, useRef, useState } from 'react';
+
+import { dictationClient, type DictationStartOptions } from '@/lib/dictation/dictation-client';
+import { DictationStreamSender } from '@/lib/dictation/dictation-stream-sender';
+import { useDictationAudioSource } from '@/lib/dictation/use-dictation-audio-source';
+import { useConfigStore } from '@/stores/useConfigStore';
+
+export type DictationStatus = 'idle' | 'recording' | 'uploading' | 'failed';
+
+export interface UseDictationOptions {
+    onTranscript?: (text: string) => void;
+    onError?: (error: Error) => void;
+    canStart?: () => boolean;
+}
+
+export interface UseDictationResult {
+    status: DictationStatus;
+    isRecording: boolean;
+    isProcessing: boolean;
+    partialTranscript: string;
+    volume: number;
+    duration: number;
+    error: string | null;
+    errorReason: string | null;
+    startDictation: () => Promise<void>;
+    confirmDictation: () => Promise<string | null>;
+    cancelDictation: () => Promise<void>;
+    retryFailedDictation: () => Promise<string | null>;
+    acceptPartialTranscript: () => string | null;
+    discardFailedDictation: () => void;
+}
+
+const DURATION_TICK_MS = 1000;
+
+const toError = (value: unknown): Error =>
+    value instanceof Error ? value : new Error(String(value));
+
+const getDictationStartOptions = (): DictationStartOptions => {
+    const state = useConfigStore.getState();
+    const language = state.sttLanguage?.trim();
+    if (state.sttProvider === 'openai-compatible') {
+        return {
+            provider: 'openai-compatible',
+            ...(language ? { language } : {}),
+            openaiCompatible: {
+                baseUrl: state.sttServerUrl,
+                model: state.sttModel,
+                ...(state.sttApiKey ? { apiKey: state.sttApiKey } : {}),
+            },
+        };
+    }
+    return {
+        provider: 'local',
+        ...(language ? { language } : {}),
+        localModel: state.sttLocalModel,
+    };
+};
+
+export function useDictation(options: UseDictationOptions = {}): UseDictationResult {
+    const { onTranscript, onError, canStart } = options;
+
+    const [status, setStatus] = useState<DictationStatus>('idle');
+    const [partialTranscript, setPartialTranscript] = useState('');
+    const [duration, setDuration] = useState(0);
+    const [error, setError] = useState<string | null>(null);
+    const [errorReason, setErrorReason] = useState<string | null>(null);
+
+    const statusRef = useRef(status);
+    useEffect(() => {
+        statusRef.current = status;
+    }, [status]);
+
+    const latestPartialRef = useRef('');
+    const durationIntervalRef = useRef<ReturnType<typeof setInterval> | null>(null);
+    const actionGateRef = useRef({ starting: false, confirming: false, cancelling: false });
+
+    const onTranscriptRef = useRef(onTranscript);
+    const onErrorRef = useRef(onError);
+    useEffect(() => {
+        onTranscriptRef.current = onTranscript;
+        onErrorRef.current = onError;
+    }, [onTranscript, onError]);
+
+    const senderRef = useRef<DictationStreamSender | null>(null);
+    if (!senderRef.current) {
+        senderRef.current = new DictationStreamSender({
+            client: dictationClient,
+            getStartOptions: getDictationStartOptions,
+        });
+    }
+
+    const stopDurationTracking = useCallback(() => {
+        if (durationIntervalRef.current) {
+            clearInterval(durationIntervalRef.current);
+            durationIntervalRef.current = null;
+        }
+    }, []);
+
+    const startDurationTracking = useCallback(() => {
+        if (durationIntervalRef.current) {
+            return;
+        }
+        durationIntervalRef.current = setInterval(() => {
+            setDuration((prev) => prev + 1);
+        }, DURATION_TICK_MS);
+    }, []);
+
+    const reportError = useCallback((err: unknown) => {
+        const normalized = toError(err);
+        setError(normalized.message);
+        const reason = (normalized as Error & { reasonCode?: string }).reasonCode;
+        setErrorReason(typeof reason === 'string' ? reason : null);
+        onErrorRef.current?.(normalized);
+    }, []);
+
+    const clearError = useCallback(() => {
+        setError(null);
+        setErrorReason(null);
+    }, []);
+
+    const clearStreamingState = useCallback(() => {
+        senderRef.current?.clearAll();
+        latestPartialRef.current = '';
+        setPartialTranscript('');
+    }, []);
+
+    // Live partial transcripts for the active dictation.
+    useEffect(() => {
+        return dictationClient.onPartial((dictationId, text) => {
+            const activeDictationId = senderRef.current?.getDictationId();
+            if (!activeDictationId || dictationId !== activeDictationId) {
+                return;
+            }
+            latestPartialRef.current = text;
+            setPartialTranscript(text);
+        });
+    }, []);
+
+    // Restart the stream (replaying buffered segments) after a reconnect.
+    useEffect(() => {
+        return dictationClient.subscribeConnectionStatus((connected) => {
+            if (!connected) {
+                return;
+            }
+            if (statusRef.current !== 'recording') {
+                return;
+            }
+            void senderRef.current?.restartStream().catch((err) => {
+                reportError(err);
+            });
+        });
+    }, [reportError]);
+
+    const audio = useDictationAudioSource({
+        onPcmSegment: (audioData) => {
+            senderRef.current?.enqueueSegment(audioData);
+        },
+        onError: (err) => {
+            onErrorRef.current?.(err);
+        },
+    });
+    const audioStopRef = useRef(audio.stop);
+    useEffect(() => {
+        audioStopRef.current = audio.stop;
+    }, [audio.stop]);
+
+    const handleSuccess = useCallback(
+        (text: string): string | null => {
+            setDuration(0);
+            setStatus('idle');
+            const transcriptText = text.trim().length > 0 ? text.trim() : latestPartialRef.current.trim();
+            clearStreamingState();
+            if (!transcriptText) {
+                return null;
+            }
+            onTranscriptRef.current?.(transcriptText);
+            return transcriptText;
+        },
+        [clearStreamingState],
+    );
+
+    const handleFailure = useCallback(
+        (failure: unknown) => {
+            if (senderRef.current?.hasSegments()) {
+                setStatus('failed');
+            } else {
+                setStatus('idle');
+            }
+            reportError(failure);
+        },
+        [reportError],
+    );
+
+    const startDictation = useCallback(async () => {
+        const gate = actionGateRef.current;
+        if (gate.starting || gate.confirming || gate.cancelling) {
+            return;
+        }
+        if (statusRef.current !== 'idle') {
+            return;
+        }
+        if (canStart && !canStart()) {
+            return;
+        }
+
+        gate.starting = true;
+        clearError();
+        setPartialTranscript('');
+        setDuration(0);
+        setStatus('recording');
+        statusRef.current = 'recording';
+        clearStreamingState();
+
+        try {
+            await audio.start();
+            startDurationTracking();
+            // Open the stream eagerly so partials start flowing immediately.
+            await senderRef.current?.restartStream().catch((err) => {
+                // Non-fatal: segments buffer locally and finish() retries the
+                // start, but surface the reason (e.g. model downloading) so
+                // the overlay can show it.
+                reportError(err);
+            });
+        } catch (err) {
+            await audio.stop().catch(() => undefined);
+            stopDurationTracking();
+            setStatus('idle');
+            statusRef.current = 'idle';
+            reportError(err);
+        } finally {
+            gate.starting = false;
+        }
+    }, [audio, canStart, clearError, clearStreamingState, reportError, startDurationTracking, stopDurationTracking]);
+
+    const cancelDictation = useCallback(async () => {
+        const gate = actionGateRef.current;
+        if (gate.cancelling) {
+            return;
+        }
+        if (statusRef.current !== 'recording' && statusRef.current !== 'uploading') {
+            return;
+        }
+        gate.cancelling = true;
+        stopDurationTracking();
+        setDuration(0);
+        clearError();
+
+        // Optimistic: dismiss the overlay immediately. Tearing down the audio
+        // graph (AudioContext.close) can take up to ~1s on mobile WebViews and
+        // must not delay the visible response to Cancel.
+        setStatus('idle');
+        statusRef.current = 'idle';
+        try {
+            senderRef.current?.cancel();
+        } catch {
+            // no-op
+        }
+        clearStreamingState();
+
+        try {
+            await audio.stop();
+        } catch {
+            // Cancelled anyway; mic teardown failures are not user-actionable.
+        } finally {
+            gate.cancelling = false;
+        }
+    }, [audio, clearError, clearStreamingState, stopDurationTracking]);
+
+    const confirmDictation = useCallback(async (): Promise<string | null> => {
+        const gate = actionGateRef.current;
+        if (gate.confirming) {
+            return null;
+        }
+        if (statusRef.current !== 'recording') {
+            return null;
+        }
+
+        gate.confirming = true;
+        clearError();
+        stopDurationTracking();
+
+        try {
+            await audio.stop();
+            setStatus('uploading');
+            statusRef.current = 'uploading';
+
+            const finalSeq = senderRef.current?.getFinalSeq() ?? -1;
+            if (finalSeq < 0) {
+                return handleSuccess('');
+            }
+
+            const result = await senderRef.current!.finish(finalSeq);
+            return handleSuccess(result.text);
+        } catch (err) {
+            handleFailure(err);
+            return null;
+        } finally {
+            gate.confirming = false;
+        }
+    }, [audio, clearError, handleFailure, handleSuccess, stopDurationTracking]);
+
+    const retryFailedDictation = useCallback(async (): Promise<string | null> => {
+        if (statusRef.current !== 'failed' || !senderRef.current?.hasSegments()) {
+            return null;
+        }
+        clearError();
+        setStatus('uploading');
+        statusRef.current = 'uploading';
+
+        try {
+            senderRef.current.resetStreamForReplay();
+            const finalSeq = senderRef.current.getFinalSeq();
+            const result = await senderRef.current.finish(finalSeq);
+            return handleSuccess(result.text);
+        } catch (err) {
+            handleFailure(err);
+            return null;
+        }
+    }, [clearError, handleFailure, handleSuccess]);
+
+    /**
+     * Failed dictations still hold the last streamed partial transcript.
+     * Accept it as-is instead of retrying the full transcription.
+     */
+    const acceptPartialTranscript = useCallback((): string | null => {
+        if (statusRef.current !== 'failed') {
+            return null;
+        }
+        const text = latestPartialRef.current.trim();
+        setDuration(0);
+        setStatus('idle');
+        statusRef.current = 'idle';
+        clearError();
+        try {
+            senderRef.current?.cancel();
+        } catch {
+            // no-op
+        }
+        clearStreamingState();
+        if (!text) {
+            return null;
+        }
+        onTranscriptRef.current?.(text);
+        return text;
+    }, [clearError, clearStreamingState]);
+
+    const discardFailedDictation = useCallback(() => {
+        setDuration(0);
+        setStatus('idle');
+        statusRef.current = 'idle';
+        clearError();
+        clearStreamingState();
+    }, [clearError, clearStreamingState]);
+
+    // While recording without an open stream (e.g. the model is still
+    // downloading), retry the stream start so live partials kick in as soon
+    // as the provider becomes ready. Buffered segments replay on success.
+    useEffect(() => {
+        if (status !== 'recording' || errorReason !== 'model_download_in_progress') {
+            return;
+        }
+        const interval = setInterval(() => {
+            if (statusRef.current !== 'recording' || senderRef.current?.getDictationId()) {
+                return;
+            }
+            senderRef.current?.restartStream().then(() => {
+                clearError();
+            }).catch((err) => {
+                reportError(err);
+            });
+        }, 3000);
+        return () => clearInterval(interval);
+    }, [status, errorReason, clearError, reportError]);
+
+    useEffect(() => {
+        return () => {
+            stopDurationTracking();
+            void audioStopRef.current().catch(() => undefined);
+            senderRef.current?.cancel();
+        };
+    }, [stopDurationTracking]);
+
+    return {
+        status,
+        isRecording: status === 'recording',
+        isProcessing: status === 'uploading',
+        partialTranscript,
+        volume: audio.volume,
+        duration,
+        error,
+        errorReason,
+        startDictation,
+        confirmDictation,
+        cancelDictation,
+        retryFailedDictation,
+        acceptPartialTranscript,
+        discardFailedDictation,
+    };
+}

packages/ui/src/hooks/useKeyboardShortcuts.ts

@@ -474,6 +474,18 @@ export const useKeyboardShortcuts = () => {
         return;
       }
 
+      if (eventMatchesShortcut(e, combo('toggle_dictation'))) {
+        const { activeMainTab, isCommandPaletteOpen, isHelpDialogOpen, isSessionSwitcherOpen, isSettingsDialogOpen } = useUIStore.getState();
+        if (activeMainTab !== 'chat' || isCommandPaletteOpen || isHelpDialogOpen || isSessionSwitcherOpen || isSettingsDialogOpen) {
+          return;
+        }
+        e.preventDefault();
+        // Dictation state lives inside the composer's isolated component;
+        // toggle it via an event instead of subscribing this hot hook to it.
+        window.dispatchEvent(new CustomEvent('openchamber:dictation-toggle'));
+        return;
+      }
+
       if (e.key === 'Escape') {
         const target = e.target as Element | null;
         const isInsideDialog = Boolean(target?.closest('[role="dialog"]'));

packages/ui/src/hooks/useLocalTTS.ts

@@ -0,0 +1,266 @@
+/**
+ * useLocalTTS Hook
+ *
+ * React hook for local (Kokoro via sherpa-onnx) text-to-speech playback.
+ * Synthesis runs on the OpenChamber server in the dictation worker.
+ *
+ * Long texts are pipelined by sentence chunks: the first chunk starts playing
+ * as soon as it is synthesized while the next chunk synthesizes in the
+ * background, so time-to-first-audio stays ~1 chunk regardless of message
+ * length.
+ */
+
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { runtimeFetch } from '@/lib/runtime-fetch';
+
+export interface LocalTTSSpeakOptions {
+    /** Kokoro speaker id (0-10) */
+    speakerId?: number;
+    /** Playback speed multiplier (1.0 = normal) */
+    speed?: number;
+    onStart?: () => void;
+    onEnd?: () => void;
+    onError?: (error: string) => void;
+}
+
+export interface UseLocalTTSReturn {
+    isPlaying: boolean;
+    error: string | null;
+    speak: (text: string, options?: LocalTTSSpeakOptions) => Promise<void>;
+    stop: () => void;
+    /** Unlock audio for mobile Safari - call this on user gesture */
+    unlockAudio: () => Promise<void>;
+}
+
+/** Target chunk size: big enough to amortize requests, small enough for low latency. */
+const MIN_CHUNK_CHARS = 60;
+const MAX_CHUNK_CHARS = 400;
+
+/**
+ * Split text into sentence-aligned chunks for pipelined synthesis.
+ * Sentences are merged until MIN_CHUNK_CHARS and hard-split at
+ * MAX_CHUNK_CHARS so a single run-on sentence cannot stall the pipeline.
+ */
+export function splitTextForSynthesis(text: string): string[] {
+    const normalized = text.replace(/\s+/g, ' ').trim();
+    if (!normalized) {
+        return [];
+    }
+
+    const sentences = normalized.match(/[^.!?…]+[.!?…]+["')\]]*\s*|[^.!?…]+$/g) ?? [normalized];
+
+    const chunks: string[] = [];
+    let current = '';
+    for (const sentence of sentences) {
+        for (let offset = 0; offset < sentence.length; offset += MAX_CHUNK_CHARS) {
+            const piece = sentence.slice(offset, offset + MAX_CHUNK_CHARS);
+            if (current && current.length + piece.length > MAX_CHUNK_CHARS) {
+                chunks.push(current.trim());
+                current = '';
+            }
+            current += piece;
+            if (current.length >= MIN_CHUNK_CHARS && /[.!?…]["')\]]*\s*$/.test(current)) {
+                chunks.push(current.trim());
+                current = '';
+            }
+        }
+    }
+    if (current.trim()) {
+        chunks.push(current.trim());
+    }
+
+    // Latency trim for the FIRST chunk only: if the opening sentence is long,
+    // split it at a clause boundary (comma/semicolon/colon) so the first
+    // audio starts sooner. Only the first chunk gets this treatment — comma
+    // splits inside every sentence would chop the prosody, and after playback
+    // starts the pipeline hides synthesis time anyway.
+    if (chunks.length > 0 && chunks[0].length > 120) {
+        const first = chunks[0];
+        const clauseBreak = /[,;:]\s/g;
+        let bestBreak = -1;
+        let match: RegExpExecArray | null;
+        while ((match = clauseBreak.exec(first)) !== null) {
+            const end = match.index + 1;
+            if (end < 40) {
+                continue;
+            }
+            if (end > 160) {
+                break;
+            }
+            bestBreak = end;
+            break;
+        }
+        if (bestBreak > 0) {
+            const head = first.slice(0, bestBreak).trim();
+            const tail = first.slice(bestBreak).trim();
+            chunks.splice(0, 1, head, tail);
+        }
+    }
+
+    return chunks;
+}
+
+let sharedAudioContext: AudioContext | null = null;
+
+function getAudioContext(): AudioContext {
+    if (!sharedAudioContext) {
+        sharedAudioContext = new (window.AudioContext || (window as unknown as { webkitAudioContext: typeof AudioContext }).webkitAudioContext)();
+    }
+    return sharedAudioContext;
+}
+
+interface PlaybackSession {
+    cancelled: boolean;
+    abort: AbortController;
+}
+
+export function useLocalTTS(): UseLocalTTSReturn {
+    const [isPlaying, setIsPlaying] = useState(false);
+    const [error, setError] = useState<string | null>(null);
+
+    const audioSourceRef = useRef<AudioBufferSourceNode | null>(null);
+    const sessionRef = useRef<PlaybackSession | null>(null);
+
+    const unlockAudio = useCallback(async (): Promise<void> => {
+        try {
+            const ctx = getAudioContext();
+            if (ctx.state === 'suspended') {
+                await ctx.resume();
+            }
+            const buffer = ctx.createBuffer(1, 1, 22050);
+            const source = ctx.createBufferSource();
+            source.buffer = buffer;
+            source.connect(ctx.destination);
+            source.start(0);
+        } catch {
+            // Unlocking is best-effort.
+        }
+    }, []);
+
+    const stop = useCallback(() => {
+        const session = sessionRef.current;
+        if (session) {
+            session.cancelled = true;
+            session.abort.abort();
+            sessionRef.current = null;
+        }
+        if (audioSourceRef.current) {
+            try {
+                audioSourceRef.current.onended = null;
+                audioSourceRef.current.stop();
+            } catch {
+                // Already stopped
+            }
+            audioSourceRef.current = null;
+        }
+        setIsPlaying(false);
+    }, []);
+
+    const speak = useCallback(async (text: string, options?: LocalTTSSpeakOptions): Promise<void> => {
+        stop();
+
+        const chunks = splitTextForSynthesis(text);
+        if (chunks.length === 0) {
+            setError('No text to speak');
+            options?.onError?.('No text to speak');
+            return;
+        }
+
+        setError(null);
+
+        const session: PlaybackSession = { cancelled: false, abort: new AbortController() };
+        sessionRef.current = session;
+
+        const fetchChunk = async (chunk: string): Promise<ArrayBuffer> => {
+            const response = await runtimeFetch('/api/dictation/tts/speak', {
+                method: 'POST',
+                headers: { 'Content-Type': 'application/json' },
+                body: JSON.stringify({
+                    text: chunk,
+                    ...(typeof options?.speakerId === 'number' ? { speakerId: options.speakerId } : {}),
+                    ...(typeof options?.speed === 'number' ? { speed: options.speed } : {}),
+                }),
+                signal: session.abort.signal,
+            });
+            if (!response.ok) {
+                const errorData = await response.json().catch(() => ({ error: 'Unknown error' }));
+                throw new Error(errorData.error || `HTTP ${response.status}`);
+            }
+            return (await response.blob()).arrayBuffer();
+        };
+
+        const playBuffer = async (arrayBuffer: ArrayBuffer): Promise<void> => {
+            const ctx = getAudioContext();
+            if (ctx.state === 'suspended') {
+                await ctx.resume();
+            }
+            const audioBuffer = await ctx.decodeAudioData(arrayBuffer);
+            if (session.cancelled) {
+                return;
+            }
+            await new Promise<void>((resolve) => {
+                const source = ctx.createBufferSource();
+                source.buffer = audioBuffer;
+                source.connect(ctx.destination);
+                audioSourceRef.current = source;
+                source.onended = () => {
+                    if (audioSourceRef.current === source) {
+                        audioSourceRef.current = null;
+                    }
+                    resolve();
+                };
+                source.start(0);
+            });
+        };
+
+        try {
+            setIsPlaying(true);
+            options?.onStart?.();
+
+            // Pipelined: synthesize chunk N+1 while chunk N is playing.
+            let nextFetch: Promise<ArrayBuffer> = fetchChunk(chunks[0]);
+            for (let i = 0; i < chunks.length; i += 1) {
+                const buffer = await nextFetch;
+                if (session.cancelled) {
+                    return;
+                }
+                if (i + 1 < chunks.length) {
+                    nextFetch = fetchChunk(chunks[i + 1]);
+                }
+                await playBuffer(buffer);
+                if (session.cancelled) {
+                    return;
+                }
+            }
+
+            setIsPlaying(false);
+            options?.onEnd?.();
+        } catch (err) {
+            if ((err as Error).name === 'AbortError' || session.cancelled) {
+                return;
+            }
+            const errorMsg = err instanceof Error ? err.message : 'Failed to speak';
+            setError(errorMsg);
+            options?.onError?.(errorMsg);
+            setIsPlaying(false);
+        } finally {
+            if (sessionRef.current === session) {
+                sessionRef.current = null;
+            }
+        }
+    }, [stop]);
+
+    useEffect(() => {
+        return () => {
+            stop();
+        };
+    }, [stop]);
+
+    return {
+        isPlaying,
+        error,
+        speak,
+        stop,
+        unlockAudio,
+    };
+}

packages/ui/src/hooks/useMessageTTS.ts

@@ -9,6 +9,7 @@ import { useCallback, useState } from 'react';
 import { useConfigStore } from '@/stores/useConfigStore';
 import { useServerTTS } from './useServerTTS';
 import { useSayTTS } from './useSayTTS';
+import { useLocalTTS } from './useLocalTTS';
 import { browserVoiceService } from '@/lib/voice/browserVoiceService';
 import { sanitizeForTTS } from '@/lib/voice/summarize';
 
@@ -29,6 +30,7 @@ export function useMessageTTS(): UseMessageTTSReturn {
     const speechPitch = useConfigStore((state) => state.speechPitch);
     const speechVolume = useConfigStore((state) => state.speechVolume);
     const sayVoice = useConfigStore((state) => state.sayVoice);
+    const localTtsVoiceId = useConfigStore((state) => state.localTtsVoiceId);
     const browserVoice = useConfigStore((state) => state.browserVoice);
     const openaiVoice = useConfigStore((state) => state.openaiVoice);
     const openaiCompatibleVoice = useConfigStore((state) => state.openaiCompatibleVoice);
@@ -48,13 +50,15 @@ export function useMessageTTS(): UseMessageTTSReturn {
     const { speak: speakSayTTS, stop: stopSayTTS, isAvailable: isSayTTSAvailable } = useSayTTS({
         enabled: shouldCheckSayAvailability,
     });
+    const { speak: speakLocalTTS, stop: stopLocalTTS } = useLocalTTS();
     
     const stop = useCallback(() => {
         setIsPlaying(false);
         stopServerTTS();
         stopSayTTS();
+        stopLocalTTS();
         browserVoiceService.cancelSpeech();
-    }, [stopServerTTS, stopSayTTS]);
+    }, [stopServerTTS, stopSayTTS, stopLocalTTS]);
     
     const play = useCallback(async (text: string) => {
         if (!text.trim()) return;
@@ -84,6 +88,13 @@ export function useMessageTTS(): UseMessageTTSReturn {
                     onEnd: () => setIsPlaying(false),
                     onError: () => setIsPlaying(false),
                 });
+            } else if (voiceProvider === 'local') {
+                await speakLocalTTS(sanitizedText, {
+                    speakerId: localTtsVoiceId,
+                    speed: speechRate,
+                    onEnd: () => setIsPlaying(false),
+                    onError: () => setIsPlaying(false),
+                });
             } else if (voiceProvider === 'say' && isSayTTSAvailable) {
                 const wordsPerMinute = Math.round(100 + (speechRate - 0.5) * 200);
                 await speakSayTTS(sanitizedText, {
@@ -129,6 +140,8 @@ export function useMessageTTS(): UseMessageTTSReturn {
         ttsInputMode,
         speakServerTTS,
         speakSayTTS,
+        speakLocalTTS,
+        localTtsVoiceId,
         stop,
     ]);
     

packages/ui/src/hooks/useVoiceContext.ts

@@ -1,57 +0,0 @@
-import { useEffect, useRef } from 'react';
-import { useSessionUIStore } from '@/sync/session-ui-store';
-import { useSessionPermissions, useSessionTextMessages } from '@/sync/sync-context';
-import { voiceHooks, isVoiceSessionStarted } from '@/lib/voice';
-
-/**
- * Hook that syncs session events (messages, permissions) to the voice agent.
- * Call this inside VoiceProvider to enable session awareness during voice.
- */
-export function useVoiceContext() {
-    const currentSessionId = useSessionUIStore((s) => s.currentSessionId);
-    const messages = useSessionTextMessages(currentSessionId ?? '');
-    const permissions = useSessionPermissions(currentSessionId ?? '');
-
-    // Track last seen message count to only forward new messages
-    const lastMessageCountRef = useRef(0);
-
-    // Forward new messages to voice agent
-    useEffect(() => {
-        if (!currentSessionId || !messages || messages.length === 0 || !isVoiceSessionStarted()) return;
-
-        const currentCount = messages.length;
-        if (currentCount <= lastMessageCountRef.current) return;
-
-        // Get only new messages (messages since last check)
-        const newMessages = messages.slice(lastMessageCountRef.current);
-        lastMessageCountRef.current = currentCount;
-
-        const formattedMessages = newMessages.map(m => ({
-            role: m.role ?? '',
-            content: m.text,
-        }));
-
-        voiceHooks.onMessages(currentSessionId, formattedMessages);
-    }, [currentSessionId, messages]);
-
-    // Forward permission requests to voice agent
-    useEffect(() => {
-        if (!currentSessionId || !permissions || permissions.length === 0) return;
-        if (!isVoiceSessionStarted()) return;
-
-        const request = permissions[0];
-        if (!request) return;
-
-        voiceHooks.onPermissionRequested(
-            currentSessionId,
-            request.id,
-            request.permission,
-            request.metadata
-        );
-    }, [currentSessionId, permissions]);
-
-    // Reset message count when session changes
-    useEffect(() => {
-        lastMessageCountRef.current = 0;
-    }, [currentSessionId]);
-}

packages/ui/src/lib/desktop.ts

@@ -178,14 +178,12 @@ export type DesktopSettings = {
   responseStyleEnabled?: boolean;
   responseStylePreset?: 'concise' | 'detailed' | 'mentor' | 'pushback' | 'noFiller' | 'matchEnergy' | 'warmPeer' | 'custom';
   responseStyleCustomInstructions?: string;
-  sttProvider?: 'browser' | 'server' | 'wasm';
+  dictationEnabled?: boolean;
+  sttProvider?: 'local' | 'openai-compatible';
   sttServerUrl?: string;
   sttModel?: string;
-  wasmSttModel?: string;
+  sttLocalModel?: string;
   sttLanguage?: string;
-  sttSilenceThresholdDb?: number;
-  sttSilenceHoldMs?: number;
-  sttTranscribeOnStop?: boolean;
   // Global draft welcome starters (pinned commands/skills), persisted to settings.json
   draftStarters?: DraftStarterRef[];
 };

packages/ui/src/lib/dictation/dictation-client.ts

@@ -0,0 +1,447 @@
+/**
+ * WebSocket client for the OpenChamber dictation endpoint (/api/dictation/ws).
+ *
+ * One shared client per app. The socket is opened lazily when a dictation
+ * starts and closed after an idle delay. URLs are resolved at connect time via
+ * the runtime URL resolver so runtime switches never leak a stale endpoint.
+ */
+
+import { getRuntimeUrlResolver } from '@/lib/runtime-url';
+import { refreshRuntimeUrlAuthToken } from '@/lib/runtime-auth';
+
+export interface DictationStartOptions {
+    provider?: 'local' | 'openai-compatible';
+    language?: string;
+    localModel?: string;
+    openaiCompatible?: {
+        baseUrl?: string;
+        model?: string;
+        apiKey?: string;
+    };
+}
+
+interface DictationServerMessage {
+    type: string;
+    dictationId?: string;
+    ackSeq?: number;
+    text?: string;
+    timeoutMs?: number;
+    error?: string;
+    retryable?: boolean;
+    reasonCode?: string;
+}
+
+interface DictationStreamError extends Error {
+    retryable: boolean;
+    reasonCode?: string;
+}
+
+const createStreamError = (message: string, retryable: boolean, reasonCode?: string): DictationStreamError => {
+    const error = new Error(message) as DictationStreamError;
+    error.name = 'DictationStreamError';
+    error.retryable = retryable;
+    if (reasonCode) {
+        error.reasonCode = reasonCode;
+    }
+    return error;
+};
+
+const CONNECT_TIMEOUT_MS = 10000;
+const START_TIMEOUT_MS = 15000;
+const IDLE_CLOSE_DELAY_MS = 30000;
+const DEFAULT_FINISH_TIMEOUT_MS = 30000;
+
+type ConnectionStatusListener = (connected: boolean) => void;
+type PartialListener = (dictationId: string, text: string) => void;
+
+interface PendingStart {
+    resolve: () => void;
+    reject: (error: Error) => void;
+    timeout: ReturnType<typeof setTimeout>;
+}
+
+interface PendingFinish {
+    resolve: (result: { text: string }) => void;
+    reject: (error: Error) => void;
+    timeout: ReturnType<typeof setTimeout> | null;
+}
+
+export class DictationClient {
+    private socket: WebSocket | null = null;
+    private connectPromise: Promise<void> | null = null;
+    private idleCloseTimer: ReturnType<typeof setTimeout> | null = null;
+    private readonly pendingStarts = new Map<string, PendingStart>();
+    private readonly pendingFinishes = new Map<string, PendingFinish>();
+    private readonly connectionListeners = new Set<ConnectionStatusListener>();
+    private readonly partialListeners = new Set<PartialListener>();
+    private activeDictations = 0;
+
+    get isConnected(): boolean {
+        return this.socket?.readyState === WebSocket.OPEN;
+    }
+
+    subscribeConnectionStatus(listener: ConnectionStatusListener): () => void {
+        this.connectionListeners.add(listener);
+        return () => {
+            this.connectionListeners.delete(listener);
+        };
+    }
+
+    onPartial(listener: PartialListener): () => void {
+        this.partialListeners.add(listener);
+        return () => {
+            this.partialListeners.delete(listener);
+        };
+    }
+
+    async ensureConnected(): Promise<void> {
+        if (this.isConnected) {
+            return;
+        }
+        if (this.connectPromise) {
+            await this.connectPromise;
+            return;
+        }
+
+        // A WebSocket upgrade can't carry an Authorization header, so it
+        // authenticates via the oc_url_token query param. Mint/await a valid
+        // token BEFORE connecting — the sync getter returns "" while the token
+        // is unminted or inside its expiry skew, and the server would reject
+        // the upgrade with 401.
+        try {
+            await refreshRuntimeUrlAuthToken();
+        } catch {
+            // No auth configured (local runtime) — proceed without a token.
+        }
+
+        this.connectPromise = new Promise<void>((resolve, reject) => {
+            let settled = false;
+            let socket: WebSocket;
+            try {
+                const url = getRuntimeUrlResolver().websocket('/api/dictation/ws');
+                socket = new WebSocket(url);
+            } catch (error) {
+                this.connectPromise = null;
+                reject(error instanceof Error ? error : new Error(String(error)));
+                return;
+            }
+
+            const timeout = setTimeout(() => {
+                if (!settled) {
+                    settled = true;
+                    this.connectPromise = null;
+                    try {
+                        socket.close();
+                    } catch {
+                        // ignore
+                    }
+                    reject(new Error('Dictation connection timed out'));
+                }
+            }, CONNECT_TIMEOUT_MS);
+
+            socket.onopen = () => {
+                // Wait for the server 'ready' frame before resolving.
+            };
+
+            socket.onmessage = (event) => {
+                let message: DictationServerMessage;
+                try {
+                    message = JSON.parse(String(event.data));
+                } catch {
+                    return;
+                }
+                if (!settled && message.type === 'ready') {
+                    settled = true;
+                    clearTimeout(timeout);
+                    this.socket = socket;
+                    this.connectPromise = null;
+                    this.notifyConnection(true);
+                    resolve();
+                    return;
+                }
+                this.handleMessage(message);
+            };
+
+            socket.onerror = () => {
+                if (!settled) {
+                    settled = true;
+                    clearTimeout(timeout);
+                    this.connectPromise = null;
+                    reject(new Error('Dictation connection failed'));
+                }
+            };
+
+            socket.onclose = () => {
+                if (!settled) {
+                    settled = true;
+                    clearTimeout(timeout);
+                    this.connectPromise = null;
+                    reject(new Error('Dictation connection closed'));
+                    return;
+                }
+                if (this.socket === socket) {
+                    this.socket = null;
+                    this.rejectAllPending(new Error('Dictation connection lost'));
+                    this.notifyConnection(false);
+                }
+            };
+        });
+
+        await this.connectPromise;
+    }
+
+    /**
+     * Start a dictation stream. Resolves once the server acks the stream.
+     */
+    async startDictationStream(
+        dictationId: string,
+        format: string,
+        options: DictationStartOptions,
+    ): Promise<void> {
+        await this.ensureConnected();
+        this.activeDictations += 1;
+        this.clearIdleCloseTimer();
+
+        return new Promise<void>((resolve, reject) => {
+            const timeout = setTimeout(() => {
+                this.pendingStarts.delete(dictationId);
+                this.releaseDictation();
+                reject(new Error('Dictation start timed out'));
+            }, START_TIMEOUT_MS);
+
+            this.pendingStarts.set(dictationId, {
+                resolve: () => {
+                    clearTimeout(timeout);
+                    resolve();
+                },
+                reject: (error) => {
+                    clearTimeout(timeout);
+                    this.releaseDictation();
+                    reject(error);
+                },
+            timeout });
+
+            if (!this.send({ type: 'start', dictationId, format, options })) {
+                clearTimeout(timeout);
+                this.pendingStarts.delete(dictationId);
+                this.releaseDictation();
+                reject(new Error('Dictation connection lost'));
+            }
+        });
+    }
+
+    sendDictationStreamChunk(dictationId: string, seq: number, audioBase64: string): boolean {
+        return this.send({ type: 'chunk', dictationId, seq, audio: audioBase64 });
+    }
+
+    /**
+     * Finish a dictation stream. Resolves with the final transcript.
+     */
+    finishDictationStream(dictationId: string, finalSeq: number): Promise<{ text: string }> {
+        return new Promise<{ text: string }>((resolve, reject) => {
+            const pending: PendingFinish = {
+                resolve: (result) => {
+                    if (pending.timeout) {
+                        clearTimeout(pending.timeout);
+                    }
+                    this.releaseDictation();
+                    resolve(result);
+                },
+                reject: (error) => {
+                    if (pending.timeout) {
+                        clearTimeout(pending.timeout);
+                    }
+                    this.releaseDictation();
+                    reject(error);
+                },
+                timeout: null,
+            };
+            pending.timeout = setTimeout(() => {
+                this.pendingFinishes.delete(dictationId);
+                this.releaseDictation();
+                reject(new Error('Timed out waiting for transcription'));
+            }, DEFAULT_FINISH_TIMEOUT_MS);
+
+            this.pendingFinishes.set(dictationId, pending);
+
+            if (!this.send({ type: 'finish', dictationId, finalSeq })) {
+                this.pendingFinishes.delete(dictationId);
+                pending.reject(new Error('Dictation connection lost'));
+            }
+        });
+    }
+
+    cancelDictationStream(dictationId: string): void {
+        this.send({ type: 'cancel', dictationId });
+        const start = this.pendingStarts.get(dictationId);
+        if (start) {
+            this.pendingStarts.delete(dictationId);
+            start.reject(new Error('Dictation cancelled'));
+        }
+        const finish = this.pendingFinishes.get(dictationId);
+        if (finish) {
+            this.pendingFinishes.delete(dictationId);
+            finish.reject(new Error('Dictation cancelled'));
+        }
+        this.releaseDictation();
+        this.scheduleIdleCloseIfReady();
+    }
+
+    private handleMessage(message: DictationServerMessage): void {
+        const dictationId = message.dictationId;
+        if (!dictationId) {
+            return;
+        }
+
+        switch (message.type) {
+            case 'ack': {
+                const pendingStart = this.pendingStarts.get(dictationId);
+                if (pendingStart) {
+                    this.pendingStarts.delete(dictationId);
+                    pendingStart.resolve();
+                }
+                return;
+            }
+            case 'partial': {
+                for (const listener of this.partialListeners) {
+                    listener(dictationId, message.text ?? '');
+                }
+                return;
+            }
+            case 'finish_accepted': {
+                const pendingFinish = this.pendingFinishes.get(dictationId);
+                if (pendingFinish && typeof message.timeoutMs === 'number') {
+                    if (pendingFinish.timeout) {
+                        clearTimeout(pendingFinish.timeout);
+                    }
+                    pendingFinish.timeout = setTimeout(() => {
+                        this.pendingFinishes.delete(dictationId);
+                        pendingFinish.reject(new Error('Timed out waiting for transcription'));
+                    }, message.timeoutMs + 5000);
+                }
+                return;
+            }
+            case 'final': {
+                const pendingFinish = this.pendingFinishes.get(dictationId);
+                if (pendingFinish) {
+                    this.pendingFinishes.delete(dictationId);
+                    pendingFinish.resolve({ text: message.text ?? '' });
+                }
+                this.scheduleIdleCloseIfReady();
+                return;
+            }
+            case 'error': {
+                const error = createStreamError(
+                    message.error || 'Dictation failed',
+                    message.retryable !== false,
+                    message.reasonCode,
+                );
+                const pendingStart = this.pendingStarts.get(dictationId);
+                if (pendingStart) {
+                    this.pendingStarts.delete(dictationId);
+                    pendingStart.reject(error);
+                }
+                const pendingFinish = this.pendingFinishes.get(dictationId);
+                if (pendingFinish) {
+                    this.pendingFinishes.delete(dictationId);
+                    pendingFinish.reject(error);
+                }
+                this.scheduleIdleCloseIfReady();
+                return;
+            }
+            default:
+        }
+    }
+
+    private send(message: object): boolean {
+        if (!this.isConnected || !this.socket) {
+            return false;
+        }
+        try {
+            this.socket.send(JSON.stringify(message));
+            return true;
+        } catch {
+            return false;
+        }
+    }
+
+    private notifyConnection(connected: boolean): void {
+        for (const listener of this.connectionListeners) {
+            listener(connected);
+        }
+    }
+
+    private rejectAllPending(error: Error): void {
+        for (const [dictationId, pending] of this.pendingStarts) {
+            this.pendingStarts.delete(dictationId);
+            pending.reject(error);
+        }
+        for (const [dictationId, pending] of this.pendingFinishes) {
+            this.pendingFinishes.delete(dictationId);
+            pending.reject(error);
+        }
+        this.activeDictations = 0;
+    }
+
+    private releaseDictation(): void {
+        this.activeDictations = Math.max(0, this.activeDictations - 1);
+        this.scheduleIdleCloseIfReady();
+    }
+
+    private scheduleIdleCloseIfReady(): void {
+        if (this.activeDictations > 0 || this.pendingFinishes.size > 0 || this.pendingStarts.size > 0) {
+            return;
+        }
+        this.clearIdleCloseTimer();
+        this.idleCloseTimer = setTimeout(() => {
+            if (this.activeDictations === 0 && this.pendingFinishes.size === 0 && this.pendingStarts.size === 0) {
+                const socket = this.socket;
+                this.socket = null;
+                if (socket) {
+                    try {
+                        socket.close(1000, 'idle');
+                    } catch {
+                        // ignore
+                    }
+                    this.notifyConnection(false);
+                }
+            }
+        }, IDLE_CLOSE_DELAY_MS);
+    }
+
+    /**
+     * Runtime switch: close the socket and fail all in-flight dictations so
+     * nothing keeps streaming to the previous runtime.
+     */
+    cancelAllForRuntimeSwitch(): void {
+        this.clearIdleCloseTimer();
+        const socket = this.socket;
+        this.socket = null;
+        this.connectPromise = null;
+        this.rejectAllPending(new Error('Runtime changed'));
+        if (socket) {
+            try {
+                socket.close(1000, 'runtime switch');
+            } catch {
+                // ignore
+            }
+            this.notifyConnection(false);
+        }
+    }
+
+    private clearIdleCloseTimer(): void {
+        if (this.idleCloseTimer) {
+            clearTimeout(this.idleCloseTimer);
+            this.idleCloseTimer = null;
+        }
+    }
+}
+
+export const dictationClient = new DictationClient();
+
+if (typeof window !== 'undefined') {
+    window.addEventListener('openchamber:runtime-endpoint-changed', () => {
+        // Drop the socket so the next dictation reconnects to the new runtime.
+        dictationClient.cancelAllForRuntimeSwitch();
+    });
+}

packages/ui/src/lib/dictation/dictation-stream-sender.ts

@@ -0,0 +1,240 @@
+/**
+ * Small, non-React state machine for dictation streaming.
+ *
+ * Responsibilities:
+ * - Maintain an ordered buffer of base64 PCM segments
+ * - Start/restart a dictation stream (dictationId)
+ * - Send missing segments (seq) when connected
+ * - Finish/cancel the stream
+ *
+ * Segments are retained until the dictation completes, which enables replay
+ * after a connection drop (`resetStreamForReplay()` + `finish()`).
+ */
+
+import type { DictationClient, DictationStartOptions } from './dictation-client';
+
+const MAX_CHUNKS_PER_FLUSH_TURN = 128;
+
+const PCM_DICTATION_FORMAT = 'audio/pcm;rate=16000;bits=16';
+
+const waitForNextFlushTurn = (): Promise<void> => new Promise((resolve) => setTimeout(resolve, 0));
+
+const createDictationIdDefault = (): string => {
+    const rand = Math.random().toString(36).slice(2, 10);
+    return `dic_${Date.now().toString(16)}${rand}`;
+};
+
+export interface DictationFinishResult {
+    dictationId: string;
+    text: string;
+}
+
+export class DictationStreamSender {
+    private readonly client: DictationClient;
+    private readonly format: string;
+    private readonly createDictationId: () => string;
+    private getStartOptions: () => DictationStartOptions;
+
+    private dictationId: string | null = null;
+    private sendSeq = 0;
+    private segments: string[] = [];
+    private streamReady = false;
+    private flushTimer: ReturnType<typeof setTimeout> | null = null;
+    private drainWaiters: Array<() => void> = [];
+
+    private startGeneration = 0;
+    private startPromise: Promise<void> | null = null;
+
+    constructor(params: {
+        client: DictationClient;
+        getStartOptions: () => DictationStartOptions;
+        format?: string;
+        createDictationId?: () => string;
+    }) {
+        this.client = params.client;
+        this.format = params.format ?? PCM_DICTATION_FORMAT;
+        this.getStartOptions = params.getStartOptions;
+        this.createDictationId = params.createDictationId ?? createDictationIdDefault;
+    }
+
+    getDictationId(): string | null {
+        return this.dictationId;
+    }
+
+    getFinalSeq(): number {
+        return this.segments.length - 1;
+    }
+
+    hasSegments(): boolean {
+        return this.segments.length > 0;
+    }
+
+    clearAll(): void {
+        this.clearScheduledFlush();
+        this.dictationId = null;
+        this.sendSeq = 0;
+        this.segments = [];
+        this.streamReady = false;
+        this.startPromise = null;
+        this.startGeneration += 1;
+    }
+
+    resetStreamForReplay(): void {
+        this.clearScheduledFlush();
+        this.dictationId = null;
+        this.sendSeq = 0;
+        this.streamReady = false;
+        this.startPromise = null;
+        this.startGeneration += 1;
+    }
+
+    enqueueSegment(base64Pcm: string): void {
+        this.segments.push(base64Pcm);
+
+        if (!this.client.isConnected) {
+            return;
+        }
+
+        if (!this.dictationId) {
+            if (!this.startPromise) {
+                void this.restartStream().catch(() => {
+                    // Start failures surface through finish(); segments are retained.
+                });
+            }
+            return;
+        }
+
+        this.flush();
+    }
+
+    flush(): number {
+        const dictationId = this.dictationId;
+        if (!this.client.isConnected || !dictationId || !this.streamReady) {
+            return 0;
+        }
+
+        let sent = 0;
+        while (this.sendSeq < this.segments.length && sent < MAX_CHUNKS_PER_FLUSH_TURN) {
+            const seq = this.sendSeq;
+            const audio = this.segments[seq];
+            if (!this.client.sendDictationStreamChunk(dictationId, seq, audio)) {
+                break;
+            }
+            this.sendSeq = seq + 1;
+            sent += 1;
+        }
+        if (this.hasPendingSegments()) {
+            this.scheduleFlush();
+        } else {
+            this.resolveDrainWaiters();
+        }
+        return sent;
+    }
+
+    async restartStream(): Promise<void> {
+        this.startGeneration += 1;
+        const generation = this.startGeneration;
+
+        const dictationId = this.createDictationId();
+        this.dictationId = dictationId;
+        this.sendSeq = 0;
+        this.streamReady = false;
+
+        const start = (async () => {
+            await this.client.startDictationStream(dictationId, this.format, this.getStartOptions());
+            if (this.startGeneration !== generation) {
+                return;
+            }
+            if (this.dictationId !== dictationId) {
+                return;
+            }
+            this.streamReady = true;
+            this.flush();
+        })()
+            .catch((error) => {
+                // Keep segments for retry, but clear the stream so finish can error cleanly.
+                if (this.startGeneration === generation && this.dictationId === dictationId) {
+                    this.dictationId = null;
+                    this.streamReady = false;
+                }
+                throw error;
+            })
+            .finally(() => {
+                if (this.startPromise === start) {
+                    this.startPromise = null;
+                }
+            });
+
+        this.startPromise = start;
+        await start;
+    }
+
+    async finish(finalSeq: number): Promise<DictationFinishResult> {
+        if (!this.dictationId) {
+            await this.restartStream();
+        }
+        if (this.startPromise) {
+            await this.startPromise;
+        }
+
+        const dictationId = this.dictationId;
+        if (!dictationId || !this.streamReady) {
+            throw new Error('Failed to start dictation stream');
+        }
+
+        this.flush();
+        await this.waitForFlushDrain();
+        const result = await this.client.finishDictationStream(dictationId, finalSeq);
+        return { dictationId, text: result.text };
+    }
+
+    cancel(): void {
+        const dictationId = this.dictationId;
+        if (this.client.isConnected && dictationId) {
+            this.client.cancelDictationStream(dictationId);
+        }
+        this.resetStreamForReplay();
+    }
+
+    private hasPendingSegments(): boolean {
+        return this.sendSeq < this.segments.length;
+    }
+
+    private scheduleFlush(): void {
+        if (this.flushTimer) {
+            return;
+        }
+        this.flushTimer = setTimeout(() => {
+            this.flushTimer = null;
+            this.flush();
+        }, 0);
+    }
+
+    private clearScheduledFlush(): void {
+        if (!this.flushTimer) {
+            return;
+        }
+        clearTimeout(this.flushTimer);
+        this.flushTimer = null;
+    }
+
+    private async waitForFlushDrain(): Promise<void> {
+        while (this.hasPendingSegments()) {
+            if (!this.client.isConnected || !this.dictationId || !this.streamReady) {
+                throw new Error('Failed to flush dictation stream');
+            }
+            await new Promise<void>((resolve) => {
+                this.drainWaiters.push(resolve);
+            });
+            await waitForNextFlushTurn();
+        }
+    }
+
+    private resolveDrainWaiters(): void {
+        const waiters = this.drainWaiters;
+        this.drainWaiters = [];
+        for (const resolve of waiters) {
+            resolve();
+        }
+    }
+}

packages/ui/src/lib/dictation/use-dictation-audio-source.ts

@@ -0,0 +1,301 @@
+/**
+ * Microphone capture for dictation.
+ *
+ * Captures mono audio via getUserMedia, taps it with a ScriptProcessorNode
+ * (universally supported, including iOS WKWebView), resamples Float32 to
+ * 16 kHz PCM16LE, and emits ~1-second base64 chunks plus a normalized RMS
+ * volume for the level meter.
+ */
+
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+
+export interface DictationAudioSourceConfig {
+    onPcmSegment: (base64Pcm: string) => void;
+    onError?: (error: Error) => void;
+}
+
+export interface DictationAudioSource {
+    start: () => Promise<void>;
+    stop: () => Promise<void>;
+    volume: number;
+}
+
+const OUTPUT_RATE = 16000;
+const CHUNK_SAMPLES = OUTPUT_RATE; // ~1s per chunk
+
+const getAudioContextCtor = (): typeof AudioContext | null => {
+    if (typeof window === 'undefined') {
+        return null;
+    }
+    const win = window as typeof window & { webkitAudioContext?: typeof AudioContext };
+    return win.AudioContext || win.webkitAudioContext || null;
+};
+
+const floatToInt16 = (sample: number): number => {
+    const clamped = Math.max(-1, Math.min(1, sample));
+    return clamped < 0 ? Math.round(clamped * 0x8000) : Math.round(clamped * 0x7fff);
+};
+
+const resampleToPcm16 = (input: Float32Array, inputRate: number, outputRate: number): Int16Array => {
+    if (input.length === 0) {
+        return new Int16Array(0);
+    }
+    if (inputRate === outputRate) {
+        const out = new Int16Array(input.length);
+        for (let i = 0; i < input.length; i++) {
+            out[i] = floatToInt16(input[i]);
+        }
+        return out;
+    }
+
+    const ratio = inputRate / outputRate;
+    const outputLength = Math.max(1, Math.round(input.length / ratio));
+    const out = new Int16Array(outputLength);
+    for (let i = 0; i < outputLength; i++) {
+        const sourceIndex = i * ratio;
+        const i0 = Math.floor(sourceIndex);
+        const i1 = Math.min(input.length - 1, i0 + 1);
+        const frac = sourceIndex - i0;
+        out[i] = floatToInt16(input[i0] * (1 - frac) + input[i1] * frac);
+    }
+    return out;
+};
+
+const concatInt16 = (a: Int16Array, b: Int16Array): Int16Array => {
+    if (a.length === 0) {
+        return b;
+    }
+    if (b.length === 0) {
+        return a;
+    }
+    const out = new Int16Array(a.length + b.length);
+    out.set(a, 0);
+    out.set(b, a.length);
+    return out;
+};
+
+const int16ToBase64 = (pcm: Int16Array): string => {
+    const bytes = new Uint8Array(pcm.buffer, pcm.byteOffset, pcm.byteLength);
+    let binary = '';
+    for (let i = 0; i < bytes.length; i++) {
+        binary += String.fromCharCode(bytes[i]);
+    }
+    return btoa(binary);
+};
+
+interface CaptureGraph {
+    stream: MediaStream | null;
+    context: AudioContext | null;
+    source: MediaStreamAudioSourceNode | null;
+    processor: ScriptProcessorNode | null;
+    gain: GainNode | null;
+    pending: Int16Array;
+    started: boolean;
+}
+
+const emptyGraph = (): CaptureGraph => ({
+    stream: null,
+    context: null,
+    source: null,
+    processor: null,
+    gain: null,
+    pending: new Int16Array(0),
+    started: false,
+});
+
+const safeDisconnect = (node: AudioNode | null): void => {
+    if (!node) {
+        return;
+    }
+    try {
+        node.disconnect();
+    } catch {
+        // no-op
+    }
+};
+
+export const isDictationCaptureSupported = (): boolean => {
+    if (typeof navigator === 'undefined' || typeof window === 'undefined') {
+        return false;
+    }
+    if (!navigator.mediaDevices || typeof navigator.mediaDevices.getUserMedia !== 'function') {
+        return false;
+    }
+    return getAudioContextCtor() !== null;
+};
+
+export function useDictationAudioSource(config: DictationAudioSourceConfig): DictationAudioSource {
+    const [volume, setVolume] = useState(0);
+
+    const onPcmSegmentRef = useRef(config.onPcmSegment);
+    const onErrorRef = useRef(config.onError);
+    useEffect(() => {
+        onPcmSegmentRef.current = config.onPcmSegment;
+        onErrorRef.current = config.onError;
+    }, [config.onPcmSegment, config.onError]);
+
+    const graphRef = useRef<CaptureGraph>(emptyGraph());
+
+    const start = useCallback(async () => {
+        if (graphRef.current.started) {
+            return;
+        }
+
+        if (
+            typeof navigator === 'undefined' ||
+            !navigator.mediaDevices ||
+            typeof navigator.mediaDevices.getUserMedia !== 'function'
+        ) {
+            throw new Error('Microphone capture is not supported in this environment');
+        }
+
+        const AudioContextCtor = getAudioContextCtor();
+        if (!AudioContextCtor) {
+            throw new Error('AudioContext unavailable');
+        }
+
+        const stream = await navigator.mediaDevices.getUserMedia({
+            audio: {
+                channelCount: 1,
+                noiseSuppression: true,
+                echoCancellation: true,
+                autoGainControl: true,
+            },
+        });
+
+        const context = new AudioContextCtor();
+        try {
+            if (context.state === 'suspended') {
+                await context.resume().catch(() => undefined);
+            }
+
+            const source = context.createMediaStreamSource(stream);
+            const processor = context.createScriptProcessor(4096, 1, 1);
+            const gain = context.createGain();
+            gain.gain.value = 0;
+
+            graphRef.current = {
+                stream,
+                context,
+                source,
+                processor,
+                gain,
+                pending: new Int16Array(0),
+                started: true,
+            };
+
+            processor.onaudioprocess = (event) => {
+                const graph = graphRef.current;
+                if (!graph.started) {
+                    return;
+                }
+                const input = event.inputBuffer.getChannelData(0);
+
+                let sumSquares = 0;
+                for (let i = 0; i < input.length; i++) {
+                    sumSquares += input[i] * input[i];
+                }
+                const rms = Math.sqrt(sumSquares / Math.max(1, input.length));
+                setVolume(Math.min(1, Math.max(0, rms * 2)));
+
+                const next = resampleToPcm16(input, context.sampleRate, OUTPUT_RATE);
+                graph.pending = concatInt16(graph.pending, next);
+
+                while (graph.pending.length >= CHUNK_SAMPLES) {
+                    const chunk = graph.pending.slice(0, CHUNK_SAMPLES);
+                    graph.pending = graph.pending.slice(CHUNK_SAMPLES);
+                    onPcmSegmentRef.current(int16ToBase64(chunk));
+                }
+            };
+
+            source.connect(processor);
+            processor.connect(gain);
+            gain.connect(context.destination);
+        } catch (error) {
+            stream.getTracks().forEach((track) => {
+                try {
+                    track.stop();
+                } catch {
+                    // no-op
+                }
+            });
+            try {
+                await context.close();
+            } catch {
+                // no-op
+            }
+            graphRef.current = emptyGraph();
+            throw error instanceof Error ? error : new Error(String(error));
+        }
+    }, []);
+
+    const stop = useCallback(async () => {
+        const graph = graphRef.current;
+        graph.started = false;
+        setVolume(0);
+
+        if (graph.processor) {
+            try {
+                graph.processor.onaudioprocess = null;
+            } catch {
+                // no-op
+            }
+        }
+        safeDisconnect(graph.processor);
+        safeDisconnect(graph.source);
+        safeDisconnect(graph.gain);
+        if (graph.stream) {
+            graph.stream.getTracks().forEach((track) => {
+                try {
+                    track.stop();
+                } catch {
+                    // no-op
+                }
+            });
+        }
+        const pending = graph.pending;
+        graph.pending = new Int16Array(0);
+        if (pending.length > 0) {
+            onPcmSegmentRef.current(int16ToBase64(pending));
+        }
+
+        if (graph.context) {
+            try {
+                await graph.context.close();
+            } catch {
+                // no-op
+            }
+        }
+
+        // A new capture may have started while the old context was closing;
+        // only clear the ref if it still points at the graph we tore down.
+        if (graphRef.current === graph) {
+            graphRef.current = emptyGraph();
+        }
+    }, []);
+
+    useEffect(() => {
+        return () => {
+            void stop().catch((err) => {
+                onErrorRef.current?.(err instanceof Error ? err : new Error(String(err)));
+            });
+        };
+    }, [stop]);
+
+    return useMemo(
+        () => ({
+            start: async () => {
+                try {
+                    await start();
+                } catch (err) {
+                    const normalized = err instanceof Error ? err : new Error(String(err));
+                    onErrorRef.current?.(normalized);
+                    throw normalized;
+                }
+            },
+            stop,
+            volume,
+        }),
+        [start, stop, volume],
+    );
+}

packages/ui/src/lib/i18n/messages/en.settings.ts

@@ -1518,16 +1518,33 @@ export const settingsDict = {
   'settings.notifications.page.toast.backgroundDisabled': 'Background notifications disabled',
   'settings.notifications.page.testNotification.title': 'Test Notification',
   'settings.notifications.page.testNotification.body': 'This is a test notification from OpenChamber.',
-  'settings.voice.page.section.voiceSetup': 'Voice Setup',
   'settings.voice.page.section.speechRecognition': 'Speech Recognition',
+  'settings.voice.page.field.enableVoiceInput': 'Enable voice input',
+  'settings.voice.page.field.enableVoiceInputAria': 'Enable voice input (dictation)',
   'settings.voice.page.section.playbackAndSummary': 'Playback',
-  'settings.voice.page.field.enableVoiceModeAria': 'Enable voice mode',
-  'settings.voice.page.field.enableVoiceMode': 'Enable Voice Mode',
   'settings.voice.page.field.provider': 'Provider',
   'settings.voice.page.provider.browser': 'Browser',
   'settings.voice.page.provider.custom': 'Custom',
   'settings.voice.page.provider.say': 'Say',
   'settings.voice.page.provider.server': 'Server',
+  'settings.voice.page.provider.local': 'Local',
+  'settings.voice.page.tooltip.sttLocal': 'On-device transcription on the OpenChamber server. Models download automatically; no API key needed.',
+  'settings.voice.page.tooltip.localTts': 'On-device synthesis on the OpenChamber server (Kokoro, English). The model downloads automatically; no API key needed.',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2 (English)',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3 (25 European languages)',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base (multilingual)',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny (multilingual)',
+  'settings.voice.page.stt.badge.bestForEnglish': 'Best for English',
+  'settings.voice.page.stt.badge.bestForMultilingual': 'Best for multilingual',
+  'settings.voice.page.stt.meta.accuracy': 'Accuracy',
+  'settings.voice.page.stt.meta.speed': 'Speed',
+  'settings.voice.page.stt.modelInstalled': 'Model installed',
+  'settings.voice.page.stt.modelDownloading': 'Downloading model...',
+  'settings.voice.page.stt.modelDownloadingProgress': 'Downloading model... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': 'Model not downloaded',
+  'settings.voice.page.stt.modelDownload': 'Download',
+  'settings.voice.page.stt.modelDelete': 'Delete model',
+  'settings.voice.page.stt.modelRetry': 'Retry',
   'settings.voice.page.provider.wasm': 'Local',
   'settings.voice.page.stt.wasmModel': 'Whisper Model',
   'settings.voice.page.stt.wasmLoaded': 'Model loaded, ready',

packages/ui/src/lib/i18n/messages/en.ts

@@ -1873,6 +1873,18 @@ export const dict = {
   'chat.chatInput.placeholder.chat': '@ for files/agents; / for commands and skills; ! for shell; # for snippets',
   'chat.chatInput.placeholder.chatCompact': 'Use @ / ! # for helpers',
   'chat.chatInput.placeholder.selectSession': 'Select or create a session to start chatting',
+  'chat.dictation.start': 'Start dictation',
+  'chat.dictation.overlayAria': 'Dictation',
+  'chat.dictation.downloadingModel': 'Downloading speech model...',
+  'chat.dictation.downloadingModelProgress': 'Downloading speech model... {percent}%',
+  'chat.dictation.listening': 'Listening...',
+  'chat.dictation.processing': 'Transcribing...',
+  'chat.dictation.failed': 'Transcription failed',
+  'chat.dictation.cancel': 'Discard dictation',
+  'chat.dictation.insert': 'Insert transcript',
+  'chat.dictation.insertAndSend': 'Insert and send',
+  'chat.dictation.retry': 'Retry transcription',
+  'chat.dictation.discard': 'Discard recording',
   'chat.snippetAutocomplete.action.addNew': '+ Add new snippet',
   'chat.snippetAutocomplete.empty': 'No snippets found',
   'chat.snippetAutocomplete.footer': '↑↓ navigate • Enter select • Esc close',

packages/ui/src/lib/i18n/messages/es.settings.ts

@@ -1485,16 +1485,33 @@ export const settingsDict = {
   "settings.notifications.page.toast.backgroundDisabled": "Notificaciones de fondo deshabilitadas",
   "settings.notifications.page.testNotification.title": "Notificación de prueba",
   "settings.notifications.page.testNotification.body": "Esta es una notificación de prueba de OpenChamber.",
-  "settings.voice.page.section.voiceSetup": "Configuración de voz",
   "settings.voice.page.section.speechRecognition": "Reconocimiento de voz",
+  "settings.voice.page.field.enableVoiceInput": "Habilitar entrada de voz",
+  "settings.voice.page.field.enableVoiceInputAria": "Habilitar entrada de voz (dictado)",
   "settings.voice.page.section.playbackAndSummary": "Reproducción",
-  "settings.voice.page.field.enableVoiceModeAria": "Habilitar modo de voz",
-  "settings.voice.page.field.enableVoiceMode": "Habilitar modo de voz",
   "settings.voice.page.field.provider": "Proveedor",
   "settings.voice.page.provider.browser": "Navegador",
   "settings.voice.page.provider.custom": "Personalizado",
   "settings.voice.page.provider.say": "Decir",
   "settings.voice.page.provider.server": "Servidor",
+  "settings.voice.page.provider.local": "Local",
+  "settings.voice.page.tooltip.sttLocal": "Transcripción local en el servidor de OpenChamber. Los modelos se descargan automáticamente; no se necesita clave de API.",
+  "settings.voice.page.tooltip.localTts": "Síntesis local en el servidor de OpenChamber (Kokoro, inglés). El modelo se descarga automáticamente; no se necesita clave de API.",
+  "settings.voice.page.stt.model.parakeetV2": "Parakeet v2 (inglés)",
+  "settings.voice.page.stt.model.parakeetV3": "Parakeet v3 (25 idiomas europeos)",
+  "settings.voice.page.stt.model.whisperBase": "Whisper base (multilingüe)",
+  "settings.voice.page.stt.model.whisperTiny": "Whisper tiny (multilingüe)",
+  "settings.voice.page.stt.badge.bestForEnglish": "Mejor para inglés",
+  "settings.voice.page.stt.badge.bestForMultilingual": "Mejor multilingüe",
+  "settings.voice.page.stt.meta.accuracy": "Precisión",
+  "settings.voice.page.stt.meta.speed": "Velocidad",
+  "settings.voice.page.stt.modelInstalled": "Modelo instalado",
+  "settings.voice.page.stt.modelDownloading": "Descargando modelo...",
+  "settings.voice.page.stt.modelDownloadingProgress": "Descargando modelo... {percent}%",
+  "settings.voice.page.stt.modelNotInstalled": "Modelo no descargado",
+  "settings.voice.page.stt.modelDownload": "Descargar",
+  "settings.voice.page.stt.modelDelete": "Eliminar modelo",
+  "settings.voice.page.stt.modelRetry": "Reintentar",
   "settings.voice.page.provider.wasm": "Local",
   "settings.voice.page.stt.wasmModel": "Modelo Whisper",
   "settings.voice.page.stt.wasmLoaded": "Modelo cargado",

packages/ui/src/lib/i18n/messages/es.ts

@@ -1245,6 +1245,18 @@ export const dict: Record<I18nKey, string> = {
   'diffView.reviewDialog.actions.starting': 'Iniciando...',
   'diffView.reviewDialog.toast.noSessionDirectory': 'El directorio de la sesión no está disponible',
   'diffView.reviewDialog.toast.startFailed': 'No se pudo iniciar el flujo de revisión',
+  'chat.dictation.start': 'Iniciar dictado',
+  'chat.dictation.overlayAria': 'Dictado',
+  'chat.dictation.downloadingModel': 'Descargando el modelo de voz...',
+  'chat.dictation.downloadingModelProgress': 'Descargando el modelo de voz... {percent}%',
+  'chat.dictation.listening': 'Escuchando...',
+  'chat.dictation.processing': 'Transcribiendo...',
+  'chat.dictation.failed': 'La transcripción falló',
+  'chat.dictation.cancel': 'Descartar dictado',
+  'chat.dictation.insert': 'Insertar transcripción',
+  'chat.dictation.insertAndSend': 'Insertar y enviar',
+  'chat.dictation.retry': 'Reintentar transcripción',
+  'chat.dictation.discard': 'Descartar grabación',
   'chat.history.loadOlder': 'Cargar mensajes anteriores',
   'chat.autoReview.title': 'El ciclo de revisión de código está en curso',
   'chat.autoReview.status.waitingForReviewer': 'Esperando al revisor',

packages/ui/src/lib/i18n/messages/fr.settings.ts

@@ -1468,16 +1468,33 @@ export const settingsDict = {
   'settings.notifications.page.toast.backgroundDisabled': 'Notifications en arrière-plan désactivées',
   'settings.notifications.page.testNotification.title': 'Avis de test',
   'settings.notifications.page.testNotification.body': 'Il s\'agit d\'une notification de test de OpenChamber.',
-  'settings.voice.page.section.voiceSetup': 'Configuration vocale',
   'settings.voice.page.section.speechRecognition': 'Reconnaissance vocale',
+  'settings.voice.page.field.enableVoiceInput': 'Activer la saisie vocale',
+  'settings.voice.page.field.enableVoiceInputAria': 'Activer la saisie vocale (dictée)',
   'settings.voice.page.section.playbackAndSummary': 'Lecture',
-  'settings.voice.page.field.enableVoiceModeAria': 'Activer le mode vocal',
-  'settings.voice.page.field.enableVoiceMode': 'Activer le mode vocal',
   'settings.voice.page.field.provider': 'Fournisseur',
   'settings.voice.page.provider.browser': 'Navigateur',
   'settings.voice.page.provider.custom': 'Personnalisé',
   'settings.voice.page.provider.say': 'Dire',
   'settings.voice.page.provider.server': 'Serveur',
+  'settings.voice.page.provider.local': 'Local',
+  'settings.voice.page.tooltip.sttLocal': 'Transcription locale sur le serveur OpenChamber. Les modèles se téléchargent automatiquement ; aucune clé d\'API requise.',
+  'settings.voice.page.tooltip.localTts': 'Synthèse locale sur le serveur OpenChamber (Kokoro, anglais). Le modèle se télécharge automatiquement ; aucune clé d’API requise.',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2 (anglais)',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3 (25 langues européennes)',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base (multilingue)',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny (multilingue)',
+  'settings.voice.page.stt.badge.bestForEnglish': 'Idéal pour l’anglais',
+  'settings.voice.page.stt.badge.bestForMultilingual': 'Idéal multilingue',
+  'settings.voice.page.stt.meta.accuracy': 'Précision',
+  'settings.voice.page.stt.meta.speed': 'Vitesse',
+  'settings.voice.page.stt.modelInstalled': 'Modèle installé',
+  'settings.voice.page.stt.modelDownloading': 'Téléchargement du modèle...',
+  'settings.voice.page.stt.modelDownloadingProgress': 'Téléchargement du modèle... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': 'Modèle non téléchargé',
+  'settings.voice.page.stt.modelDownload': 'Télécharger',
+  'settings.voice.page.stt.modelDelete': 'Supprimer le modèle',
+  'settings.voice.page.stt.modelRetry': 'Réessayer',
   'settings.voice.page.provider.wasm': 'Local',
   'settings.voice.page.stt.wasmModel': 'Modèle Whisper',
   'settings.voice.page.stt.wasmLoaded': 'Modèle chargé, prêt',

packages/ui/src/lib/i18n/messages/fr.ts

@@ -1669,6 +1669,18 @@ export const dict = {
   'chat.chatInput.placeholder.chat': '@ pour les fichiers/agents ; / pour les commandes et les skills ; ! pour shell ; # pour les extraits',
   'chat.chatInput.placeholder.chatCompact': 'Utiliser @ / ! # pour les aides',
   'chat.chatInput.placeholder.selectSession': 'Sélectionnez ou créez une session pour commencer à discuter',
+  'chat.dictation.start': 'Démarrer la dictée',
+  'chat.dictation.overlayAria': 'Dictée',
+  'chat.dictation.downloadingModel': 'Téléchargement du modèle vocal...',
+  'chat.dictation.downloadingModelProgress': 'Téléchargement du modèle vocal... {percent}%',
+  'chat.dictation.listening': 'Écoute...',
+  'chat.dictation.processing': 'Transcription...',
+  'chat.dictation.failed': 'Échec de la transcription',
+  'chat.dictation.cancel': 'Abandonner la dictée',
+  'chat.dictation.insert': 'Insérer la transcription',
+  'chat.dictation.insertAndSend': 'Insérer et envoyer',
+  'chat.dictation.retry': 'Réessayer la transcription',
+  'chat.dictation.discard': 'Abandonner l\'enregistrement',
   'chat.snippetAutocomplete.action.addNew': '+ Ajouter un nouvel extrait',
   'chat.snippetAutocomplete.empty': 'Aucun extrait trouvé',
   'chat.snippetAutocomplete.footer': '↑↓ naviguer • Entrer sélectionner • Esc fermer',

packages/ui/src/lib/i18n/messages/ja.settings.ts

@@ -1518,16 +1518,33 @@ export const settingsDict = {
   'settings.notifications.page.toast.backgroundDisabled': 'バックグラウンド通知を無効化しました',
   'settings.notifications.page.testNotification.title': 'テスト通知',
   'settings.notifications.page.testNotification.body': 'これは OpenChamber からのテスト通知です。',
-  'settings.voice.page.section.voiceSetup': '音声設定',
   'settings.voice.page.section.speechRecognition': '音声認識',
+  'settings.voice.page.field.enableVoiceInput': '音声入力を有効にする',
+  'settings.voice.page.field.enableVoiceInputAria': '音声入力（ディクテーション）を有効にする',
   'settings.voice.page.section.playbackAndSummary': '再生',
-  'settings.voice.page.field.enableVoiceModeAria': '音声モードを有効化',
-  'settings.voice.page.field.enableVoiceMode': '音声モードを有効化',
   'settings.voice.page.field.provider': 'プロバイダー',
   'settings.voice.page.provider.browser': 'ブラウザ',
   'settings.voice.page.provider.custom': 'カスタム',
   'settings.voice.page.provider.say': 'Say',
   'settings.voice.page.provider.server': 'サーバー',
+  'settings.voice.page.provider.local': 'ローカル',
+  'settings.voice.page.tooltip.sttLocal': 'OpenChamber サーバー上でローカルに文字起こしします。モデルは自動でダウンロードされ、API キーは不要です。',
+  'settings.voice.page.tooltip.localTts': 'OpenChamber サーバー上でローカルに音声合成します（Kokoro、英語）。モデルは自動でダウンロードされ、API キーは不要です。',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2（英語）',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3（ヨーロッパ25言語）',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base（多言語）',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny（多言語）',
+  'settings.voice.page.stt.badge.bestForEnglish': '英語に最適',
+  'settings.voice.page.stt.badge.bestForMultilingual': '多言語に最適',
+  'settings.voice.page.stt.meta.accuracy': '精度',
+  'settings.voice.page.stt.meta.speed': '速度',
+  'settings.voice.page.stt.modelInstalled': 'モデルはインストール済み',
+  'settings.voice.page.stt.modelDownloading': 'モデルをダウンロード中...',
+  'settings.voice.page.stt.modelDownloadingProgress': 'モデルをダウンロード中... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': 'モデル未ダウンロード',
+  'settings.voice.page.stt.modelDownload': 'ダウンロード',
+  'settings.voice.page.stt.modelDelete': 'モデルを削除',
+  'settings.voice.page.stt.modelRetry': '再試行',
   'settings.voice.page.provider.wasm': 'ローカル',
   'settings.voice.page.stt.wasmModel': 'Whisper モデル',
   'settings.voice.page.stt.wasmLoaded': 'モデル読み込み完了',

packages/ui/src/lib/i18n/messages/ja.ts

@@ -1869,6 +1869,18 @@ export const dict: Record<I18nKey, string> = {
   'chat.chatInput.placeholder.chat': '@でファイル/エージェント、/でコマンド/スキル、!でシェル、#でスニペット',
   'chat.chatInput.placeholder.chatCompact': '@ / ! # でヘルパーを使用',
   'chat.chatInput.placeholder.selectSession': 'セッションを選択または作成してチャットを開始',
+  'chat.dictation.start': '音声入力を開始',
+  'chat.dictation.overlayAria': '音声入力',
+  'chat.dictation.downloadingModel': '音声モデルをダウンロード中...',
+  'chat.dictation.downloadingModelProgress': '音声モデルをダウンロード中... {percent}%',
+  'chat.dictation.listening': '聞き取り中...',
+  'chat.dictation.processing': '文字起こし中...',
+  'chat.dictation.failed': '文字起こしに失敗しました',
+  'chat.dictation.cancel': '音声入力を破棄',
+  'chat.dictation.insert': 'テキストを挿入',
+  'chat.dictation.insertAndSend': '挿入して送信',
+  'chat.dictation.retry': '文字起こしを再試行',
+  'chat.dictation.discard': '録音を破棄',
   'chat.snippetAutocomplete.action.addNew': '+ 新しいスニペットを追加',
   'chat.snippetAutocomplete.empty': 'スニペットが見つかりません',
   'chat.snippetAutocomplete.footer': '↑↓ 移動 • Enter 選択 • Esc 閉じる',

packages/ui/src/lib/i18n/messages/ko.settings.ts

@@ -1485,16 +1485,33 @@ export const settingsDict = {
   'settings.notifications.page.toast.backgroundDisabled': '백그라운드 알림이 비활성화되었습니다',
   'settings.notifications.page.testNotification.title': '테스트 알림',
   'settings.notifications.page.testNotification.body': 'OpenChamber의 테스트 알림입니다.',
-  'settings.voice.page.section.voiceSetup': '음성',
   'settings.voice.page.section.speechRecognition': '음성 인식',
+  'settings.voice.page.field.enableVoiceInput': '음성 입력 사용',
+  'settings.voice.page.field.enableVoiceInputAria': '음성 입력(받아쓰기) 사용',
   'settings.voice.page.section.playbackAndSummary': '재생',
-  'settings.voice.page.field.enableVoiceModeAria': '음성 모드 활성화',
-  'settings.voice.page.field.enableVoiceMode': '음성 모드 활성화',
   'settings.voice.page.field.provider': '프로바이더',
   'settings.voice.page.provider.browser': '브라우저',
   'settings.voice.page.provider.custom': '사용자 정의',
   'settings.voice.page.provider.say': 'Say',
   'settings.voice.page.provider.server': '서버',
+  'settings.voice.page.provider.local': '로컬',
+  'settings.voice.page.tooltip.sttLocal': 'OpenChamber 서버에서 로컬로 변환합니다. 모델은 자동으로 다운로드되며 API 키가 필요 없습니다.',
+  'settings.voice.page.tooltip.localTts': 'OpenChamber 서버에서 로컬로 음성을 합성합니다(Kokoro, 영어). 모델은 자동으로 다운로드되며 API 키가 필요 없습니다.',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2 (영어)',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3 (유럽 25개 언어)',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base (다국어)',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny (다국어)',
+  'settings.voice.page.stt.badge.bestForEnglish': '영어에 최적',
+  'settings.voice.page.stt.badge.bestForMultilingual': '다국어에 최적',
+  'settings.voice.page.stt.meta.accuracy': '정확도',
+  'settings.voice.page.stt.meta.speed': '속도',
+  'settings.voice.page.stt.modelInstalled': '모델 설치됨',
+  'settings.voice.page.stt.modelDownloading': '모델 다운로드 중...',
+  'settings.voice.page.stt.modelDownloadingProgress': '모델 다운로드 중... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': '모델이 다운로드되지 않음',
+  'settings.voice.page.stt.modelDownload': '다운로드',
+  'settings.voice.page.stt.modelDelete': '모델 삭제',
+  'settings.voice.page.stt.modelRetry': '다시 시도',
   'settings.voice.page.provider.wasm': '로컬',
   'settings.voice.page.stt.wasmModel': 'Whisper 모델',
   'settings.voice.page.stt.wasmLoaded': '모델 로드됨',

packages/ui/src/lib/i18n/messages/ko.ts

@@ -1873,6 +1873,18 @@ export const dict: Record<I18nKey, string> = {
   'chat.chatInput.placeholder.chat': '@ 파일/에이전트; / 명령 및 스킬; ! shell; # 스니펫',
   'chat.chatInput.placeholder.chatCompact': '@ / ! # 도우미 사용',
   'chat.chatInput.placeholder.selectSession': '채팅을 시작할 세션을 선택하거나 새로 만드세요',
+  'chat.dictation.start': '받아쓰기 시작',
+  'chat.dictation.overlayAria': '받아쓰기',
+  'chat.dictation.downloadingModel': '음성 모델 다운로드 중...',
+  'chat.dictation.downloadingModelProgress': '음성 모델 다운로드 중... {percent}%',
+  'chat.dictation.listening': '듣는 중...',
+  'chat.dictation.processing': '변환 중...',
+  'chat.dictation.failed': '변환에 실패했습니다',
+  'chat.dictation.cancel': '받아쓰기 취소',
+  'chat.dictation.insert': '텍스트 삽입',
+  'chat.dictation.insertAndSend': '삽입 후 전송',
+  'chat.dictation.retry': '변환 다시 시도',
+  'chat.dictation.discard': '녹음 삭제',
   'chat.snippetAutocomplete.action.addNew': '+ 새 스니펫 추가',
   'chat.snippetAutocomplete.empty': '스니펫을 찾을 수 없음',
   'chat.snippetAutocomplete.footer': '↑↓ 이동 • Enter 선택 • Esc 닫기',

packages/ui/src/lib/i18n/messages/pl.settings.ts

@@ -1743,8 +1743,6 @@ export const settingsDict = {
   'settings.voice.page.field.apiKeyHintUsingConfig': 'Użyto klucza z konfiguracji',
   'settings.voice.page.field.auto': 'Auto',
   'settings.voice.page.field.configuredAbove': 'Skonfigurowano powyżej',
-  'settings.voice.page.field.enableVoiceMode': 'Włącz tryb głosowy',
-  'settings.voice.page.field.enableVoiceModeAria': 'Włącz tryb głosowy',
   'settings.voice.page.field.language': 'Język',
   'settings.voice.page.field.messageReadAloudButton': 'Przycisk czytania wiadomości na głos',
   'settings.voice.page.field.messageReadAloudButtonAria': 'Przycisk czytania wiadomości na głos',
@@ -1782,6 +1780,24 @@ export const settingsDict = {
   'settings.voice.page.provider.custom': 'Własny',
   'settings.voice.page.provider.say': 'Say',
   'settings.voice.page.provider.server': 'Serwer',
+  'settings.voice.page.provider.local': 'Lokalny',
+  'settings.voice.page.tooltip.sttLocal': 'Transkrypcja lokalna na serwerze OpenChamber. Modele pobierają się automatycznie; klucz API nie jest potrzebny.',
+  'settings.voice.page.tooltip.localTts': 'Lokalna synteza na serwerze OpenChamber (Kokoro, angielski). Model pobiera się automatycznie; klucz API nie jest potrzebny.',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2 (angielski)',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3 (25 języków europejskich)',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base (wielojęzyczny)',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny (wielojęzyczny)',
+  'settings.voice.page.stt.badge.bestForEnglish': 'Najlepszy do angielskiego',
+  'settings.voice.page.stt.badge.bestForMultilingual': 'Najlepszy wielojęzyczny',
+  'settings.voice.page.stt.meta.accuracy': 'Dokładność',
+  'settings.voice.page.stt.meta.speed': 'Szybkość',
+  'settings.voice.page.stt.modelInstalled': 'Model zainstalowany',
+  'settings.voice.page.stt.modelDownloading': 'Pobieranie modelu...',
+  'settings.voice.page.stt.modelDownloadingProgress': 'Pobieranie modelu... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': 'Model niepobrany',
+  'settings.voice.page.stt.modelDownload': 'Pobierz',
+  'settings.voice.page.stt.modelDelete': 'Usuń model',
+  'settings.voice.page.stt.modelRetry': 'Ponów',
   'settings.voice.page.provider.wasm': 'Lokalny',
   'settings.voice.page.stt.wasmModel': 'Model Whisper',
   'settings.voice.page.stt.wasmLoaded': 'Model załadowany',
@@ -1792,7 +1808,8 @@ export const settingsDict = {
   'settings.voice.page.stt.wasmRetry': 'Spróbuj ponownie',
   'settings.voice.page.section.playbackAndSummary': 'Odtwarzanie',
   'settings.voice.page.section.speechRecognition': 'Rozpoznawanie mowy',
-  'settings.voice.page.section.voiceSetup': 'Konfiguracja głosu',
+  'settings.voice.page.field.enableVoiceInput': 'Włącz wprowadzanie głosowe',
+  'settings.voice.page.field.enableVoiceInputAria': 'Włącz wprowadzanie głosowe (dyktowanie)',
   'settings.voice.page.tooltip.browser': 'Darmowe, offline, ograniczone wsparcie mobilne.',
   'settings.voice.page.tooltip.custom': 'Serwer zgodny z OpenAI (na przykład Kokoro).',
   'settings.voice.page.tooltip.openai': 'Wysoka jakość, gotowe na urządzenia mobilne, wymaga klucza API.',

packages/ui/src/lib/i18n/messages/pl.ts

@@ -1091,6 +1091,18 @@ export const dict: Record<I18nKey, string> = {
   'chat.chatInput.placeholder.chatCompact': 'Użyj @ / ! # dla pomocników',
   'chat.chatInput.placeholder.selectSession': 'Wybierz lub utwórz sesję, aby zacząć czatować',
   'chat.chatInput.placeholder.shell': 'Wpisz polecenie powłoki...',
+  'chat.dictation.start': 'Rozpocznij dyktowanie',
+  'chat.dictation.overlayAria': 'Dyktowanie',
+  'chat.dictation.downloadingModel': 'Pobieranie modelu mowy...',
+  'chat.dictation.downloadingModelProgress': 'Pobieranie modelu mowy... {percent}%',
+  'chat.dictation.listening': 'Słucham...',
+  'chat.dictation.processing': 'Transkrybowanie...',
+  'chat.dictation.failed': 'Transkrypcja nie powiodła się',
+  'chat.dictation.cancel': 'Odrzuć dyktowanie',
+  'chat.dictation.insert': 'Wstaw transkrypcję',
+  'chat.dictation.insertAndSend': 'Wstaw i wyślij',
+  'chat.dictation.retry': 'Ponów transkrypcję',
+  'chat.dictation.discard': 'Odrzuć nagranie',
   'chat.snippetAutocomplete.action.addNew': '+ Dodaj nowy fragment',
   'chat.snippetAutocomplete.empty': 'Nie znaleziono fragmentów',
   'chat.snippetAutocomplete.footer': '↑↓ nawigacja • Enter wybierz • Esc zamknij',

packages/ui/src/lib/i18n/messages/pt-BR.settings.ts

@@ -1485,16 +1485,33 @@ export const settingsDict = {
   "settings.notifications.page.toast.backgroundDisabled": "Notificações em segundo plano desabilitadas",
   "settings.notifications.page.testNotification.title": "Notificação de teste",
   "settings.notifications.page.testNotification.body": "Esta é uma notificação de teste do OpenChamber.",
-  "settings.voice.page.section.voiceSetup": "Configurações de voz",
   "settings.voice.page.section.speechRecognition": "Reconhecimento de voz",
+  "settings.voice.page.field.enableVoiceInput": "Ativar entrada por voz",
+  "settings.voice.page.field.enableVoiceInputAria": "Ativar entrada por voz (ditado)",
   "settings.voice.page.section.playbackAndSummary": "Reprodução",
-  "settings.voice.page.field.enableVoiceModeAria": "Ativar modo de voz",
-  "settings.voice.page.field.enableVoiceMode": "Ativar modo de voz",
   "settings.voice.page.field.provider": "Provedor",
   "settings.voice.page.provider.browser": "Navegador",
   "settings.voice.page.provider.custom": "Personalizado",
   "settings.voice.page.provider.say": "Falar",
   "settings.voice.page.provider.server": "Servidor",
+  "settings.voice.page.provider.local": "Local",
+  "settings.voice.page.tooltip.sttLocal": "Transcrição local no servidor do OpenChamber. Os modelos são baixados automaticamente; não é necessária chave de API.",
+  "settings.voice.page.tooltip.localTts": "Síntese local no servidor do OpenChamber (Kokoro, inglês). O modelo é baixado automaticamente; não é necessária chave de API.",
+  "settings.voice.page.stt.model.parakeetV2": "Parakeet v2 (inglês)",
+  "settings.voice.page.stt.model.parakeetV3": "Parakeet v3 (25 idiomas europeus)",
+  "settings.voice.page.stt.model.whisperBase": "Whisper base (multilíngue)",
+  "settings.voice.page.stt.model.whisperTiny": "Whisper tiny (multilíngue)",
+  "settings.voice.page.stt.badge.bestForEnglish": "Melhor para inglês",
+  "settings.voice.page.stt.badge.bestForMultilingual": "Melhor multilíngue",
+  "settings.voice.page.stt.meta.accuracy": "Precisão",
+  "settings.voice.page.stt.meta.speed": "Velocidade",
+  "settings.voice.page.stt.modelInstalled": "Modelo instalado",
+  "settings.voice.page.stt.modelDownloading": "Baixando modelo...",
+  "settings.voice.page.stt.modelDownloadingProgress": "Baixando modelo... {percent}%",
+  "settings.voice.page.stt.modelNotInstalled": "Modelo não baixado",
+  "settings.voice.page.stt.modelDownload": "Baixar",
+  "settings.voice.page.stt.modelDelete": "Excluir modelo",
+  "settings.voice.page.stt.modelRetry": "Tentar novamente",
   "settings.voice.page.provider.wasm": "Local",
   "settings.voice.page.stt.wasmModel": "Modelo Whisper",
   "settings.voice.page.stt.wasmLoaded": "Modelo carregado",

packages/ui/src/lib/i18n/messages/pt-BR.ts

@@ -1245,6 +1245,18 @@ export const dict: Record<I18nKey, string> = {
   'diffView.reviewDialog.actions.starting': 'Iniciando...',
   'diffView.reviewDialog.toast.noSessionDirectory': 'O diretório da sessão está indisponível',
   'diffView.reviewDialog.toast.startFailed': 'Falha ao iniciar o fluxo de revisão',
+  'chat.dictation.start': 'Iniciar ditado',
+  'chat.dictation.overlayAria': 'Ditado',
+  'chat.dictation.downloadingModel': 'Baixando o modelo de fala...',
+  'chat.dictation.downloadingModelProgress': 'Baixando o modelo de fala... {percent}%',
+  'chat.dictation.listening': 'Ouvindo...',
+  'chat.dictation.processing': 'Transcrevendo...',
+  'chat.dictation.failed': 'Falha na transcrição',
+  'chat.dictation.cancel': 'Descartar ditado',
+  'chat.dictation.insert': 'Inserir transcrição',
+  'chat.dictation.insertAndSend': 'Inserir e enviar',
+  'chat.dictation.retry': 'Tentar transcrever novamente',
+  'chat.dictation.discard': 'Descartar gravação',
   'chat.history.loadOlder': 'Carregar mensagens anteriores',
   'chat.autoReview.title': 'O ciclo de revisão de código está em andamento',
   'chat.autoReview.status.waitingForReviewer': 'Aguardando o revisor',

packages/ui/src/lib/i18n/messages/uk.settings.ts

@@ -1485,16 +1485,33 @@ export const settingsDict = {
   "settings.notifications.page.toast.backgroundDisabled": "Фонові сповіщення вимкнено",
   "settings.notifications.page.testNotification.title": "Тестове сповіщення",
   "settings.notifications.page.testNotification.body": "Це тестове сповіщення від OpenChamber.",
-  "settings.voice.page.section.voiceSetup": "Налаштування голосу",
   "settings.voice.page.section.speechRecognition": "Розпізнавання мовлення",
+  "settings.voice.page.field.enableVoiceInput": "Увімкнути голосовий ввід",
+  "settings.voice.page.field.enableVoiceInputAria": "Увімкнути голосовий ввід (диктування)",
   "settings.voice.page.section.playbackAndSummary": "Відтворення",
-  "settings.voice.page.field.enableVoiceModeAria": "Увімкнути голосовий режим",
-  "settings.voice.page.field.enableVoiceMode": "Увімкнути голосовий режим",
   "settings.voice.page.field.provider": "Провайдер",
   "settings.voice.page.provider.browser": "Браузер",
   "settings.voice.page.provider.custom": "Власний",
   "settings.voice.page.provider.say": "Say",
   "settings.voice.page.provider.server": "Сервер",
+  "settings.voice.page.provider.local": "Локальний",
+  "settings.voice.page.tooltip.sttLocal": "Локальна розшифровка на сервері OpenChamber. Моделі завантажуються автоматично; ключ API не потрібен.",
+  "settings.voice.page.tooltip.localTts": "Локальний синтез на сервері OpenChamber (Kokoro, англійська). Модель завантажується автоматично; ключ API не потрібен.",
+  "settings.voice.page.stt.model.parakeetV2": "Parakeet v2 (англійська)",
+  "settings.voice.page.stt.model.parakeetV3": "Parakeet v3 (25 європейських мов)",
+  "settings.voice.page.stt.model.whisperBase": "Whisper base (мультимовна)",
+  "settings.voice.page.stt.model.whisperTiny": "Whisper tiny (мультимовна)",
+  "settings.voice.page.stt.badge.bestForEnglish": "Найкраща для англійської",
+  "settings.voice.page.stt.badge.bestForMultilingual": "Найкраща мультимовна",
+  "settings.voice.page.stt.meta.accuracy": "Точність",
+  "settings.voice.page.stt.meta.speed": "Швидкість",
+  "settings.voice.page.stt.modelInstalled": "Модель встановлено",
+  "settings.voice.page.stt.modelDownloading": "Завантаження моделі...",
+  "settings.voice.page.stt.modelDownloadingProgress": "Завантаження моделі... {percent}%",
+  "settings.voice.page.stt.modelNotInstalled": "Модель не завантажено",
+  "settings.voice.page.stt.modelDownload": "Завантажити",
+  "settings.voice.page.stt.modelDelete": "Видалити модель",
+  "settings.voice.page.stt.modelRetry": "Повторити",
   "settings.voice.page.provider.wasm": "Локально",
   "settings.voice.page.stt.wasmModel": "Модель Whisper",
   "settings.voice.page.stt.wasmLoaded": "Модель завантажено",

packages/ui/src/lib/i18n/messages/uk.ts

@@ -1245,6 +1245,18 @@ export const dict: Record<I18nKey, string> = {
   'diffView.reviewDialog.actions.starting': 'Запуск...',
   'diffView.reviewDialog.toast.noSessionDirectory': 'Директорія сесії недоступна',
   'diffView.reviewDialog.toast.startFailed': 'Не вдалося запустити review flow',
+  'chat.dictation.start': 'Почати диктування',
+  'chat.dictation.overlayAria': 'Диктування',
+  'chat.dictation.downloadingModel': 'Завантаження мовної моделі...',
+  'chat.dictation.downloadingModelProgress': 'Завантаження мовної моделі... {percent}%',
+  'chat.dictation.listening': 'Слухаю...',
+  'chat.dictation.processing': 'Розшифровка...',
+  'chat.dictation.failed': 'Не вдалося розшифрувати запис',
+  'chat.dictation.cancel': 'Відхилити диктування',
+  'chat.dictation.insert': 'Вставити текст',
+  'chat.dictation.insertAndSend': 'Вставити й надіслати',
+  'chat.dictation.retry': 'Повторити розшифровку',
+  'chat.dictation.discard': 'Відхилити запис',
   'chat.history.loadOlder': 'Завантажити ще',
   'chat.autoReview.title': 'Цикл код-ревʼю триває',
   'chat.autoReview.status.waitingForReviewer': 'Очікуємо ревʼювера',

packages/ui/src/lib/i18n/messages/zh-CN.settings.ts

@@ -1485,16 +1485,33 @@ export const settingsDict = {
   'settings.notifications.page.toast.backgroundDisabled': '后台通知已关闭',
   'settings.notifications.page.testNotification.title': '测试通知',
   'settings.notifications.page.testNotification.body': '这是来自 OpenChamber 的测试通知。',
-  'settings.voice.page.section.voiceSetup': '语音设置',
   'settings.voice.page.section.speechRecognition': '语音识别',
+  'settings.voice.page.field.enableVoiceInput': '启用语音输入',
+  'settings.voice.page.field.enableVoiceInputAria': '启用语音输入（听写）',
   'settings.voice.page.section.playbackAndSummary': '播放',
-  'settings.voice.page.field.enableVoiceModeAria': '启用语音模式',
-  'settings.voice.page.field.enableVoiceMode': '启用语音模式',
   'settings.voice.page.field.provider': '提供方',
   'settings.voice.page.provider.browser': '浏览器',
   'settings.voice.page.provider.custom': '自定义',
   'settings.voice.page.provider.say': 'Say',
   'settings.voice.page.provider.server': '服务器',
+  'settings.voice.page.provider.local': '本地',
+  'settings.voice.page.tooltip.sttLocal': '在 OpenChamber 服务器上本地转写。模型自动下载，无需 API 密钥。',
+  'settings.voice.page.tooltip.localTts': '在 OpenChamber 服务器上本地合成语音（Kokoro，英语）。模型自动下载，无需 API 密钥。',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2（英语）',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3（25 种欧洲语言）',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base（多语言）',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny（多语言）',
+  'settings.voice.page.stt.badge.bestForEnglish': '英语最佳',
+  'settings.voice.page.stt.badge.bestForMultilingual': '多语言最佳',
+  'settings.voice.page.stt.meta.accuracy': '准确度',
+  'settings.voice.page.stt.meta.speed': '速度',
+  'settings.voice.page.stt.modelInstalled': '模型已安装',
+  'settings.voice.page.stt.modelDownloading': '正在下载模型...',
+  'settings.voice.page.stt.modelDownloadingProgress': '正在下载模型... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': '模型未下载',
+  'settings.voice.page.stt.modelDownload': '下载',
+  'settings.voice.page.stt.modelDelete': '删除模型',
+  'settings.voice.page.stt.modelRetry': '重试',
   'settings.voice.page.provider.wasm': '本地',
   'settings.voice.page.stt.wasmModel': 'Whisper 模型',
   'settings.voice.page.stt.wasmLoaded': '模型已加载',

packages/ui/src/lib/i18n/messages/zh-CN.ts

@@ -1839,6 +1839,18 @@ export const dict: Record<I18nKey, string> = {
   'chat.chatInput.placeholder.chat': '@ 用于文件/智能体；/ 用于命令和技能；! 用于 shell；# 用于代码片段',
   'chat.chatInput.placeholder.chatCompact': '使用 @ / ! # 辅助',
   'chat.chatInput.placeholder.selectSession': '选择或创建会话以开始聊天',
+  'chat.dictation.start': '开始语音输入',
+  'chat.dictation.overlayAria': '语音输入',
+  'chat.dictation.downloadingModel': '正在下载语音模型...',
+  'chat.dictation.downloadingModelProgress': '正在下载语音模型... {percent}%',
+  'chat.dictation.listening': '正在聆听...',
+  'chat.dictation.processing': '正在转写...',
+  'chat.dictation.failed': '转写失败',
+  'chat.dictation.cancel': '放弃语音输入',
+  'chat.dictation.insert': '插入文本',
+  'chat.dictation.insertAndSend': '插入并发送',
+  'chat.dictation.retry': '重试转写',
+  'chat.dictation.discard': '放弃录音',
   'chat.snippetAutocomplete.action.addNew': '+ 新建代码片段',
   'chat.snippetAutocomplete.empty': '未找到代码片段',
   'chat.snippetAutocomplete.footer': '↑↓ 导航 • Enter 选择 • Esc 关闭',

packages/ui/src/lib/i18n/messages/zh-TW.settings.ts

@@ -1401,16 +1401,33 @@
   'settings.notifications.page.toast.backgroundDisabled': '背景通知已關閉',
   'settings.notifications.page.testNotification.title': '測試通知',
   'settings.notifications.page.testNotification.body': '這是來自 OpenChamber 的測試通知。',
-  'settings.voice.page.section.voiceSetup': '語音設定',
   'settings.voice.page.section.speechRecognition': '語音辨識',
+  'settings.voice.page.field.enableVoiceInput': '啟用語音輸入',
+  'settings.voice.page.field.enableVoiceInputAria': '啟用語音輸入（聽寫）',
   'settings.voice.page.section.playbackAndSummary': '播放',
-  'settings.voice.page.field.enableVoiceModeAria': '啟用語音模式',
-  'settings.voice.page.field.enableVoiceMode': '啟用語音模式',
   'settings.voice.page.field.provider': '提供方',
   'settings.voice.page.provider.browser': '瀏覽器',
   'settings.voice.page.provider.custom': '自訂',
   'settings.voice.page.provider.say': 'Say',
   'settings.voice.page.provider.server': '伺服器',
+  'settings.voice.page.provider.local': '本機',
+  'settings.voice.page.tooltip.sttLocal': '在 OpenChamber 伺服器上本機轉寫。模型會自動下載，無需 API 金鑰。',
+  'settings.voice.page.tooltip.localTts': '在 OpenChamber 伺服器上本機合成語音（Kokoro，英文）。模型會自動下載，無需 API 金鑰。',
+  'settings.voice.page.stt.model.parakeetV2': 'Parakeet v2（英文）',
+  'settings.voice.page.stt.model.parakeetV3': 'Parakeet v3（25 種歐洲語言）',
+  'settings.voice.page.stt.model.whisperBase': 'Whisper base（多語言）',
+  'settings.voice.page.stt.model.whisperTiny': 'Whisper tiny（多語言）',
+  'settings.voice.page.stt.badge.bestForEnglish': '英文最佳',
+  'settings.voice.page.stt.badge.bestForMultilingual': '多語言最佳',
+  'settings.voice.page.stt.meta.accuracy': '準確度',
+  'settings.voice.page.stt.meta.speed': '速度',
+  'settings.voice.page.stt.modelInstalled': '模型已安裝',
+  'settings.voice.page.stt.modelDownloading': '正在下載模型...',
+  'settings.voice.page.stt.modelDownloadingProgress': '正在下載模型... {percent}%',
+  'settings.voice.page.stt.modelNotInstalled': '模型未下載',
+  'settings.voice.page.stt.modelDownload': '下載',
+  'settings.voice.page.stt.modelDelete': '刪除模型',
+  'settings.voice.page.stt.modelRetry': '重試',
   'settings.voice.page.provider.wasm': '本機',
   'settings.voice.page.stt.wasmModel': 'Whisper 模型',
   'settings.voice.page.stt.wasmLoaded': '模型已載入',

packages/ui/src/lib/i18n/messages/zh-TW.ts

@@ -1843,6 +1843,18 @@ export const dict: Record<I18nKey, string> = {
   'chat.chatInput.placeholder.chat': '@ 用於檔案/Agent；/ 用於命令；! 用於 shell；# 用於程式片段',
   'chat.chatInput.placeholder.chatCompact': '使用 @ / ! # 輔助',
   'chat.chatInput.placeholder.selectSession': '選擇或建立會話以開始聊天',
+  'chat.dictation.start': '開始語音輸入',
+  'chat.dictation.overlayAria': '語音輸入',
+  'chat.dictation.downloadingModel': '正在下載語音模型...',
+  'chat.dictation.downloadingModelProgress': '正在下載語音模型... {percent}%',
+  'chat.dictation.listening': '正在聆聽...',
+  'chat.dictation.processing': '正在轉寫...',
+  'chat.dictation.failed': '轉寫失敗',
+  'chat.dictation.cancel': '放棄語音輸入',
+  'chat.dictation.insert': '插入文字',
+  'chat.dictation.insertAndSend': '插入並傳送',
+  'chat.dictation.retry': '重試轉寫',
+  'chat.dictation.discard': '放棄錄音',
   'chat.snippetAutocomplete.action.addNew': '+ 新建程式片段',
   'chat.snippetAutocomplete.empty': '未找到程式片段',
   'chat.snippetAutocomplete.footer': '↑↓ 導航 • Enter 選擇 • Esc 關閉',

packages/ui/src/lib/persistence.ts

@@ -113,7 +113,10 @@ const persistToLocalStorage = (settings: DesktopSettings) => {
       localStorage.removeItem('opencode-update-toast-dismissed-version');
     }
   }
-  if (settings.sttProvider === 'browser' || settings.sttProvider === 'server') {
+  if (typeof settings.dictationEnabled === 'boolean') {
+    localStorage.setItem('dictationEnabled', String(settings.dictationEnabled));
+  }
+  if (settings.sttProvider === 'local' || settings.sttProvider === 'openai-compatible') {
     localStorage.setItem('sttProvider', settings.sttProvider);
   }
   if (typeof settings.sttServerUrl === 'string') {
@@ -122,18 +125,12 @@ const persistToLocalStorage = (settings: DesktopSettings) => {
   if (typeof settings.sttModel === 'string') {
     localStorage.setItem('sttModel', settings.sttModel);
   }
+  if (typeof settings.sttLocalModel === 'string') {
+    localStorage.setItem('sttLocalModel', settings.sttLocalModel);
+  }
   if (typeof settings.sttLanguage === 'string') {
     localStorage.setItem('sttLanguage', settings.sttLanguage);
   }
-  if (typeof settings.sttSilenceThresholdDb === 'number' && Number.isFinite(settings.sttSilenceThresholdDb)) {
-    localStorage.setItem('sttSilenceThresholdDb', String(settings.sttSilenceThresholdDb));
-  }
-  if (typeof settings.sttSilenceHoldMs === 'number' && Number.isFinite(settings.sttSilenceHoldMs)) {
-    localStorage.setItem('sttSilenceHoldMs', String(settings.sttSilenceHoldMs));
-  }
-  if (typeof settings.sttTranscribeOnStop === 'boolean') {
-    localStorage.setItem('sttTranscribeOnStop', String(settings.sttTranscribeOnStop));
-  }
 };
 
 const dispatchSettingsSynced = (settings: DesktopSettings): void => {
@@ -614,7 +611,10 @@ const applyDesktopUiPreferences = (settings: DesktopSettings) => {
   }
   if (configStoreApi && configStore) {
     const nextConfigState: Partial<typeof configStore> = {};
-    if ((settings.sttProvider === 'browser' || settings.sttProvider === 'server') && settings.sttProvider !== configStore.sttProvider) {
+    if (typeof settings.dictationEnabled === 'boolean' && settings.dictationEnabled !== configStore.dictationEnabled) {
+      nextConfigState.dictationEnabled = settings.dictationEnabled;
+    }
+    if ((settings.sttProvider === 'local' || settings.sttProvider === 'openai-compatible') && settings.sttProvider !== configStore.sttProvider) {
       nextConfigState.sttProvider = settings.sttProvider;
     }
     if (typeof settings.sttServerUrl === 'string' && settings.sttServerUrl !== configStore.sttServerUrl) {
@@ -623,18 +623,12 @@ const applyDesktopUiPreferences = (settings: DesktopSettings) => {
     if (typeof settings.sttModel === 'string' && settings.sttModel !== configStore.sttModel) {
       nextConfigState.sttModel = settings.sttModel;
     }
+    if (typeof settings.sttLocalModel === 'string' && settings.sttLocalModel !== configStore.sttLocalModel) {
+      nextConfigState.sttLocalModel = settings.sttLocalModel;
+    }
     if (typeof settings.sttLanguage === 'string' && settings.sttLanguage !== configStore.sttLanguage) {
       nextConfigState.sttLanguage = settings.sttLanguage;
     }
-    if (typeof settings.sttSilenceThresholdDb === 'number' && Number.isFinite(settings.sttSilenceThresholdDb) && settings.sttSilenceThresholdDb !== configStore.sttSilenceThresholdDb) {
-      nextConfigState.sttSilenceThresholdDb = settings.sttSilenceThresholdDb;
-    }
-    if (typeof settings.sttSilenceHoldMs === 'number' && Number.isFinite(settings.sttSilenceHoldMs) && settings.sttSilenceHoldMs !== configStore.sttSilenceHoldMs) {
-      nextConfigState.sttSilenceHoldMs = settings.sttSilenceHoldMs;
-    }
-    if (typeof settings.sttTranscribeOnStop === 'boolean' && settings.sttTranscribeOnStop !== configStore.sttTranscribeOnStop) {
-      nextConfigState.sttTranscribeOnStop = settings.sttTranscribeOnStop;
-    }
     if (Object.keys(nextConfigState).length > 0) {
       configStoreApi.setState(nextConfigState);
     }
@@ -1200,8 +1194,16 @@ const sanitizeWebSettings = (payload: unknown): DesktopSettings | null => {
   if (typeof candidate.responseStyleCustomInstructions === 'string') {
     result.responseStyleCustomInstructions = candidate.responseStyleCustomInstructions;
   }
-  if (candidate.sttProvider === 'browser' || candidate.sttProvider === 'server') {
+  if (typeof candidate.dictationEnabled === 'boolean') {
+    result.dictationEnabled = candidate.dictationEnabled;
+  }
+  if (candidate.sttProvider === 'local' || candidate.sttProvider === 'openai-compatible') {
     result.sttProvider = candidate.sttProvider;
+  } else if (candidate.sttProvider === 'server') {
+    // Legacy provider migration: 'server' was the OpenAI-compatible endpoint.
+    result.sttProvider = 'openai-compatible';
+  } else if (candidate.sttProvider === 'browser' || candidate.sttProvider === 'wasm') {
+    result.sttProvider = 'local';
   }
   if (typeof candidate.sttServerUrl === 'string') {
     result.sttServerUrl = candidate.sttServerUrl.trim();
@@ -1209,18 +1211,12 @@ const sanitizeWebSettings = (payload: unknown): DesktopSettings | null => {
   if (typeof candidate.sttModel === 'string') {
     result.sttModel = candidate.sttModel.trim();
   }
+  if (typeof candidate.sttLocalModel === 'string') {
+    result.sttLocalModel = candidate.sttLocalModel.trim();
+  }
   if (typeof candidate.sttLanguage === 'string') {
     result.sttLanguage = candidate.sttLanguage.trim();
   }
-  if (typeof candidate.sttSilenceThresholdDb === 'number' && Number.isFinite(candidate.sttSilenceThresholdDb)) {
-    result.sttSilenceThresholdDb = candidate.sttSilenceThresholdDb;
-  }
-  if (typeof candidate.sttSilenceHoldMs === 'number' && Number.isFinite(candidate.sttSilenceHoldMs)) {
-    result.sttSilenceHoldMs = candidate.sttSilenceHoldMs;
-  }
-  if (typeof candidate.sttTranscribeOnStop === 'boolean') {
-    result.sttTranscribeOnStop = candidate.sttTranscribeOnStop;
-  }
 
   return result;
 };

packages/ui/src/lib/settings/search.ts

@@ -678,22 +678,16 @@ const SETTINGS_SEARCH_ITEMS: readonly SettingsSearchItem[] = [
     keywords: ['keyboard', 'hotkeys', 'bindings'],
   },
   {
-    id: 'voice.voice-setup',
+    id: 'voice.playback',
     page: 'voice',
-    titleKey: 'settings.voice.page.section.voiceSetup',
+    titleKey: 'settings.voice.page.section.playbackAndSummary',
-    keywords: ['tts', 'voice mode', 'provider', 'speech rate', 'speech pitch', 'speech volume', 'language'],
+    keywords: ['tts', 'read aloud', 'voice', 'provider', 'speech rate', 'speech pitch', 'speech volume', 'tts input mode', 'markdown'],
   },
   {
     id: 'voice.speech-recognition',
     page: 'voice',
     titleKey: 'settings.voice.page.section.speechRecognition',
-    keywords: ['stt', 'transcribe', 'whisper', 'microphone', 'silence threshold'],
+    keywords: ['stt', 'dictation', 'voice input', 'transcribe', 'whisper', 'parakeet', 'microphone'],
-  },
-  {
-    id: 'voice.playback',
-    page: 'voice',
-    titleKey: 'settings.voice.page.section.playbackAndSummary',
-    keywords: ['read aloud', 'tts input mode', 'summary', 'markdown'],
   },
   {
     id: 'tunnel.provider',

packages/ui/src/lib/shortcuts.ts

@@ -327,6 +327,13 @@ const SHORTCUT_ACTIONS: ReadonlyArray<ShortcutAction> = [
     description: 'Toggle focus mode for the chat input',
     customizable: true,
   },
+  {
+    id: 'toggle_dictation',
+    defaultCombo: 'mod+alt+v',
+    label: 'Voice input',
+    description: 'Start dictation; press again to confirm and insert the transcript',
+    customizable: true,
+  },
   {
     id: 'abort_run',
     defaultCombo: 'escape',

packages/ui/src/lib/voice/audioStreamService.ts

@@ -1,397 +0,0 @@
-/**
- * Audio Stream Service
- *
- * Captures microphone audio using MediaRecorder, detects utterance boundaries
- * via an AnalyserNode-based silence detector (VAD), then POSTs each utterance
- * as a raw audio blob to the OpenChamber server's /api/stt/transcribe endpoint.
- *
- * Mimics the BrowserVoiceService.startListening interface so useBrowserVoice
- * can swap providers without changing its internal logic.
- *
- * @example
- * ```typescript
- * audioStreamService.configure({ baseURL: 'http://localhost:8001/v1', model: 'whisper-1' });
- * audioStreamService.startListening('en', (text, isFinal) => {
- *   if (isFinal) console.log('transcript:', text);
- * });
- * audioStreamService.stopListening();
- * ```
- */
-
-import { runtimeFetch } from '@/lib/runtime-fetch';
-
-type SpeechResultCallback = (text: string, isFinal: boolean) => void;
-type ErrorCallback = (error: string) => void;
-
-interface AudioStreamConfig {
-  /** Base URL of the OpenAI-compatible STT server (e.g. http://localhost:8001/v1) */
-  baseURL: string;
-  /** Whisper-compatible model name */
-  model: string;
-  /** Optional BCP-47 language hint (e.g. 'en'). Empty string = auto-detect. */
-  language?: string;
-  /**
-   * Silence threshold in dB below which audio is considered silence.
-   * Lower (more negative) = only very quiet audio counts as silence.
-   * Default: -45
-   */
-  silenceThresholdDb?: number;
-  /**
-   * How long continuous silence must last (ms) before the utterance is finalised.
-   * Default: 1500
-   */
-  silenceHoldMs?: number;
-  /** Optional API key for the STT server. */
-  apiKey?: string;
-}
-
-// How often (ms) the VAD samples the analyser
-const VAD_POLL_MS = 80;
-// Minimum audio duration (ms) to bother uploading (avoids blank clips)
-const MIN_UTTERANCE_MS = 300;
-
-class AudioStreamService {
-  private stream: MediaStream | null = null;
-  private mediaRecorder: MediaRecorder | null = null;
-  private audioContext: AudioContext | null = null;
-  private analyser: AnalyserNode | null = null;
-  private vadTimer: ReturnType<typeof setInterval> | null = null;
-  private chunks: Blob[] = [];
-  private recordingStartMs = 0;
-  private isActive = false;
-  private isSpeaking = false;
-  private silenceSince: number | null = null;
-  private onResult: SpeechResultCallback | null = null;
-  private onError: ErrorCallback | null = null;
-  private finishResolver: (() => void) | null = null;
-  private lang = 'en';
-
-  // Configurable parameters
-  private cfg: Required<AudioStreamConfig> = {
-    baseURL: '',
-    model: 'deepdml/faster-whisper-large-v3-turbo-ct2',
-    language: '',
-    silenceThresholdDb: -45,
-    silenceHoldMs: 1500,
-    apiKey: '',
-  };
-
-  /** Update service configuration. Can be called before or after startListening. */
-  configure(config: AudioStreamConfig): void {
-    this.cfg = {
-      silenceThresholdDb: -45,
-      silenceHoldMs: 1500,
-      language: '',
-      apiKey: '',
-      ...config,
-    };
-    this.cfg.apiKey = config.apiKey ?? '';
-  }
-
-  /** Whether the browser supports the required APIs. */
-  isSupported(): boolean {
-    return (
-      typeof window !== 'undefined' &&
-      typeof navigator !== 'undefined' &&
-      typeof navigator.mediaDevices?.getUserMedia === 'function' &&
-      typeof window.MediaRecorder !== 'undefined' &&
-      typeof window.AudioContext !== 'undefined'
-    );
-  }
-
-  /**
-   * Start listening. Requests microphone access if not already held.
-   * Calls onResult(text, true) for each completed utterance.
-   */
-  async startListening(
-    lang: string,
-    onResult: SpeechResultCallback,
-    onError?: ErrorCallback
-  ): Promise<void> {
-    if (this.isActive) {
-      this.stopListening();
-    }
-
-    this.lang = lang;
-    this.onResult = onResult;
-    this.onError = onError ?? null;
-    this.isActive = true;
-
-    try {
-      this.stream = await navigator.mediaDevices.getUserMedia({ audio: true, video: false });
-    } catch (err) {
-      this.isActive = false;
-      const msg = err instanceof Error ? err.message : 'Microphone access denied';
-      onError?.(msg);
-      return;
-    }
-
-    this._setupAudioContext();
-    this._startRecorder();
-    this._startVAD();
-  }
-
-  /** Stop listening and clean up all resources. */
-  stopListening(): void {
-    this._stopVAD();
-    this._stopRecorder();
-    this._cleanupAfterStop(true);
-  }
-
-  async finishListening(): Promise<void> {
-    if (!this.isActive) return;
-
-    this._stopVAD();
-    this.isSpeaking = false;
-    this.silenceSince = null;
-
-    if (!this.mediaRecorder || this.mediaRecorder.state === 'inactive') {
-      this._cleanupAfterStop(true);
-      return;
-    }
-
-    await new Promise<void>((resolve) => {
-      this.finishResolver = resolve;
-      this._finaliseUtterance(false);
-    });
-
-    this._cleanupAfterStop(true);
-  }
-
-  /** Whether currently listening. */
-  getIsListening(): boolean {
-    return this.isActive;
-  }
-
-  // ── Private helpers ──────────────────────────────────────────────────────
-
-  private _setupAudioContext(): void {
-    if (!this.stream) return;
-    const AudioContextClass = window.AudioContext ?? (window as unknown as { webkitAudioContext: typeof AudioContext }).webkitAudioContext;
-    this.audioContext = new AudioContextClass();
-    const source = this.audioContext.createMediaStreamSource(this.stream);
-    this.analyser = this.audioContext.createAnalyser();
-    this.analyser.fftSize = 512;
-    source.connect(this.analyser);
-  }
-
-  private _teardownAudioContext(): void {
-    try {
-      this.audioContext?.close();
-    } catch {
-      // ignore
-    }
-    this.audioContext = null;
-    this.analyser = null;
-  }
-
-  private _startRecorder(): void {
-    if (!this.stream) return;
-
-    const mimeType = this._pickMimeType();
-    const options: MediaRecorderOptions = {};
-    if (mimeType && MediaRecorder.isTypeSupported(mimeType)) {
-      options.mimeType = mimeType;
-    }
-
-    this.mediaRecorder = new MediaRecorder(this.stream, options);
-    this.chunks = [];
-    this.recordingStartMs = Date.now();
-
-    this.mediaRecorder.ondataavailable = (e) => {
-      if (e.data && e.data.size > 0) {
-        this.chunks.push(e.data);
-      }
-    };
-
-    this.mediaRecorder.onstop = () => {
-      const blobs = this.chunks.splice(0);
-      const durationMs = Date.now() - this.recordingStartMs;
-      if (blobs.length === 0 || durationMs < MIN_UTTERANCE_MS) {
-        this.finishResolver?.();
-        this.finishResolver = null;
-        return;
-      }
-
-      const mType = blobs[0].type || mimeType || 'audio/webm';
-      const blob = new Blob(blobs, { type: mType });
-      void this._upload(blob, mType).finally(() => {
-        this.finishResolver?.();
-        this.finishResolver = null;
-      });
-    };
-
-    // Collect data every 250 ms so we don't lose the tail on stop()
-    this.mediaRecorder.start(250);
-  }
-
-  private _stopRecorder(): void {
-    if (this.mediaRecorder && this.mediaRecorder.state !== 'inactive') {
-      try {
-        this.mediaRecorder.stop();
-      } catch {
-        this.finishResolver?.();
-        this.finishResolver = null;
-        // ignore
-      }
-    }
-    this.mediaRecorder = null;
-  }
-
-  private _releaseStream(): void {
-    if (this.stream) {
-      this.stream.getTracks().forEach((t) => t.stop());
-      this.stream = null;
-    }
-  }
-
-  private _startVAD(): void {
-    this._stopVAD();
-    this.silenceSince = null;
-    this.isSpeaking = false;
-
-    this.vadTimer = setInterval(() => {
-      if (!this.isActive || !this.analyser) return;
-      const db = this._getRmsDb();
-      const isSilent = db < this.cfg.silenceThresholdDb;
-
-      if (!isSilent) {
-        // Audio detected
-        this.silenceSince = null;
-        if (!this.isSpeaking) {
-          this.isSpeaking = true;
-          // Restart recorder to capture from the start of speech
-          if (this.mediaRecorder?.state === 'recording') {
-            this.recordingStartMs = Date.now();
-          }
-        }
-      } else {
-        // Silence detected
-        if (this.isSpeaking) {
-          if (this.silenceSince === null) {
-            this.silenceSince = Date.now();
-          } else if (Date.now() - this.silenceSince >= this.cfg.silenceHoldMs) {
-            // End of utterance — stop recorder (triggers onstop → upload)
-            this.isSpeaking = false;
-            this.silenceSince = null;
-            this._finaliseUtterance(true);
-          }
-        }
-      }
-    }, VAD_POLL_MS);
-  }
-
-  private _stopVAD(): void {
-    if (this.vadTimer !== null) {
-      clearInterval(this.vadTimer);
-      this.vadTimer = null;
-    }
-  }
-
-  private _cleanupAfterStop(clearChunks: boolean): void {
-    const pendingResolver = this.finishResolver;
-    this.isActive = false;
-    this.finishResolver = null;
-    this.mediaRecorder = null;
-    this._teardownAudioContext();
-    this._releaseStream();
-    if (clearChunks) {
-      this.chunks = [];
-    }
-    this.isSpeaking = false;
-    this.silenceSince = null;
-    this.onResult = null;
-    this.onError = null;
-    pendingResolver?.();
-  }
-
-  /** Stop the current recorder to flush the utterance, optionally restarting for the next one. */
-  private _finaliseUtterance(restart: boolean): void {
-    if (!this.isActive) return;
-    if (this.mediaRecorder && this.mediaRecorder.state === 'recording') {
-      this.mediaRecorder.stop();
-    }
-    if (!restart) return;
-
-    // Restart recorder for the next utterance after a short delay
-    // (MediaRecorder.onstop fires asynchronously; we wait for it to complete)
-    setTimeout(() => {
-      if (this.isActive && this.stream) {
-        this._startRecorder();
-      }
-    }, 100);
-  }
-
-  /** Compute RMS of current analyser frame in dBFS. */
-  private _getRmsDb(): number {
-    if (!this.analyser) return -Infinity;
-    const buf = new Float32Array(this.analyser.fftSize);
-    this.analyser.getFloatTimeDomainData(buf);
-    let sumSq = 0;
-    for (const s of buf) sumSq += s * s;
-    const rms = Math.sqrt(sumSq / buf.length);
-    return rms === 0 ? -Infinity : 20 * Math.log10(rms);
-  }
-
-  /** POST utterance blob to server, call onResult with transcript. */
-  private async _upload(blob: Blob, mimeType: string): Promise<void> {
-    if (!this.onResult) return;
-
-    try {
-      const headers: Record<string, string> = {
-        'Content-Type': mimeType,
-        'X-Base-URL': this.cfg.baseURL,
-        'X-Model': this.cfg.model,
-      };
-      if (this.cfg.apiKey) {
-        headers['Authorization'] = `Bearer ${this.cfg.apiKey}`;
-      }
-      if (this.cfg.language) {
-        headers['X-Language'] = this.cfg.language;
-      } else if (this.lang && this.lang !== 'auto') {
-        // Use BCP-47 base language code (e.g. 'en' from 'en-US')
-        const baseLang = this.lang.split('-')[0];
-        headers['X-Language'] = baseLang;
-      }
-
-      const response = await runtimeFetch('/api/stt/transcribe', {
-        method: 'POST',
-        headers,
-        body: blob,
-      });
-
-      if (!response.ok) {
-        const errData = await response.json().catch(() => ({ error: 'Unknown error' }));
-        throw new Error(errData.error ?? `HTTP ${response.status}`);
-      }
-
-      const data = await response.json();
-      const transcript: string = (data.transcript ?? '').trim();
-      if (transcript) {
-        this.onResult(transcript, true);
-      }
-    } catch (err) {
-      if (!this.isActive) return; // Stopped — ignore
-      const msg = err instanceof Error ? err.message : 'Transcription upload failed';
-      console.error('[AudioStreamService] Upload error:', msg);
-      this.onError?.(msg);
-    }
-  }
-
-  /** Pick the best supported MIME type for MediaRecorder. */
-  private _pickMimeType(): string {
-    const candidates = [
-      'audio/webm;codecs=opus',
-      'audio/webm',
-      'audio/ogg;codecs=opus',
-      'audio/ogg',
-      'audio/mp4',
-    ];
-    if (typeof MediaRecorder !== 'undefined' && MediaRecorder.isTypeSupported) {
-      return candidates.find((t) => MediaRecorder.isTypeSupported(t)) ?? '';
-    }
-    return '';
-  }
-}
-
-export const audioStreamService = new AudioStreamService();

packages/ui/src/lib/voice/contextFormatters.ts

@@ -1,131 +0,0 @@
-/**
- * Context formatters for voice-native output
- * Formats session events (messages, permissions, ready events) into natural language
- * for the ElevenLabs voice agent to speak aloud.
- *
- * @example
- * ```typescript
- * import { formatMessage, formatPermissionRequest } from '@/lib/voice';
- *
- * const voiceText = formatMessage({ role: 'assistant', content: 'Hello!' });
- * // Returns: "Claude Code: Hello!"
- * ```
- */
-
-import { VOICE_CONFIG } from "./voiceConfig";
-
-/** Message type for voice formatting */
-export interface VoiceMessage {
-    role: string;
-    content: string;
-}
-
-/**
- * Format a single message for voice output
- * - Assistant messages: Code blocks replaced with "[code block]", prefixed with "Claude Code: "
- * - User messages: Prefixed with "User: "
- * - Other roles: Returns null (not spoken)
- *
- * @param message - The message to format
- * @returns Formatted text for voice, or null if should not be spoken
- */
-function formatMessage(message: VoiceMessage): string | null {
-    // Handle edge cases
-    if (!message || typeof message.content !== "string") {
-        return null;
-    }
-
-    const content = message.content.trim();
-    if (!content) {
-        return null;
-    }
-
-    if (message.role === "assistant") {
-        // Replace code blocks with description (don't read code aloud)
-        const textOnly = content.replace(/```[\s\S]*?```/g, "[code block]");
-        return `Claude Code: ${textOnly}`;
-    }
-
-    if (message.role === "user") {
-        return `User: ${content}`;
-    }
-
-    // Skip system, tool, and other roles for voice
-    return null;
-}
-
-/**
- * Format multiple new messages for voice output
- * - Maps messages through formatMessage
- * - Filters out nulls (unspoken roles)
- * - Joins with newlines
- *
- * @param sessionId - The session ID (for future use/debugging)
- * @param messages - Array of messages to format
- * @returns Formatted text for voice, or null if no speakable messages
- */
-export function formatNewMessages(
-    sessionId: string,
-    messages: VoiceMessage[]
-): string | null {
-    // Handle edge cases
-    if (!Array.isArray(messages) || messages.length === 0) {
-        return null;
-    }
-
-    if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-        console.log(`[Voice] Formatting ${messages.length} messages for session ${sessionId}`);
-    }
-
-    // Format each message and filter out nulls
-    const formattedMessages = messages
-        .map(formatMessage)
-        .filter((msg): msg is string => msg !== null);
-
-    if (formattedMessages.length === 0) {
-        return null;
-    }
-
-    return formattedMessages.join("\n");
-}
-
-/**
- * Format a permission request for voice announcement
- * - Per CONTEXT.md: Only tool name, not arguments (LIMITED_TOOL_CALLS)
- * - Prompts user to say "allow" or "deny"
- *
- * @param sessionId - The session ID
- * @param requestId - The permission request ID
- * @param toolName - Name of the tool requesting permission
- * @param toolArgs - Tool arguments (not included in voice output per config)
- * @returns Formatted permission request for voice
- */
-export function formatPermissionRequest(
-    sessionId: string,
-    requestId: string,
-    toolName: string,
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
-    toolArgs: unknown
-): string {
-    if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-        console.log(`[Voice] Formatting permission request ${requestId} for session ${sessionId}`);
-    }
-
-    // Per VOICE_CONFIG.LIMITED_TOOL_CALLS, we don't include toolArgs in voice output
-    return `Claude Code is requesting permission to use ${toolName}. Say "allow" or "deny".`;
-}
-
-/**
- * Format a ready event for voice announcement
- * - Indicates the AI has finished working and is ready for next instruction
- *
- * @param sessionId - The session ID
- * @returns Formatted ready event for voice
- */
-export function formatReadyEvent(sessionId: string): string {
-    if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-        console.log(`[Voice] Formatting ready event for session ${sessionId}`);
-    }
-
-    return "Claude Code finished working. Ready for next instruction.";
-}

packages/ui/src/lib/voice/index.ts

@@ -1,17 +0,0 @@
-/**
- * Voice module barrel export
- * Provides a clean import path for voice session hooks.
- *
- * @example
- * ```typescript
- * import { voiceHooks } from '@/lib/voice';
- * ```
- */
-
-// Voice session registry (from voiceSession.ts)
-export {
-    isVoiceSessionStarted,
-} from "./voiceSession";
-
-// Voice hooks for session-to-voice event routing (from voiceHooks.ts)
-export { voiceHooks } from "./voiceHooks";

packages/ui/src/lib/voice/summarize.ts

@@ -6,15 +6,21 @@
 export function sanitizeForTTS(text: string): string {
     if (!text) return '';
     return text
-        // Remove code blocks
+        // Remove fenced code blocks entirely (multi-line code is unreadable),
+        // but keep inline-code CONTENT and only strip the backticks: agents
+        // routinely inline meaningful words ("You are on `main`").
         .replace(/```[\s\S]*?```/g, '')
-        .replace(/`[^`]*`/g, '')
+        .replace(/`([^`\n]*)`/g, '$1')
         // Remove markdown formatting
         .replace(/[*_~#]/g, '')
         // Remove URLs
         .replace(/https?:\/\/[^\s]+/g, '')
-        // Remove file paths
+        // Remove absolute file paths (leading slash, one or more segments).
-        .replace(/\/[\w\-./]+/g, '')
+        // Deliberately NOT matching interword slashes: "iOS/Android" and
+        // "origin/main" are speech, not paths.
+        .replace(/(^|\s)\/(?:[\w.-]+\/)*[\w.-]+/g, '$1')
+        // Read remaining interword slashes out loud ("iOS slash Android").
+        .replace(/([\w.])\/([\w.])/g, '$1 slash $2')
         // Remove shell-like patterns
         .replace(/^\s*[$#>]\s*/gm, '')
         // Remove brackets and special chars

packages/ui/src/lib/voice/voiceConfig.ts

@@ -1,32 +0,0 @@
-/**
- * Static voice context configuration
- * Controls voice behavior and feature flags for the ElevenLabs voice agent
- */
-export const VOICE_CONFIG = {
-    /** Disable all tool call information from being sent to voice context */
-    DISABLE_TOOL_CALLS: false,
-
-    /** Send only tool names and descriptions, exclude arguments */
-    LIMITED_TOOL_CALLS: true,
-
-    /** Disable permission request forwarding */
-    DISABLE_PERMISSION_REQUESTS: false,
-
-    /** Disable session online/offline notifications */
-    DISABLE_SESSION_STATUS: true,
-
-    /** Disable message forwarding */
-    DISABLE_MESSAGES: false,
-
-    /** Disable session focus notifications */
-    DISABLE_SESSION_FOCUS: false,
-
-    /** Disable ready event notifications */
-    DISABLE_READY_EVENTS: false,
-
-    /** Maximum number of messages to include in session history */
-    MAX_HISTORY_MESSAGES: 50,
-
-    /** Enable debug logging for voice context updates */
-    ENABLE_DEBUG_LOGGING: true,
-} as const;

packages/ui/src/lib/voice/voiceHooks.ts

@@ -1,125 +0,0 @@
-/**
- * Voice hooks for session-to-voice event routing
- * Routes session events (messages, permissions, ready events) to the ElevenLabs
- * voice agent via contextual updates.
- *
- * This module provides hooks that can be called when session events occur,
- * using the voice session registry from voiceSession.ts.
- *
- * @example
- * ```typescript
- * import { voiceHooks } from '@/lib/voice';
- *
- * // Route session messages to voice
- * voiceHooks.onMessages(sessionId, messages);
- * ```
- */
-
-import { VOICE_CONFIG } from "./voiceConfig";
-import {
-    formatNewMessages,
-    formatPermissionRequest,
-    formatReadyEvent,
-    type VoiceMessage,
-} from "./contextFormatters";
-import { getVoiceSession, isVoiceSessionStarted } from "./voiceSession";
-
-/**
- * Report a contextual update to the voice session
- * Internal helper that checks preconditions and handles errors
- *
- * @param update - The text update to send, or null/undefined to skip
- */
-function reportContextualUpdate(update: string | null | undefined): void {
-    // Skip empty/null/undefined updates
-    if (!update || update.trim().length === 0) {
-        return;
-    }
-
-    // Skip if no voice session or not started
-    const voiceSession = getVoiceSession();
-    if (!voiceSession || !isVoiceSessionStarted()) {
-        if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-            console.log("[Voice] Skipping contextual update - no active session");
-        }
-        return;
-    }
-
-    try {
-        if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-            console.log("[Voice] Sending contextual update:", update.substring(0, 100));
-        }
-        voiceSession.sendContextualUpdate(update);
-    } catch (error) {
-        // Log error but don't throw - voice updates shouldn't break the app
-        console.error("[Voice] Failed to send contextual update:", error);
-    }
-}
-
-/**
- * Voice hooks - exported functions to route session events to voice
- *
- * These hooks should be called when corresponding session events occur.
- * They respect VOICE_CONFIG feature flags to enable/disable specific
- * event types.
- */
-export const voiceHooks = {
-    /**
-     * Called when new messages arrive in the session
-     * Formats and sends messages to voice agent (if not disabled)
-     *
-     * @param sessionId - The session ID
-     * @param messages - Array of messages to format and send
-     */
-    onMessages(sessionId: string, messages: VoiceMessage[]): void {
-        if (VOICE_CONFIG.DISABLE_MESSAGES) {
-            if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-                console.log("[Voice] Message forwarding disabled");
-            }
-            return;
-        }
-        reportContextualUpdate(formatNewMessages(sessionId, messages));
-    },
-
-    /**
-     * Called when a permission request is made
-     * Announces the permission request to the voice agent (if not disabled)
-     *
-     * @param sessionId - The session ID
-     * @param requestId - The permission request ID
-     * @param toolName - Name of the tool requesting permission
-     * @param toolArgs - Arguments for the tool (not sent to voice per LIMITED_TOOL_CALLS)
-     */
-    onPermissionRequested(
-        sessionId: string,
-        requestId: string,
-        toolName: string,
-        toolArgs: unknown
-    ): void {
-        if (VOICE_CONFIG.DISABLE_PERMISSION_REQUESTS) {
-            if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-                console.log("[Voice] Permission request forwarding disabled");
-            }
-            return;
-        }
-        reportContextualUpdate(
-            formatPermissionRequest(sessionId, requestId, toolName, toolArgs)
-        );
-    },
-
-    /**
-     * Called when the AI is ready for the next instruction
-     * Announces ready state to the voice agent (if not disabled)
-     *
-     * @param sessionId - The session ID
-     */
-    onReady(sessionId: string): void {
-        if (VOICE_CONFIG.DISABLE_READY_EVENTS) {
-            if (VOICE_CONFIG.ENABLE_DEBUG_LOGGING) {
-                console.log("[Voice] Ready event forwarding disabled");
-            }
-            return;
-        }
-        reportContextualUpdate(formatReadyEvent(sessionId));
-    },
-};

packages/ui/src/lib/voice/voiceSession.ts

@@ -1,28 +0,0 @@
-/**
- * Voice session interface
- * Used for type safety without importing ReturnType from SDK
- */
-interface VoiceSession {
-    sendContextualUpdate: (text: string) => void;
-}
-
-/**
- * Global storage for the active voice session.
- * Used by voiceHooks to send contextual updates to the voice agent.
- */
-const activeVoiceSession: VoiceSession | null = null;
-
-/**
- * Get the currently registered voice session.
- * Used by voiceHooks to send contextual updates.
- */
-export function getVoiceSession(): VoiceSession | null {
-    return activeVoiceSession;
-}
-
-/**
- * Check if a voice session is currently active.
- */
-export function isVoiceSessionStarted(): boolean {
-    return activeVoiceSession !== null;
-}

packages/ui/src/lib/voice/wasmSttService.ts

@@ -1,555 +0,0 @@
-/**
- * WASM Speech-to-Text Service
- *
- * Local Whisper transcription via Transformers.js (ONNX Runtime Web).
- * Captures microphone audio, detects utterance boundaries via silence-based
- * VAD, then transcribes each utterance locally — no cloud API required.
- *
- * Works in Electron and all modern browsers that support Web Audio API.
- * First use downloads a Whisper model (~40–166 MB, cached).
- */
-
-export type WasmModelStatus =
-  | { state: 'unloaded' }
-  | { state: 'downloading'; progress: number }
-  | { state: 'loading' }
-  | { state: 'ready' }
-  | { state: 'error'; error: string };
-
-export interface WasmModelInfo {
-  id: string;
-  name: string;
-  size: string;
-  languages: string;
-  description: string;
-}
-
-export const WASM_MODELS: WasmModelInfo[] = [
-  {
-    id: 'Xenova/whisper-tiny.en',
-    name: 'Whisper Tiny (EN)',
-    size: '~39 MB',
-    languages: 'English',
-    description: 'Fastest, lowest accuracy. Good for quick dictation.',
-  },
-  {
-    id: 'Xenova/whisper-base.en',
-    name: 'Whisper Base (EN)',
-    size: '~73 MB',
-    languages: 'English',
-    description: 'Balanced speed and accuracy. Default for English.',
-  },
-  {
-    id: 'Xenova/whisper-small.en',
-    name: 'Whisper Small (EN)',
-    size: '~166 MB',
-    languages: 'English',
-    description: 'Higher accuracy, slower. Best for noisy environments.',
-  },
-];
-
-type SpeechResultCallback = (text: string, isFinal: boolean) => void;
-type ErrorCallback = (error: string) => void;
-
-const VAD_POLL_MS = 80;
-const MIN_UTTERANCE_MS = 300;
-const WHISPER_SAMPLE_RATE = 16000;
-
-interface WasmSttConfig {
-  silenceThresholdDb?: number;
-  silenceHoldMs?: number;
-}
-
-class WasmSttService {
-  private transcriber: unknown = null;
-  private worker: Worker | null = null;
-  private modelStatus: WasmModelStatus = { state: 'unloaded' };
-  private currentModelId: string | null = null;
-
-  private stream: MediaStream | null = null;
-  private mediaRecorder: MediaRecorder | null = null;
-  private audioContext: AudioContext | null = null;
-  private analyser: AnalyserNode | null = null;
-  private vadTimer: ReturnType<typeof setInterval> | null = null;
-  private chunks: Blob[] = [];
-  private recordingStartMs = 0;
-  private isActive = false;
-  private isSpeaking = false;
-  private silenceSince: number | null = null;
-  private onResult: SpeechResultCallback | null = null;
-  private onError: ErrorCallback | null = null;
-  private finishResolver: (() => void) | null = null;
-  private lang = 'en';
-
-  private cfg: Required<WasmSttConfig> = {
-    silenceThresholdDb: -45,
-    silenceHoldMs: 1500,
-  };
-
-  public onModelStatusChange: ((status: WasmModelStatus) => void) | null = null;
-
-  configure(config: WasmSttConfig): void {
-    this.cfg = { ...this.cfg, ...config };
-  }
-
-  isSupported(): boolean {
-    return (
-      typeof window !== 'undefined' &&
-      typeof navigator !== 'undefined' &&
-      typeof navigator.mediaDevices?.getUserMedia === 'function' &&
-      typeof window.MediaRecorder !== 'undefined' &&
-      typeof window.AudioContext !== 'undefined'
-    );
-  }
-
-  getModelStatus(): WasmModelStatus {
-    return this.modelStatus;
-  }
-
-  getCurrentModelId(): string | null {
-    return this.currentModelId;
-  }
-
-  private setModelStatus(status: WasmModelStatus): void {
-    this.modelStatus = status;
-    this.onModelStatusChange?.(status);
-  }
-
-  async loadModel(modelId: string): Promise<void> {
-    if (this.currentModelId === modelId && this.modelStatus.state === 'ready') {
-      return;
-    }
-
-    if (this.modelStatus.state === 'downloading' || this.modelStatus.state === 'loading') {
-      return;
-    }
-
-    this._terminateWorker();
-    this.transcriber = null;
-
-    this.setModelStatus({ state: 'downloading', progress: 0 });
-    this.currentModelId = modelId;
-
-    // Try Web Worker first — inference off main thread = no UI freeze.
-    try {
-      const WasmWorkerMod = await import('./wasmSttWorker?worker');
-      const WasmWorker = WasmWorkerMod.default as new () => Worker;
-      this.worker = new WasmWorker();
-
-      await new Promise<void>((resolve, reject) => {
-        const timer = setTimeout(() => reject(new Error('Worker init timed out')), 10000);
-
-        this.worker!.onmessage = (e: MessageEvent) => {
-          const data = e.data as { type: string; progress?: number; error?: string; text?: string };
-          if (data.type === 'progress') {
-            this.setModelStatus({ state: 'downloading', progress: data.progress ?? 0 });
-          } else if (data.type === 'loaded') {
-            clearTimeout(timer);
-            resolve();
-          } else if (data.type === 'error') {
-            clearTimeout(timer);
-            reject(new Error(data.error ?? 'Worker load failed'));
-          }
-        };
-
-        this.worker!.onerror = (err) => {
-          clearTimeout(timer);
-          reject(new Error(err.message || 'Worker error'));
-        };
-
-        this.worker!.postMessage({ type: 'load', modelId });
-      });
-
-      this.setModelStatus({ state: 'ready' });
-      return;
-    } catch (err) {
-      console.warn('[WasmStt] Worker failed, using main-thread:', err instanceof Error ? err.message : err);
-      this._terminateWorker();
-    }
-
-    // Fallback: main-thread pipeline (causes brief UI freeze during inference).
-    try {
-      const { pipeline, env } = await import('@xenova/transformers');
-      env.backends.onnx.wasm.numThreads = 1;
-      env.allowLocalModels = false;
-
-      const fileDoneBytes = new Map<string, number>();
-      let totalDone = 0;
-      let totalEstimate = 0;
-
-      this.transcriber = await pipeline('automatic-speech-recognition', modelId, {
-        progress_callback: (info: { status?: string; file?: string; loaded?: number; total?: number }) => {
-          if (info.status === 'progress' && info.file) {
-            const prevDone = fileDoneBytes.get(info.file) ?? 0;
-            const currentDone = info.loaded ?? 0;
-            const delta = Math.max(0, currentDone - prevDone);
-            fileDoneBytes.set(info.file, currentDone);
-            totalDone += delta;
-            if (info.total && info.total > totalEstimate) totalEstimate = info.total;
-            const effectiveTotal = Math.max(totalEstimate, totalDone);
-            const pct = effectiveTotal > 0 ? Math.min(100, Math.round((totalDone / effectiveTotal) * 100)) : 0;
-            this.setModelStatus({ state: 'downloading', progress: pct });
-          }
-        },
-      });
-
-      this.setModelStatus({ state: 'ready' });
-    } catch (err) {
-      const msg = err instanceof Error ? err.message : 'Unknown error loading model';
-      this.setModelStatus({ state: 'error', error: msg });
-      this.transcriber = null;
-      this.currentModelId = null;
-      throw err;
-    }
-  }
-
-  private _terminateWorker(): void {
-    if (this.worker) {
-      this.worker.terminate();
-      this.worker = null;
-    }
-  }
-
-  async unloadModel(): Promise<void> {
-    this._terminateWorker();
-    this.transcriber = null;
-    this.currentModelId = null;
-    this.setModelStatus({ state: 'unloaded' });
-  }
-
-  async startListening(
-    lang: string,
-    onResult: SpeechResultCallback,
-    onError?: ErrorCallback,
-  ): Promise<void> {
-    if (this.isActive) {
-      this.stopListening();
-    }
-
-    if (!this.transcriber && !this.worker) {
-      onError?.('Whisper model not loaded. Select a model in Voice Settings first.');
-      return;
-    }
-
-    this.lang = lang;
-    this.onResult = onResult;
-    this.onError = onError ?? null;
-    this.isActive = true;
-
-    try {
-      this.stream = await navigator.mediaDevices.getUserMedia({ audio: true, video: false });
-    } catch (err) {
-      this.isActive = false;
-      const msg = err instanceof Error ? err.message : 'Microphone access denied';
-      onError?.(msg);
-      return;
-    }
-
-    this._setupAudioContext();
-    this._startRecorder();
-    this._startVAD();
-  }
-
-  stopListening(): void {
-    this._stopVAD();
-    if (this.mediaRecorder && this.mediaRecorder.state === 'recording') {
-      try { this.mediaRecorder.stop(); } catch { /* ignore */ }
-    }
-    this._cleanupAfterStop(true);
-  }
-
-  async finishListening(): Promise<void> {
-    if (!this.isActive) return;
-
-    this._stopVAD();
-    this.isSpeaking = false;
-    this.silenceSince = null;
-
-    if (!this.mediaRecorder || this.mediaRecorder.state === 'inactive') {
-      this._cleanupAfterStop(true);
-      return;
-    }
-
-    await new Promise<void>((resolve) => {
-      this.finishResolver = resolve;
-      this._finaliseUtterance(false);
-    });
-
-    this._cleanupAfterStop(true);
-  }
-
-  getIsListening(): boolean {
-    return this.isActive;
-  }
-
-  // ── Audio capture ────────────────────────────────────────────────────
-
-  private _setupAudioContext(): void {
-    if (!this.stream) return;
-    const AudioContextClass = window.AudioContext ?? (window as unknown as { webkitAudioContext: typeof AudioContext }).webkitAudioContext;
-    this.audioContext = new AudioContextClass();
-    const source = this.audioContext.createMediaStreamSource(this.stream);
-    this.analyser = this.audioContext.createAnalyser();
-    this.analyser.fftSize = 512;
-    source.connect(this.analyser);
-  }
-
-  private _teardownAudioContext(): void {
-    try { this.audioContext?.close(); } catch { /* ignore */ }
-    this.audioContext = null;
-    this.analyser = null;
-  }
-
-  private _startRecorder(): void {
-    if (!this.stream) return;
-    const mimeType = this._pickMimeType();
-    const options: MediaRecorderOptions = {};
-    if (mimeType && MediaRecorder.isTypeSupported(mimeType)) {
-      options.mimeType = mimeType;
-    }
-    this.mediaRecorder = new MediaRecorder(this.stream, options);
-    this.chunks = [];
-    this.recordingStartMs = Date.now();
-
-    this.mediaRecorder.ondataavailable = (e) => {
-      if (e.data && e.data.size > 0) {
-        this.chunks.push(e.data);
-      }
-    };
-
-    this.mediaRecorder.onstop = () => {
-      const blobs = this.chunks.splice(0);
-      const durationMs = Date.now() - this.recordingStartMs;
-      if (blobs.length === 0 || durationMs < MIN_UTTERANCE_MS) {
-        this.finishResolver?.();
-        this.finishResolver = null;
-        return;
-      }
-      const mType = blobs[0].type || mimeType || 'audio/webm';
-      const blob = new Blob(blobs, { type: mType });
-      void this._transcribe(blob).finally(() => {
-        this.finishResolver?.();
-        this.finishResolver = null;
-      });
-    };
-
-    this.mediaRecorder.start(250);
-  }
-
-  private _releaseStream(): void {
-    if (this.stream) {
-      this.stream.getTracks().forEach((t) => t.stop());
-      this.stream = null;
-    }
-  }
-
-  // ── VAD ──────────────────────────────────────────────────────────────
-
-  private _startVAD(): void {
-    this._stopVAD();
-    this.silenceSince = null;
-    this.isSpeaking = false;
-
-    this.vadTimer = setInterval(() => {
-      if (!this.isActive || !this.analyser) return;
-      const db = this._getRmsDb();
-      const isSilent = db < this.cfg.silenceThresholdDb;
-
-      if (!isSilent) {
-        this.silenceSince = null;
-        if (!this.isSpeaking) {
-          this.isSpeaking = true;
-          if (this.mediaRecorder?.state === 'recording') {
-            this.recordingStartMs = Date.now();
-          }
-        }
-      } else {
-        if (this.isSpeaking) {
-          if (this.silenceSince === null) {
-            this.silenceSince = Date.now();
-          } else if (Date.now() - this.silenceSince >= this.cfg.silenceHoldMs) {
-            this.isSpeaking = false;
-            this.silenceSince = null;
-            this._finaliseUtterance(true);
-          }
-        }
-      }
-    }, VAD_POLL_MS);
-  }
-
-  private _stopVAD(): void {
-    if (this.vadTimer !== null) {
-      clearInterval(this.vadTimer);
-      this.vadTimer = null;
-    }
-  }
-
-  private _cleanupAfterStop(clearChunks: boolean): void {
-    const pendingResolver = this.finishResolver;
-    this.isActive = false;
-    this.finishResolver = null;
-    this.mediaRecorder = null;
-    this._teardownAudioContext();
-    this._releaseStream();
-    if (clearChunks) this.chunks = [];
-    this.isSpeaking = false;
-    this.silenceSince = null;
-    this.onResult = null;
-    this.onError = null;
-    pendingResolver?.();
-  }
-
-  private _finaliseUtterance(restart: boolean): void {
-    if (!this.isActive) return;
-    if (this.mediaRecorder && this.mediaRecorder.state === 'recording') {
-      this.mediaRecorder.stop();
-    }
-    if (!restart) return;
-    setTimeout(() => {
-      if (this.isActive && this.stream) {
-        this._startRecorder();
-      }
-    }, 100);
-  }
-
-  private _getRmsDb(): number {
-    if (!this.analyser) return -Infinity;
-    const buf = new Float32Array(this.analyser.fftSize);
-    this.analyser.getFloatTimeDomainData(buf);
-    let sumSq = 0;
-    for (const s of buf) sumSq += s * s;
-    const rms = Math.sqrt(sumSq / buf.length);
-    return rms === 0 ? -Infinity : 20 * Math.log10(rms);
-  }
-
-  // ── Transcription ────────────────────────────────────────────────────
-
-  private async _transcribe(blob: Blob): Promise<void> {
-    if (!this.onResult) return;
-    if (!this.transcriber && !this.worker) {
-      this.onError?.('Model not loaded');
-      return;
-    }
-
-    try {
-      const audioData = await this._decodeToFloat32(blob);
-      if (!audioData || audioData.length === 0) {
-        this.onError?.(`Failed to decode audio (${blob.size} bytes)`);
-        return;
-      }
-
-      const langHint = this._resolveLanguageHint();
-
-      // Prefer worker (non-blocking); fall back to main-thread pipeline.
-      const transcript = this.worker
-        ? await this._transcribeViaWorker(audioData, langHint)
-        : await this._transcribeMainThread(audioData, langHint);
-
-      if (transcript) {
-        this.onResult(transcript, true);
-      }
-    } catch (err) {
-      if (!this.isActive) return;
-      const msg = err instanceof Error ? err.message : 'Local transcription failed';
-      this.onError?.(msg);
-    }
-  }
-
-  private _transcribeViaWorker(audioData: Float32Array, langHint: string | undefined): Promise<string> {
-    return new Promise((resolve, reject) => {
-      if (!this.worker) return reject(new Error('Worker gone'));
-
-      const onMessage = (e: MessageEvent) => {
-        const data = e.data as { type: string; error?: string; transcript?: string; text?: string };
-        if (data.type === 'result') {
-          this.worker!.removeEventListener('message', onMessage);
-          resolve(data.transcript ?? '');
-        } else if (data.type === 'log') {
-          console.log('[WasmStt Worker]', data.text);
-        } else if (data.type === 'error') {
-          this.worker!.removeEventListener('message', onMessage);
-          reject(new Error(data.error ?? 'Transcription failed'));
-        }
-      };
-
-      this.worker.addEventListener('message', onMessage);
-      this.worker.postMessage(
-        { type: 'transcribe', audio: audioData.buffer, language: langHint },
-        [audioData.buffer],
-      );
-
-      setTimeout(() => {
-        this.worker?.removeEventListener('message', onMessage);
-        reject(new Error('Transcription timed out'));
-      }, 30000);
-    });
-  }
-
-  private async _transcribeMainThread(audioData: Float32Array, langHint: string | undefined): Promise<string> {
-    const pipelineFn = this.transcriber as (
-      input: Float32Array,
-      options?: Record<string, unknown>,
-    ) => Promise<{ text: string }>;
-
-    const result = await pipelineFn(audioData, {
-      task: 'transcribe',
-      ...(langHint ? { language: langHint } : {}),
-    });
-
-    return (result?.text ?? '').trim();
-  }
-
-  private async _decodeToFloat32(blob: Blob): Promise<Float32Array | null> {
-    if (!this.audioContext) return null;
-    const arrayBuffer = await blob.arrayBuffer();
-    let audioBuffer: AudioBuffer;
-    try {
-      audioBuffer = await this.audioContext.decodeAudioData(arrayBuffer);
-    } catch {
-      return null;
-    }
-
-    const origRate = audioBuffer.sampleRate;
-    const origData = audioBuffer.getChannelData(0);
-    const targetRate = WHISPER_SAMPLE_RATE;
-
-    if (origRate === targetRate) {
-      return new Float32Array(origData);
-    }
-
-    const ratio = origRate / targetRate;
-    const newLength = Math.ceil(origData.length / ratio);
-    const result = new Float32Array(newLength);
-    for (let i = 0; i < newLength; i++) {
-      const origIdx = i * ratio;
-      const idx0 = Math.floor(origIdx);
-      const idx1 = Math.min(idx0 + 1, origData.length - 1);
-      const frac = origIdx - idx0;
-      result[i] = origData[idx0] * (1 - frac) + origData[idx1] * frac;
-    }
-    return result;
-  }
-
-  private _resolveLanguageHint(): string | undefined {
-    if (this.lang && this.lang !== 'auto') {
-      return this.lang.split('-')[0];
-    }
-    return undefined;
-  }
-
-  private _pickMimeType(): string {
-    const candidates = [
-      'audio/webm;codecs=opus',
-      'audio/webm',
-      'audio/ogg;codecs=opus',
-      'audio/ogg',
-      'audio/mp4',
-    ];
-    if (typeof MediaRecorder !== 'undefined' && MediaRecorder.isTypeSupported) {
-      return candidates.find((t) => MediaRecorder.isTypeSupported(t)) ?? '';
-    }
-    return '';
-  }
-}
-
-export const wasmSttService = new WasmSttService();

packages/ui/src/lib/voice/wasmSttWorker.ts

@@ -1,92 +0,0 @@
-/**
- * Web Worker for off-main-thread Whisper transcription.
- *
- * Receives `{ type: 'load', modelId }` to load a model, then
- * `{ type: 'transcribe', audio: Float32Array (transferred buffer), language? }`
- * to run inference. Posts progress, results, and errors back.
- */
-
-import { pipeline, env } from '@xenova/transformers';
-
-let transcriber: unknown = null;
-
-self.onmessage = async (e: MessageEvent) => {
-  const { type } = e.data as { type: string };
-
-  if (type === 'load') {
-    const { modelId } = e.data as { modelId: string };
-    try {
-      env.backends.onnx.wasm.numThreads = 1;
-
-      const fileDoneBytes = new Map<string, number>();
-      let totalDone = 0;
-      let totalEstimate = 0;
-
-      transcriber = await pipeline('automatic-speech-recognition', modelId, {
-        progress_callback: (info: { status?: string; file?: string; loaded?: number; total?: number }) => {
-          if (info.status === 'progress' && info.file) {
-            const prevDone = fileDoneBytes.get(info.file) ?? 0;
-            const currentDone = info.loaded ?? 0;
-            const delta = Math.max(0, currentDone - prevDone);
-            fileDoneBytes.set(info.file, currentDone);
-            totalDone += delta;
-
-            if (info.total && info.total > totalEstimate) {
-              totalEstimate = info.total;
-            }
-
-            const effectiveTotal = Math.max(totalEstimate, totalDone);
-            const pct = effectiveTotal > 0 ? Math.min(100, Math.round((totalDone / effectiveTotal) * 100)) : 0;
-            self.postMessage({ type: 'progress', progress: pct });
-          }
-        },
-      });
-
-      self.postMessage({ type: 'loaded' });
-    } catch (err) {
-      self.postMessage({
-        type: 'error',
-        error: err instanceof Error ? err.message : 'Failed to load model',
-      });
-    }
-  } else if (type === 'transcribe') {
-    if (!transcriber) {
-      self.postMessage({ type: 'error', error: 'Model not loaded', seq: (e.data as { seq?: number }).seq });
-      return;
-    }
-
-    const { audio, language, seq } = e.data as { audio: ArrayBuffer; language?: string; seq?: number };
-
-    try {
-      const samples = new Float32Array(audio);
-      if (samples.length === 0) {
-        self.postMessage({ type: 'error', error: 'Empty audio received', seq });
-        return;
-      }
-
-      self.postMessage({ type: 'log', text: `Transcribing ${samples.length} samples (${(samples.length / 16000).toFixed(1)}s)` });
-
-      const pipelineFn = transcriber as (
-        input: Float32Array,
-        options?: Record<string, unknown>,
-      ) => Promise<{ text: string }>;
-
-      const result = await pipelineFn(samples, {
-        task: 'transcribe',
-        ...(language ? { language } : {}),
-      });
-
-      self.postMessage({
-        type: 'result',
-        transcript: (result?.text ?? '').trim(),
-        seq,
-      });
-    } catch (err) {
-      self.postMessage({
-        type: 'error',
-        error: err instanceof Error ? err.message : 'Transcription failed',
-        seq,
-      });
-    }
-  }
-};

packages/ui/src/stores/useConfigStore.ts

@@ -23,10 +23,6 @@ import { getSyncConfig, subscribeToSyncConfigChanges } from "@/sync/sync-refs";
 
 const MODELS_DEV_API_URL = "https://models.dev/api.json";
 const MODELS_DEV_PROXY_URL = "/api/openchamber/models-metadata";
-const STT_SILENCE_THRESHOLD_DB_MIN = -100;
-const STT_SILENCE_THRESHOLD_DB_MAX = 0;
-const STT_SILENCE_HOLD_MS_MIN = 250;
-const STT_SILENCE_HOLD_MS_MAX = 10000;
 
 const FALLBACK_PROVIDER_ID = "opencode";
 const FALLBACK_MODEL_ID = "big-pickle";
@@ -38,18 +34,19 @@ const GIT_UTILITY_PROVIDER_ID = "zen";
 const GIT_UTILITY_PREFERRED_MODEL_ID = "big-pickle";
 const PROVIDER_CONFIG_REFRESH_CONCURRENCY = 4;
 
-const normalizeSttSilenceThresholdDb = (value: unknown): number | undefined => {
+const normalizeSttProvider = (value: unknown): 'local' | 'openai-compatible' | undefined => {
-    if (typeof value !== 'number' || !Number.isFinite(value)) {
+    if (value === 'local' || value === 'openai-compatible') {
-        return undefined;
+        return value;
     }
-    return Math.max(STT_SILENCE_THRESHOLD_DB_MIN, Math.min(STT_SILENCE_THRESHOLD_DB_MAX, value));
+    // Legacy providers: 'server' used an OpenAI-compatible endpoint;
-};
+    // 'browser' and 'wasm' map to the local default.
-
+    if (value === 'server') {
-const normalizeSttSilenceHoldMs = (value: unknown): number | undefined => {
+        return 'openai-compatible';
-    if (typeof value !== 'number' || !Number.isFinite(value)) {
-        return undefined;
     }
-    return Math.max(STT_SILENCE_HOLD_MS_MIN, Math.min(STT_SILENCE_HOLD_MS_MAX, Math.round(value)));
+    if (value === 'browser' || value === 'wasm') {
+        return 'local';
+    }
+    return undefined;
 };
 
 interface OpenChamberDefaults {
@@ -61,13 +58,11 @@ interface OpenChamberDefaults {
     defaultFileViewerPreview?: boolean;
     zenModel?: string;
     messageStreamTransport?: 'auto' | 'ws' | 'sse';
-    sttProvider?: 'browser' | 'server' | 'wasm';
+    sttProvider?: 'local' | 'openai-compatible';
     sttServerUrl?: string;
-    wasmSttModel?: string;
     sttModel?: string;
+    sttLocalModel?: string;
     sttLanguage?: string;
-    sttSilenceThresholdDb?: number;
-    sttSilenceHoldMs?: number;
 }
 
 const fetchOpenChamberDefaults = async (): Promise<OpenChamberDefaults> => {
@@ -101,12 +96,11 @@ const fetchOpenChamberDefaults = async (): Promise<OpenChamberDefaults> => {
                         data?.messageStreamTransport === 'ws' || data?.messageStreamTransport === 'sse' || data?.messageStreamTransport === 'auto'
                             ? data.messageStreamTransport
                             : undefined;
-                    const sttProvider = data?.sttProvider === 'browser' || data?.sttProvider === 'server' || data?.sttProvider === 'wasm' ? data.sttProvider : undefined;
+                    const sttProvider = normalizeSttProvider(data?.sttProvider);
                     const sttServerUrl = typeof data?.sttServerUrl === 'string' ? data.sttServerUrl.trim() : undefined;
                     const sttModel = typeof data?.sttModel === 'string' ? data.sttModel.trim() : undefined;
+                    const sttLocalModel = typeof data?.sttLocalModel === 'string' ? data.sttLocalModel.trim() : undefined;
                     const sttLanguage = typeof data?.sttLanguage === 'string' ? data.sttLanguage.trim() : undefined;
-                    const sttSilenceThresholdDb = normalizeSttSilenceThresholdDb(data?.sttSilenceThresholdDb);
-                    const sttSilenceHoldMs = normalizeSttSilenceHoldMs(data?.sttSilenceHoldMs);
 
                     return finish('runtime-settings', {
                         defaultModel: defaultModel.length > 0 ? defaultModel : undefined,
@@ -120,9 +114,8 @@ const fetchOpenChamberDefaults = async (): Promise<OpenChamberDefaults> => {
                         sttProvider,
                         sttServerUrl,
                         sttModel,
+                        sttLocalModel,
                         sttLanguage,
-                        sttSilenceThresholdDb,
-                        sttSilenceHoldMs,
                     });
                 }
             } catch {
@@ -149,12 +142,11 @@ const fetchOpenChamberDefaults = async (): Promise<OpenChamberDefaults> => {
             data?.messageStreamTransport === 'ws' || data?.messageStreamTransport === 'sse' || data?.messageStreamTransport === 'auto'
                 ? data.messageStreamTransport
                 : undefined;
-        const sttProvider = data?.sttProvider === 'browser' || data?.sttProvider === 'server' ? data.sttProvider : undefined;
+        const sttProvider = normalizeSttProvider(data?.sttProvider);
         const sttServerUrl = typeof data?.sttServerUrl === 'string' ? data.sttServerUrl.trim() : undefined;
         const sttModel = typeof data?.sttModel === 'string' ? data.sttModel.trim() : undefined;
+        const sttLocalModel = typeof data?.sttLocalModel === 'string' ? data.sttLocalModel.trim() : undefined;
         const sttLanguage = typeof data?.sttLanguage === 'string' ? data.sttLanguage.trim() : undefined;
-        const sttSilenceThresholdDb = normalizeSttSilenceThresholdDb(data?.sttSilenceThresholdDb);
-        const sttSilenceHoldMs = normalizeSttSilenceHoldMs(data?.sttSilenceHoldMs);
 
         return finish('settings-route', {
             defaultModel: defaultModel.length > 0 ? defaultModel : undefined,
@@ -168,9 +160,8 @@ const fetchOpenChamberDefaults = async (): Promise<OpenChamberDefaults> => {
             sttProvider,
             sttServerUrl,
             sttModel,
+            sttLocalModel,
             sttLanguage,
-            sttSilenceThresholdDb,
-            sttSilenceHoldMs,
         });
     } catch (error) {
         markStartupTrace('config.defaults:error', { error: error instanceof Error ? error.message : String(error) });
@@ -989,33 +980,31 @@ interface ConfigStore {
     settingsZenModel: string | undefined;
     settingsMessageStreamTransport: 'auto' | 'ws' | 'sse';
     // Voice provider preference ('browser', 'openai', 'openai-compatible', or 'say' for macOS)
-    voiceProvider: 'browser' | 'openai' | 'openai-compatible' | 'say';
+    voiceProvider: 'browser' | 'local' | 'openai' | 'openai-compatible' | 'say';
-    setVoiceProvider: (provider: 'browser' | 'openai' | 'openai-compatible' | 'say') => void;
+    setVoiceProvider: (provider: 'browser' | 'local' | 'openai' | 'openai-compatible' | 'say') => void;
     // TTS settings
     speechRate: number;
     speechPitch: number;
     speechVolume: number;
     sayVoice: string;
     browserVoice: string;
+    localTtsVoiceId: number;
     openaiVoice: string;
     openaiApiKey: string;
     openaiCompatibleUrl: string;
     openaiCompatibleApiKey: string;
     openaiCompatibleVoice: string;
     openaiCompatibleTtsModel: string;
-    // STT (speech-to-text) settings
+    // STT (dictation) settings
-    sttProvider: 'browser' | 'server' | 'wasm';
+    dictationEnabled: boolean;
+    sttProvider: 'local' | 'openai-compatible';
     sttServerUrl: string;
     sttApiKey: string;
     sttModel: string;
-    wasmSttModel: string;
+    sttLocalModel: string;
     sttLanguage: string;
-    sttSilenceThresholdDb: number;
-    sttSilenceHoldMs: number;
-    sttTranscribeOnStop: boolean;
     showMessageTTSButtons: boolean;
     ttsInputMode: 'sanitized' | 'raw';
-    voiceModeEnabled: boolean;
     // Summarization settings
     summarizeMessageTTS: boolean;
     summarizeVoiceConversation: boolean;
@@ -1026,24 +1015,22 @@ interface ConfigStore {
     setSpeechVolume: (volume: number) => void;
     setSayVoice: (voice: string) => void;
     setBrowserVoice: (voice: string) => void;
+    setLocalTtsVoiceId: (voiceId: number) => void;
     setOpenaiVoice: (voice: string) => void;
     setOpenaiApiKey: (apiKey: string) => void;
     setOpenaiCompatibleUrl: (url: string) => void;
     setOpenaiCompatibleApiKey: (apiKey: string) => void;
     setOpenaiCompatibleVoice: (voice: string) => void;
     setOpenaiCompatibleTtsModel: (model: string) => void;
-    setSttProvider: (provider: 'browser' | 'server' | 'wasm') => void;
+    setDictationEnabled: (enabled: boolean) => void;
+    setSttProvider: (provider: 'local' | 'openai-compatible') => void;
     setSttServerUrl: (url: string) => void;
     setSttApiKey: (apiKey: string) => void;
     setSttModel: (model: string) => void;
-    setWasmSttModel: (model: string) => void;
+    setSttLocalModel: (model: string) => void;
     setSttLanguage: (lang: string) => void;
-    setSttSilenceThresholdDb: (db: number) => void;
-    setSttSilenceHoldMs: (ms: number) => void;
-    setSttTranscribeOnStop: (enabled: boolean) => void;
     setShowMessageTTSButtons: (show: boolean) => void;
     setTtsInputMode: (mode: 'sanitized' | 'raw') => void;
-    setVoiceModeEnabled: (enabled: boolean) => void;
     setSummarizeMessageTTS: (enabled: boolean) => void;
     setSummarizeVoiceConversation: (enabled: boolean) => void;
     setSummarizeCharacterThreshold: (threshold: number) => void;
@@ -1136,7 +1123,7 @@ export const useConfigStore = create<ConfigStore>()(
                 voiceProvider: (() => {
                     if (typeof window !== 'undefined') {
                         const saved = localStorage.getItem('voiceProvider');
-                        if (saved === 'openai' || saved === 'browser' || saved === 'say' || saved === 'openai-compatible') return saved;
+                        if (saved === 'openai' || saved === 'browser' || saved === 'local' || saved === 'say' || saved === 'openai-compatible') return saved;
                     }
                     return 'browser';
                 })(),
@@ -1179,6 +1166,17 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                     return 'Samantha';
                 })(),
+                // Local (Kokoro) TTS speaker id - load from localStorage or default to 0
+                localTtsVoiceId: (() => {
+                    if (typeof window !== 'undefined') {
+                        const saved = localStorage.getItem('localTtsVoiceId');
+                        if (saved !== null) {
+                            const parsed = Number.parseInt(saved, 10);
+                            if (Number.isInteger(parsed) && parsed >= 0) return parsed;
+                        }
+                    }
+                    return 0;
+                })(),
                 // Browser voice - load from localStorage or default to empty (auto-select)
                 browserVoice: (() => {
                     if (typeof window !== 'undefined') {
@@ -1235,17 +1233,24 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                     return 'kokoro';
                 })(),
-                // STT provider: 'browser' (Web Speech API), 'server' (OpenAI-compat), 'wasm' (local Whisper)
+                // Voice input (dictation) master toggle - default enabled
+                dictationEnabled: (() => {
+                    if (typeof window !== 'undefined') {
+                        const saved = localStorage.getItem('dictationEnabled');
+                        if (saved === 'false') return false;
+                    }
+                    return true;
+                })(),
+                // STT provider: 'local' (server-side sherpa-onnx) or 'openai-compatible'
                 sttProvider: (() => {
                     if (typeof window !== 'undefined') {
                         const saved = localStorage.getItem('sttProvider');
-                        if (saved === 'browser' || saved === 'server' || saved === 'wasm') return saved;
+                        if (saved === 'local' || saved === 'openai-compatible') return saved;
-                        // Electron/Chromium's Web Speech API requires Google API keys
+                        // Migrate legacy providers: 'server' used an OpenAI-compatible
-                        // not available in Electron, so default to WASM local Whisper.
+                        // endpoint; 'browser' and 'wasm' map to the local default.
-                        const electron = (window as unknown as { __OPENCHAMBER_ELECTRON__?: { runtime?: string } }).__OPENCHAMBER_ELECTRON__;
+                        if (saved === 'server') return 'openai-compatible' as const;
-                        if (electron?.runtime === 'electron') return 'wasm' as const;
                     }
-                    return 'browser' as const;
+                    return 'local' as const;
                 })(),
                 sttServerUrl: (() => {
                     if (typeof window !== 'undefined') {
@@ -1268,12 +1273,12 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                     return 'deepdml/faster-whisper-large-v3-turbo-ct2';
                 })(),
-                wasmSttModel: (() => {
+                sttLocalModel: (() => {
                     if (typeof window !== 'undefined') {
-                        const saved = localStorage.getItem('wasmSttModel');
+                        const saved = localStorage.getItem('sttLocalModel');
                         if (saved) return saved;
                     }
-                    return 'Xenova/whisper-base.en';
+                    return 'parakeet-tdt-0.6b-v2-int8';
                 })(),
                 sttLanguage: (() => {
                     if (typeof window !== 'undefined') {
@@ -1282,33 +1287,6 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                     return '';
                 })(),
-                sttSilenceThresholdDb: (() => {
-                    if (typeof window !== 'undefined') {
-                        const saved = localStorage.getItem('sttSilenceThresholdDb');
-                        if (saved) {
-                            const parsed = parseFloat(saved);
-                            if (!isNaN(parsed)) return parsed;
-                        }
-                    }
-                    return -45;
-                })(),
-                sttSilenceHoldMs: (() => {
-                    if (typeof window !== 'undefined') {
-                        const saved = localStorage.getItem('sttSilenceHoldMs');
-                        if (saved) {
-                            const parsed = parseInt(saved, 10);
-                            if (!isNaN(parsed)) return parsed;
-                        }
-                    }
-                    return 1500;
-                })(),
-                sttTranscribeOnStop: (() => {
-                    if (typeof window !== 'undefined') {
-                        const saved = localStorage.getItem('sttTranscribeOnStop');
-                        if (saved === 'true') return true;
-                    }
-                    return false;
-                })(),
                 // Show TTS buttons on messages - disabled by default until user enables it
                 showMessageTTSButtons: (() => {
                     if (typeof window !== 'undefined') {
@@ -1324,14 +1302,6 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                     return 'sanitized' as const;
                 })(),
-                // Voice mode enabled - load from localStorage or default to false
-                voiceModeEnabled: (() => {
-                    if (typeof window !== 'undefined') {
-                        const saved = localStorage.getItem('voiceModeEnabled');
-                        if (saved === 'true') return true;
-                    }
-                    return false;
-                })(),
                 // Summarization settings
                 summarizeMessageTTS: (() => {
                     if (typeof window !== 'undefined') {
@@ -2060,9 +2030,8 @@ export const useConfigStore = create<ConfigStore>()(
                                     sttProvider: openChamberDefaults.sttProvider ?? state.sttProvider,
                                     sttServerUrl: openChamberDefaults.sttServerUrl ?? state.sttServerUrl,
                                     sttModel: openChamberDefaults.sttModel ?? state.sttModel,
+                                    sttLocalModel: openChamberDefaults.sttLocalModel ?? state.sttLocalModel,
                                     sttLanguage: openChamberDefaults.sttLanguage ?? state.sttLanguage,
-                                    sttSilenceThresholdDb: openChamberDefaults.sttSilenceThresholdDb ?? state.sttSilenceThresholdDb,
-                                    sttSilenceHoldMs: openChamberDefaults.sttSilenceHoldMs ?? state.sttSilenceHoldMs,
                                     directoryScoped: {
                                         ...state.directoryScoped,
                                         [directoryKey]: nextSnapshot,
@@ -2800,7 +2769,7 @@ export const useConfigStore = create<ConfigStore>()(
                     });
                 },
 
-                setVoiceProvider: (provider: 'browser' | 'openai' | 'openai-compatible' | 'say') => {
+                setVoiceProvider: (provider: 'browser' | 'local' | 'openai' | 'openai-compatible' | 'say') => {
                     set({ voiceProvider: provider });
                     if (typeof window !== 'undefined') {
                         localStorage.setItem('voiceProvider', provider);
@@ -2838,6 +2807,13 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                 },
 
+                setLocalTtsVoiceId: (voiceId: number) => {
+                    set({ localTtsVoiceId: voiceId });
+                    if (typeof window !== 'undefined') {
+                        localStorage.setItem('localTtsVoiceId', String(voiceId));
+                    }
+                },
+
                 setBrowserVoice: (voice: string) => {
                     set({ browserVoice: voice });
                     if (typeof window !== 'undefined') {
@@ -2887,7 +2863,15 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                 },
 
-                setSttProvider: (provider: 'browser' | 'server' | 'wasm') => {
+                setDictationEnabled: (enabled: boolean) => {
+                    set({ dictationEnabled: enabled });
+                    if (typeof window !== 'undefined') {
+                        localStorage.setItem('dictationEnabled', String(enabled));
+                    }
+                    updateDesktopSettings({ dictationEnabled: enabled }).catch(() => {});
+                },
+
+                setSttProvider: (provider: 'local' | 'openai-compatible') => {
                     set({ sttProvider: provider });
                     if (typeof window !== 'undefined') {
                         localStorage.setItem('sttProvider', provider);
@@ -2918,12 +2902,12 @@ export const useConfigStore = create<ConfigStore>()(
                     updateDesktopSettings({ sttModel: model }).catch(() => {});
                 },
 
-                setWasmSttModel: (model: string) => {
+                setSttLocalModel: (model: string) => {
-                    set({ wasmSttModel: model });
+                    set({ sttLocalModel: model });
                     if (typeof window !== 'undefined') {
-                        localStorage.setItem('wasmSttModel', model);
+                        localStorage.setItem('sttLocalModel', model);
                     }
-                    updateDesktopSettings({ wasmSttModel: model }).catch(() => {});
+                    updateDesktopSettings({ sttLocalModel: model }).catch(() => {});
                 },
 
                 setSttLanguage: (lang: string) => {
@@ -2934,30 +2918,6 @@ export const useConfigStore = create<ConfigStore>()(
                     updateDesktopSettings({ sttLanguage: lang }).catch(() => {});
                 },
 
-                setSttSilenceThresholdDb: (db: number) => {
-                    set({ sttSilenceThresholdDb: db });
-                    if (typeof window !== 'undefined') {
-                        localStorage.setItem('sttSilenceThresholdDb', String(db));
-                    }
-                    updateDesktopSettings({ sttSilenceThresholdDb: db }).catch(() => {});
-                },
-
-                setSttSilenceHoldMs: (ms: number) => {
-                    set({ sttSilenceHoldMs: ms });
-                    if (typeof window !== 'undefined') {
-                        localStorage.setItem('sttSilenceHoldMs', String(ms));
-                    }
-                    updateDesktopSettings({ sttSilenceHoldMs: ms }).catch(() => {});
-                },
-
-                setSttTranscribeOnStop: (enabled: boolean) => {
-                    set({ sttTranscribeOnStop: enabled });
-                    if (typeof window !== 'undefined') {
-                        localStorage.setItem('sttTranscribeOnStop', String(enabled));
-                    }
-                    updateDesktopSettings({ sttTranscribeOnStop: enabled }).catch(() => {});
-                },
-
                 setShowMessageTTSButtons: (show: boolean) => {
                     set({ showMessageTTSButtons: show });
                     if (typeof window !== 'undefined') {
@@ -2972,13 +2932,6 @@ export const useConfigStore = create<ConfigStore>()(
                     }
                 },
 
-                setVoiceModeEnabled: (enabled: boolean) => {
-                    set({ voiceModeEnabled: enabled });
-                    if (typeof window !== 'undefined') {
-                        localStorage.setItem('voiceModeEnabled', String(enabled));
-                    }
-                },
-
                 setSummarizeMessageTTS: (enabled: boolean) => {
                     set({ summarizeMessageTTS: enabled });
                     if (typeof window !== 'undefined') {

packages/ui/src/sync/session-ui-store.ts

@@ -190,7 +190,6 @@ function notifyMessageSent(sessionId: string): void {
 
 export type { SyntheticContextPart } from "./input-store"
 export type { SessionMemoryState } from "./viewport-store"
-export type { VoiceStatus, VoiceMode } from "./voice-store"
 
 export type NewSessionDraftState = {
   open: boolean

packages/ui/src/sync/sync-context.tsx

@@ -8,7 +8,6 @@ import type { OpencodeClient } from "@opencode-ai/sdk/v2/client"
 import { createEventPipeline } from "./event-pipeline"
 import { isVSCodeRuntime } from "@/lib/desktop"
 import { isMobileSurfaceRuntime } from "@/lib/runtimeSurface"
-import { isCapacitorApp } from "@/lib/platform"
 import { reduceGlobalEvent, applyGlobalProject, applyDirectoryEvent, type SessionMaterializationReason } from "./event-reducer"
 import { useGlobalSyncStore } from "./global-sync-store"
 import { ChildStoreManager, type DirectoryStore } from "./child-store"
@@ -1692,12 +1691,13 @@ export function SyncProvider(props: {
   directory: string
   children: React.ReactNode
 }) {
-  const storedMessageStreamTransport = useConfigStore((state) => state.settingsMessageStreamTransport)
+  // Capacitor apps were previously locked to SSE because Android WebSocket
-  // Capacitor apps are locked to SSE: native WebSocket streaming is unreliable there (on
+  // upgrades appeared broken. Root cause was server-side: the Android WebView
-  // Android events only arrive once the run finishes), while SSE streams correctly. The Chat
+  // origin (https://localhost, androidScheme 'https') was missing from the
-  // settings UI disables the other options on mobile, but force it here too so the effective
+  // packaged-client origin allowlist, so every WS upgrade was rejected with
-  // transport can't drift. Remove this override (and the UI lock) to re-enable WS on mobile.
+  // 403. With the origin allowlisted, mobile uses the same transport
-  const messageStreamTransport: 'auto' | 'ws' | 'sse' = isCapacitorApp() ? 'sse' : storedMessageStreamTransport
+  // selection as everywhere else ('auto' falls back to SSE on WS failure).
+  const messageStreamTransport = useConfigStore((state) => state.settingsMessageStreamTransport)
   const childStoresRef = useRef<ChildStoreManager | null>(null)
   if (!childStoresRef.current) childStoresRef.current = new ChildStoreManager()
   const childStores = childStoresRef.current

packages/ui/src/sync/voice-store.ts

@@ -1,7 +0,0 @@
-/**
- * Voice Store — voice connection and activity state.
- * Extracted from session-ui-store for subscription isolation.
- */
-
-export type VoiceStatus = "disconnected" | "connecting" | "connected" | "error"
-export type VoiceMode = "idle" | "speaking" | "listening"

packages/vscode/webview/main.tsx

@@ -520,6 +520,23 @@ const handleLocalApiRequest = async (input: RequestInfo | URL, url: URL, init: R
     });
   }
 
+  // Dictation runs on the OpenChamber web server (WebSocket + worker); the VS
+  // Code bridge has no server process, so report it deterministically
+  // unavailable. The mic button hides itself when capture is unsupported.
+  if (normalizedPathname === '/api/dictation/status' && method === 'GET') {
+    return new Response(JSON.stringify({ provider: 'local', available: false, reasonCode: 'unsupported_runtime', models: [] }), {
+      status: 200,
+      headers: { 'Content-Type': 'application/json' },
+    });
+  }
+
+  if (normalizedPathname.startsWith('/api/dictation/') ) {
+    return new Response(JSON.stringify({ error: 'Dictation is not available in VS Code runtime' }), {
+      status: 501,
+      headers: { 'Content-Type': 'application/json' },
+    });
+  }
+
   // Health endpoints: reflect actual connection status
   if (pathname === '/health' || pathname === '/api/health') {
     const connectionStatus = window.__OPENCHAMBER_CONNECTION__?.status;

packages/web/.gitignore

@@ -1,2 +1,3 @@
 server/prompt-templates.js
 prompt-enhancer-config.json
+test-results/

packages/web/package.json

@@ -41,6 +41,7 @@
     "openai": "^4.79.0",
     "qrcode-terminal": "^0.12.0",
     "reflect-metadata": "^0.2.2",
+    "sherpa-onnx-node": "1.12.28",
     "simple-git": "^3.28.0",
     "web-push": "^3.6.7",
     "ws": "^8.18.3",

packages/web/server/index.js

@@ -38,6 +38,7 @@ import { prepareNotificationLastMessage } from './lib/notifications/index.js';
 import { registerTtsRoutes } from './lib/tts/routes.js';
 import { detectSayTtsCapability } from './lib/tts/capability-runtime.js';
 import { createTerminalRuntime } from './lib/terminal/runtime.js';
+import { createDictationRuntime } from './lib/dictation/runtime.js';
 import {
   createGlobalUiEventBroadcaster,
   createGlobalMessageStreamHub,
@@ -495,6 +496,7 @@ const tunnelAuthController = createTunnelAuth();
 let runtimeManagedRemoteTunnelToken = '';
 let runtimeManagedRemoteTunnelHostname = '';
 let terminalRuntime = null;
+let dictationRuntime = null;
 let messageStreamRuntime = null;
 const userProvidedOpenCodePassword = hmrStateRuntime.getUserProvidedOpenCodePassword(hmrState);
 const initialOpenCodeAuthState = hmrStateRuntime.resolveOpenCodeAuthFromState({
@@ -898,6 +900,7 @@ const tunnelWiringRuntime = createTunnelWiringRuntime({
 });
 const startupPipelineRuntime = createStartupPipelineRuntime({
   createTerminalRuntime,
+  createDictationRuntime,
   createMessageStreamWsRuntime,
   createServerStartupRuntime,
 });
@@ -1357,8 +1360,10 @@ async function main(options = {}) {
     tunnelRuntimeContext,
     attachSignals,
     apiOnly,
+    dictationModelsDir: path.join(OPENCHAMBER_USER_CONFIG_ROOT, 'speech-models'),
   });
   terminalRuntime = startupPipelineResult.terminalRuntime;
+  dictationRuntime = startupPipelineResult.dictationRuntime;
   messageStreamRuntime = startupPipelineResult.messageStreamRuntime;
 
   try {
@@ -1397,6 +1402,11 @@ async function main(options = {}) {
     },
     stop: (shutdownOptions = {}) => {
       realtimeProxyRuntime.stop();
+      try {
+        dictationRuntime?.stop?.();
+      } catch {
+        // best-effort shutdown of the dictation worker
+      }
       return gracefulShutdown({ exitProcess: shutdownOptions.exitProcess ?? false });
     }
   };

packages/web/server/lib/dictation/DOCUMENTATION.md

@@ -0,0 +1,63 @@
+# Dictation module
+
+Server-authoritative streaming speech-to-text for the chat composer, plus
+local text-to-speech. The client streams 16 kHz mono PCM16 chunks (base64)
+over a WebSocket; the server runs the transcription and streams live partial
+transcripts back.
+
+Local TTS (Kokoro via sherpa-onnx OfflineTts) runs in the same worker process
+and is exposed as `POST /api/dictation/tts/speak` (JSON `{text, speakerId?,
+speed?, model?}` → WAV bytes; 503 with `reasonCode` while the model is
+downloading). TTS models live in the same catalog/downloader as STT models
+(`local/model-catalog.js` `LOCAL_TTS_MODEL_CATALOG`) and are managed by the
+same status/download/delete routes.
+
+## Ownership
+
+- `runtime.js` — registers `GET /api/dictation/status`,
+  `POST /api/dictation/models/:modelId/download`, and the
+  `/api/dictation/ws` WebSocket endpoint (auth-gated the same way as the
+  terminal WS: UI session token or `oc_url_token`, plus origin check).
+  Created from the startup pipeline (`startup-pipeline-runtime.js`) before
+  the generic OpenCode proxy so routes are not shadowed.
+- `stream-manager.js` — `DictationStreamManager`, one per WS connection.
+  Chunk reordering by `seq` + ack, resampling to the provider rate,
+  auto-commit every ~15 s of audio, silence suppression by PCM peak,
+  partial-transcript concatenation, adaptive finalization timeout.
+- `service.js` — provider resolution and readiness. Providers:
+  - `local` (default): sherpa-onnx Parakeet TDT in a forked worker process.
+    Models auto-download in the background on first use; while missing, the
+    stream fails with `reasonCode: 'model_download_in_progress'` and the
+    status route reports per-model install/download state.
+  - `openai-compatible`: buffered per-segment transcription against any
+    OpenAI-compatible `/v1/audio/transcriptions` endpoint
+    (`openai-compatible-session.js`, reuses `../tts/stt.js`).
+- `local/` — worker process + client (IPC, idle shutdown TTL), sherpa
+  recognizer engine and realtime session (throttled re-decode for partials),
+  model catalog and downloader. The native `sherpa-onnx-node` addon is only
+  ever loaded inside the worker process.
+- `audio.js` — PCM16 helpers: format parsing, peak, WAV wrapping, streaming
+  linear resampler.
+
+## WebSocket protocol (JSON text frames)
+
+Client → server: `start {dictationId, format, options}`,
+`chunk {dictationId, seq, audio}`, `finish {dictationId, finalSeq}`,
+`cancel {dictationId}`, `ping`.
+
+Server → client: `ready`, `ack {ackSeq}`, `partial {text}`,
+`finish_accepted {timeoutMs}`, `final {text}`,
+`error {error, retryable, reasonCode?}`, `pong`.
+
+`options` in `start` carries the client-selected provider config:
+`{ provider: 'local' | 'openai-compatible', language?, localModel?,
+openaiCompatible?: { baseUrl, model, apiKey } }`.
+
+## Invariants
+
+- Never load `sherpa-onnx-node` in the main server process.
+- The stream manager acks only the highest contiguous seq; the client is
+  expected to retain unacked segments for retry/replay.
+- Silence-only segments (peak < 300) are cleared, never committed, so
+  Whisper-style providers do not hallucinate on silence.
+- Model files live under `~/.config/openchamber/speech-models`.

packages/web/server/lib/dictation/audio.js

@@ -0,0 +1,195 @@
+/**
+ * PCM16 audio helpers for the dictation streaming pipeline.
+ *
+ * All dictation audio travels as 16-bit little-endian mono PCM. The client
+ * captures at 16 kHz; providers may require a different rate, so chunks are
+ * resampled with Pcm16MonoResampler before being appended to an STT session.
+ */
+
+/**
+ * Parse the sample rate out of a format string like "audio/pcm;rate=16000;bits=16".
+ * @param {string} format
+ * @param {number|null} [fallback]
+ * @returns {number|null}
+ */
+export function parsePcmRateFromFormat(format, fallback = null) {
+  const match = /(?:^|[;,\s])rate\s*=\s*(\d+)(?:$|[;,\s])/i.exec(String(format || ''));
+  if (!match) {
+    return fallback;
+  }
+  const rate = Number.parseInt(match[1], 10);
+  return Number.isFinite(rate) && rate > 0 ? rate : fallback;
+}
+
+/**
+ * Return an Int16Array view over a PCM16LE buffer, copying when the buffer's
+ * byteOffset is not 2-byte aligned (IPC-transferred buffers can be views at
+ * odd offsets, and Int16Array requires an even start offset).
+ * @param {Buffer} pcm16le
+ * @returns {Int16Array}
+ */
+function toInt16Samples(pcm16le) {
+  if (pcm16le.byteOffset % 2 !== 0) {
+    const copy = Buffer.from(pcm16le);
+    return new Int16Array(copy.buffer, copy.byteOffset, copy.byteLength / 2);
+  }
+  return new Int16Array(pcm16le.buffer, pcm16le.byteOffset, pcm16le.byteLength / 2);
+}
+
+/**
+ * Peak absolute sample value of a PCM16LE buffer. Used for silence detection.
+ * @param {Buffer} pcm16le
+ * @returns {number}
+ */
+export function pcm16lePeakAbs(pcm16le) {
+  if (!pcm16le || pcm16le.length === 0) {
+    return 0;
+  }
+  if (pcm16le.length % 2 !== 0) {
+    throw new Error(`PCM16 chunk byteLength must be even, got ${pcm16le.length}`);
+  }
+  const samples = toInt16Samples(pcm16le);
+  let peak = 0;
+  for (let i = 0; i < samples.length; i += 1) {
+    const v = samples[i];
+    const abs = v < 0 ? -v : v;
+    if (abs > peak) {
+      peak = abs;
+      if (peak >= 32767) {
+        break;
+      }
+    }
+  }
+  return peak;
+}
+
+/**
+ * Convert PCM16LE to Float32 samples in [-1, 1], with optional gain.
+ * @param {Buffer} pcm16le
+ * @param {number} [gain]
+ * @returns {Float32Array}
+ */
+export function pcm16leToFloat32(pcm16le, gain = 1) {
+  if (pcm16le.length % 2 !== 0) {
+    throw new Error(`PCM16 chunk byteLength must be even, got ${pcm16le.length}`);
+  }
+  const int16 = toInt16Samples(pcm16le);
+  const out = new Float32Array(int16.length);
+  for (let i = 0; i < int16.length; i += 1) {
+    const v = (int16[i] / 32768.0) * gain;
+    out[i] = Math.max(-1, Math.min(1, v));
+  }
+  return out;
+}
+
+/**
+ * Wrap raw PCM16LE mono audio in a WAV container.
+ * @param {Buffer} pcmBuffer
+ * @param {number} sampleRate
+ * @returns {Buffer}
+ */
+export function pcm16ToWav(pcmBuffer, sampleRate) {
+  const channels = 1;
+  const bitsPerSample = 16;
+  const headerSize = 44;
+  const wavBuffer = Buffer.alloc(headerSize + pcmBuffer.length);
+  const byteRate = (sampleRate * channels * bitsPerSample) / 8;
+  const blockAlign = (channels * bitsPerSample) / 8;
+
+  wavBuffer.write('RIFF', 0);
+  wavBuffer.writeUInt32LE(36 + pcmBuffer.length, 4);
+  wavBuffer.write('WAVE', 8);
+  wavBuffer.write('fmt ', 12);
+  wavBuffer.writeUInt32LE(16, 16);
+  wavBuffer.writeUInt16LE(1, 20);
+  wavBuffer.writeUInt16LE(channels, 22);
+  wavBuffer.writeUInt32LE(sampleRate, 24);
+  wavBuffer.writeUInt32LE(byteRate, 28);
+  wavBuffer.writeUInt16LE(blockAlign, 32);
+  wavBuffer.writeUInt16LE(bitsPerSample, 34);
+  wavBuffer.write('data', 36);
+  wavBuffer.writeUInt32LE(pcmBuffer.length, 40);
+  pcmBuffer.copy(wavBuffer, 44);
+
+  return wavBuffer;
+}
+
+/**
+ * Streaming linear-interpolation resampler for PCM16LE mono audio.
+ * Carries one sample across chunk boundaries so consecutive chunks resample
+ * without seams.
+ */
+export class Pcm16MonoResampler {
+  /**
+   * @param {{ inputRate: number, outputRate: number }} params
+   */
+  constructor({ inputRate, outputRate }) {
+    this.inputRate = inputRate;
+    this.outputRate = outputRate;
+    this.step = inputRate / outputRate;
+    this.pos = 0;
+    this.carrySample = null;
+  }
+
+  reset() {
+    this.pos = 0;
+    this.carrySample = null;
+  }
+
+  /**
+   * @param {Buffer} pcm16le
+   * @returns {Buffer}
+   */
+  processChunk(pcm16le) {
+    if (pcm16le.length === 0) {
+      return Buffer.alloc(0);
+    }
+    if (pcm16le.length % 2 !== 0) {
+      throw new Error(`PCM16 chunk byteLength must be even, got ${pcm16le.length}`);
+    }
+
+    const srcChunk = toInt16Samples(pcm16le);
+
+    const hasCarry = this.carrySample !== null;
+    const srcLen = srcChunk.length + (hasCarry ? 1 : 0);
+    if (srcLen < 2) {
+      this.carrySample = srcChunk.length ? srcChunk[srcChunk.length - 1] : this.carrySample;
+      return Buffer.alloc(0);
+    }
+
+    const src = new Float32Array(srcLen);
+    let offset = 0;
+    if (hasCarry) {
+      src[0] = this.carrySample / 32768;
+      offset = 1;
+    }
+    for (let i = 0; i < srcChunk.length; i += 1) {
+      src[offset + i] = srcChunk[i] / 32768;
+    }
+
+    const out = [];
+    const maxPos = src.length - 1;
+
+    while (this.pos < maxPos) {
+      const i = Math.floor(this.pos);
+      const frac = this.pos - i;
+      const s0 = src[i];
+      const s1 = src[i + 1];
+      const sample = s0 + (s1 - s0) * frac;
+      const clamped = Math.max(-1, Math.min(1, sample));
+      out.push(Math.round(clamped * 32767));
+      this.pos += this.step;
+    }
+
+    this.carrySample = srcChunk[srcChunk.length - 1];
+
+    const shift = src.length - 1;
+    this.pos = this.pos - shift;
+    if (this.pos < 0) {
+      this.pos = 0;
+    }
+
+    const outArr = Int16Array.from(out);
+    return Buffer.from(outArr.buffer, outArr.byteOffset, outArr.byteLength);
+  }
+}

packages/web/server/lib/dictation/local/model-catalog.js

@@ -0,0 +1,141 @@
+/**
+ * Catalog of local sherpa-onnx STT models available for dictation.
+ * Models are downloaded on demand from the k2-fsa GitHub releases and
+ * extracted under the OpenChamber speech-models directory.
+ *
+ * `type` selects the recognizer construction path in the worker:
+ * - 'nemo_transducer': encoder/decoder/joiner transducer (Parakeet)
+ * - 'whisper': encoder/decoder Whisper export
+ * `files` maps logical roles to file names inside the extracted directory.
+ */
+
+import path from 'path';
+
+export const LOCAL_STT_MODEL_CATALOG = {
+  'parakeet-tdt-0.6b-v2-int8': {
+    type: 'nemo_transducer',
+    archiveUrl:
+      'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2',
+    extractedDir: 'sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8',
+    files: {
+      encoder: 'encoder.int8.onnx',
+      decoder: 'decoder.int8.onnx',
+      joiner: 'joiner.int8.onnx',
+      tokens: 'tokens.txt',
+    },
+    description: 'NVIDIA Parakeet TDT v2 (English)',
+  },
+  'parakeet-tdt-0.6b-v3-int8': {
+    type: 'nemo_transducer',
+    archiveUrl:
+      'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2',
+    extractedDir: 'sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8',
+    files: {
+      encoder: 'encoder.int8.onnx',
+      decoder: 'decoder.int8.onnx',
+      joiner: 'joiner.int8.onnx',
+      tokens: 'tokens.txt',
+    },
+    description: 'NVIDIA Parakeet TDT v3 (25 European languages, auto-detected)',
+  },
+  'whisper-base-int8': {
+    type: 'whisper',
+    archiveUrl:
+      'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-base.tar.bz2',
+    extractedDir: 'sherpa-onnx-whisper-base',
+    files: {
+      encoder: 'base-encoder.int8.onnx',
+      decoder: 'base-decoder.int8.onnx',
+      tokens: 'base-tokens.txt',
+    },
+    description: 'OpenAI Whisper base (multilingual, smaller and lighter)',
+  },
+  'whisper-tiny-int8': {
+    type: 'whisper',
+    archiveUrl:
+      'https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-tiny.tar.bz2',
+    extractedDir: 'sherpa-onnx-whisper-tiny',
+    files: {
+      encoder: 'tiny-encoder.int8.onnx',
+      decoder: 'tiny-decoder.int8.onnx',
+      tokens: 'tiny-tokens.txt',
+    },
+    description: 'OpenAI Whisper tiny (multilingual, fastest and lightest)',
+  },
+};
+
+/**
+ * Local text-to-speech models (sherpa-onnx OfflineTts). Downloaded and
+ * managed through the same pipeline as the STT models.
+ */
+export const LOCAL_TTS_MODEL_CATALOG = {
+  'kokoro-en-v0_19': {
+    type: 'kokoro',
+    archiveUrl:
+      'https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/kokoro-en-v0_19.tar.bz2',
+    extractedDir: 'kokoro-en-v0_19',
+    files: {
+      model: 'model.onnx',
+      voices: 'voices.bin',
+      tokens: 'tokens.txt',
+      espeakData: 'espeak-ng-data',
+    },
+    description: 'Kokoro TTS (English, natural voices)',
+  },
+};
+
+export const DEFAULT_LOCAL_STT_MODEL = 'parakeet-tdt-0.6b-v2-int8';
+export const DEFAULT_LOCAL_TTS_MODEL = 'kokoro-en-v0_19';
+
+export const LOCAL_STT_MODEL_IDS = Object.keys(LOCAL_STT_MODEL_CATALOG);
+export const LOCAL_TTS_MODEL_IDS = Object.keys(LOCAL_TTS_MODEL_CATALOG);
+
+/**
+ * @param {string} modelId
+ * @returns {boolean}
+ */
+export function isLocalSttModelId(modelId) {
+  return typeof modelId === 'string' && Object.hasOwn(LOCAL_STT_MODEL_CATALOG, modelId);
+}
+
+/**
+ * @param {string} modelId
+ * @returns {boolean}
+ */
+export function isLocalTtsModelId(modelId) {
+  return typeof modelId === 'string' && Object.hasOwn(LOCAL_TTS_MODEL_CATALOG, modelId);
+}
+
+/**
+ * Any managed local model (STT or TTS).
+ * @param {string} modelId
+ * @returns {boolean}
+ */
+export function isLocalModelId(modelId) {
+  return isLocalSttModelId(modelId) || isLocalTtsModelId(modelId);
+}
+
+/**
+ * Spec lookup across both catalogs (STT and TTS).
+ * @param {string} modelId
+ */
+export function getLocalSttModelSpec(modelId) {
+  const spec = LOCAL_STT_MODEL_CATALOG[modelId] ?? LOCAL_TTS_MODEL_CATALOG[modelId];
+  if (!spec) {
+    throw new Error(`Unknown local speech model id: ${modelId}`);
+  }
+  return {
+    id: modelId,
+    ...spec,
+    requiredFiles: Object.values(spec.files),
+  };
+}
+
+/**
+ * @param {string} modelsDir
+ * @param {string} modelId
+ * @returns {string}
+ */
+export function getLocalSttModelDir(modelsDir, modelId) {
+  return path.join(modelsDir, getLocalSttModelSpec(modelId).extractedDir);
+}

packages/web/server/lib/dictation/local/model-downloader.js

@@ -0,0 +1,163 @@
+/**
+ * Downloads and extracts local sherpa-onnx STT model archives.
+ * Archives (.tar.bz2) come from the k2-fsa GitHub releases and are extracted
+ * with the system `tar` into the speech-models directory.
+ */
+
+import { createWriteStream } from 'fs';
+import { mkdir, rename, rm, stat } from 'fs/promises';
+import path from 'path';
+import { Readable } from 'stream';
+import { pipeline } from 'stream/promises';
+import { spawn } from 'child_process';
+
+import { getLocalSttModelSpec } from './model-catalog.js';
+
+async function hasRequiredFiles(modelDir, requiredFiles) {
+  const results = await Promise.all(
+    requiredFiles.map(async (rel) => {
+      try {
+        const s = await stat(path.join(modelDir, rel));
+        if (s.isDirectory()) {
+          return true;
+        }
+        return s.isFile() && s.size > 0;
+      } catch {
+        return false;
+      }
+    }),
+  );
+  return results.every(Boolean);
+}
+
+async function downloadToFile(url, outputPath, onProgress) {
+  const res = await fetch(url);
+  if (!res.ok) {
+    throw new Error(`Failed to download ${url}: ${res.status} ${res.statusText}`);
+  }
+  if (!res.body) {
+    throw new Error(`Failed to download ${url}: missing response body`);
+  }
+
+  const totalBytes = Number.parseInt(res.headers.get('content-length') || '', 10) || null;
+  let downloadedBytes = 0;
+
+  const tmpPath = `${outputPath}.tmp-${Date.now()}`;
+  await mkdir(path.dirname(outputPath), { recursive: true });
+
+  const nodeStream = Readable.fromWeb(res.body);
+  if (typeof onProgress === 'function') {
+    nodeStream.on('data', (chunk) => {
+      downloadedBytes += chunk.length;
+      onProgress(downloadedBytes, totalBytes);
+    });
+  }
+
+  try {
+    await pipeline(nodeStream, createWriteStream(tmpPath));
+    await rename(tmpPath, outputPath);
+  } catch (error) {
+    await rm(tmpPath, { force: true }).catch(() => undefined);
+    throw error;
+  }
+}
+
+async function extractTarArchive(archivePath, destDir) {
+  await mkdir(destDir, { recursive: true });
+
+  await new Promise((resolve, reject) => {
+    const child = spawn('tar', ['xf', archivePath, '-C', destDir], {
+      stdio: 'ignore',
+      windowsHide: true,
+    });
+    child.on('error', reject);
+    child.on('exit', (code) => {
+      if (code === 0) {
+        resolve();
+      } else {
+        reject(new Error(`tar exited with code ${code}`));
+      }
+    });
+  });
+}
+
+async function isNonEmptyFile(filePath) {
+  try {
+    const s = await stat(filePath);
+    return s.isFile() && s.size > 0;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Check whether a model is fully installed (all required files present).
+ * @param {string} modelsDir
+ * @param {string} modelId
+ * @returns {Promise<boolean>}
+ */
+export async function isLocalSttModelInstalled(modelsDir, modelId) {
+  const spec = getLocalSttModelSpec(modelId);
+  return hasRequiredFiles(path.join(modelsDir, spec.extractedDir), spec.requiredFiles);
+}
+
+/**
+ * Ensure a model is downloaded and extracted. Resolves with the model dir.
+ *
+ * Extraction is staged: the archive unpacks into a temporary directory and is
+ * verified before being renamed into place. An interrupted or failed tar must
+ * never leave partial files at the final path — the installed check only
+ * verifies file presence, so a truncated .onnx there would be treated as an
+ * installed model forever ("Protobuf parsing failed" at load time).
+ *
+ * @param {{ modelsDir: string, modelId: string,
+ *           onProgress?: (downloadedBytes: number, totalBytes: number | null) => void }} options
+ * @returns {Promise<string>}
+ */
+export async function ensureLocalSttModel({ modelsDir, modelId, onProgress }) {
+  const spec = getLocalSttModelSpec(modelId);
+  const modelDir = path.join(modelsDir, spec.extractedDir);
+  if (await hasRequiredFiles(modelDir, spec.requiredFiles)) {
+    return modelDir;
+  }
+
+  // A directory that exists but fails the required-files check is a partial
+  // extraction from an earlier interrupted attempt — remove it before retrying.
+  await rm(modelDir, { recursive: true, force: true }).catch(() => undefined);
+
+  const downloadsDir = path.join(modelsDir, '.downloads');
+  const archiveFilename = path.basename(new URL(spec.archiveUrl).pathname);
+  const archivePath = path.join(downloadsDir, archiveFilename);
+
+  if (!(await isNonEmptyFile(archivePath))) {
+    await downloadToFile(spec.archiveUrl, archivePath, onProgress);
+  }
+
+  const stagingDir = path.join(modelsDir, `.staging-${spec.extractedDir}-${Date.now()}`);
+  try {
+    await extractTarArchive(archivePath, stagingDir);
+
+    const stagedModelDir = path.join(stagingDir, spec.extractedDir);
+    if (!(await hasRequiredFiles(stagedModelDir, spec.requiredFiles))) {
+      // Bad archive (truncated download / corrupt cache): drop it so the next
+      // attempt re-downloads instead of re-extracting the same broken bytes.
+      await rm(archivePath, { force: true }).catch(() => undefined);
+      throw new Error(
+        `Extracted ${archiveFilename}, but required model files are missing or empty. The archive was discarded; retry to re-download.`,
+      );
+    }
+
+    await rename(stagedModelDir, modelDir);
+  } catch (error) {
+    await rm(stagingDir, { recursive: true, force: true }).catch(() => undefined);
+    // Any extraction failure means the cached archive can't be trusted
+    // (corrupt bz2, truncated download). Discard it so retry re-downloads.
+    await rm(archivePath, { force: true }).catch(() => undefined);
+    throw error;
+  }
+  await rm(stagingDir, { recursive: true, force: true }).catch(() => undefined);
+
+  await rm(archivePath, { force: true }).catch(() => undefined);
+
+  return modelDir;
+}

packages/web/server/lib/dictation/local/sherpa-loader.js

@@ -0,0 +1,136 @@
+/**
+ * Loader for the sherpa-onnx-node native addon.
+ *
+ * sherpa-onnx-node ships its native addon and shared libraries in a
+ * platform-specific package (e.g. sherpa-onnx-darwin-arm64). The shared
+ * libraries must be findable via the platform's dynamic-loader search path,
+ * so the loader prepends the platform package directory to LD_LIBRARY_PATH /
+ * DYLD_LIBRARY_PATH / PATH before requiring the addon.
+ */
+
+import { createRequire } from 'module';
+import path from 'path';
+import { existsSync } from 'fs';
+
+const require = createRequire(import.meta.url);
+
+let cached = null;
+
+function sherpaPlatformPackageName(platform = process.platform, arch = process.arch) {
+  const normalizedPlatform = platform === 'win32' ? 'win' : platform;
+  return `sherpa-onnx-${normalizedPlatform}-${arch}`;
+}
+
+function sherpaLoaderEnvKey(platform = process.platform) {
+  if (platform === 'linux') {
+    return 'LD_LIBRARY_PATH';
+  }
+  if (platform === 'darwin') {
+    return 'DYLD_LIBRARY_PATH';
+  }
+  if (platform === 'win32') {
+    return 'PATH';
+  }
+  return null;
+}
+
+function prependEnvPath(existing, value) {
+  const parts = String(existing ?? '').split(path.delimiter).filter(Boolean);
+  if (parts.includes(value)) {
+    return parts.join(path.delimiter);
+  }
+  return [value, ...parts].join(path.delimiter);
+}
+
+/**
+ * Case-insensitive env key lookup: on Windows `{...process.env}` yields a
+ * plain object where PATH may be stored as `Path`. Using a hardcoded 'PATH'
+ * would create a duplicate key and break the child process PATH.
+ */
+function findEnvKey(env, key) {
+  const lower = key.toLowerCase();
+  for (const k of Object.keys(env)) {
+    if (k.toLowerCase() === lower) {
+      return k;
+    }
+  }
+  return key;
+}
+
+function resolveSherpaLibDir(platform = process.platform, arch = process.arch) {
+  const packageName = sherpaPlatformPackageName(platform, arch);
+  try {
+    const pkgJson = require.resolve(`${packageName}/package.json`);
+    // Electron packages node_modules inside app.asar, but native addons and
+    // their shared libraries are extracted to app.asar.unpacked. The dynamic
+    // loader (dlopen/DYLD/LD) cannot read from the asar archive, so point the
+    // search path at the unpacked copy.
+    const dir = path.dirname(pkgJson);
+    const unpacked = dir.replace(`app.asar${path.sep}`, `app.asar.unpacked${path.sep}`);
+    return existsSync(unpacked) ? unpacked : dir;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Prepend the sherpa platform package dir to the loader search path env var.
+ * Mutates the provided env object.
+ * @param {NodeJS.ProcessEnv} env
+ */
+export function applySherpaLoaderEnv(env) {
+  const key = sherpaLoaderEnvKey();
+  const libDir = resolveSherpaLibDir();
+  if (!key || !libDir) {
+    return { key: null, libDir: null };
+  }
+  const actualKey = findEnvKey(env, key);
+  env[actualKey] = prependEnvPath(env[actualKey], libDir);
+  return { key, libDir };
+}
+
+/**
+ * Load the sherpa-onnx-node module, trying the upstream entry first and then
+ * the platform addon directly.
+ */
+export function loadSherpaOnnxNode() {
+  if (cached) {
+    return cached;
+  }
+
+  const attempts = [];
+
+  try {
+    cached = require('sherpa-onnx-node');
+    return cached;
+  } catch (error) {
+    attempts.push(`sherpa-onnx-node: ${error?.message || String(error)}`);
+  }
+
+  const libDir = resolveSherpaLibDir();
+  if (libDir) {
+    applySherpaLoaderEnv(process.env);
+    const addonPath = path.join(libDir, 'sherpa-onnx.node');
+    if (existsSync(addonPath)) {
+      try {
+        cached = require(addonPath);
+        return cached;
+      } catch (error) {
+        attempts.push(`${addonPath}: ${error?.message || String(error)}`);
+      }
+    } else {
+      attempts.push(`${addonPath}: file not found`);
+    }
+  } else {
+    attempts.push(`${sherpaPlatformPackageName()}: platform package not installed`);
+  }
+
+  throw new Error(
+    [
+      `Failed to load sherpa-onnx-node for ${process.platform}-${process.arch}.`,
+      `Node ${process.version} (ABI ${process.versions.modules}).`,
+      'Load attempts:',
+      ...attempts.map((line) => `- ${line}`),
+    ].join('\n'),
+  );
+}

packages/web/server/lib/dictation/local/sherpa-recognizer.js

@@ -0,0 +1,277 @@
+/**
+ * Sherpa-onnx offline recognizer engine (NeMo transducer / Parakeet) plus a
+ * realtime streaming transcription session that re-decodes the accumulated
+ * segment audio on a throttle to produce live partial transcripts.
+ *
+ * Runs inside the dictation worker process only — never load the native
+ * addon in the main server process.
+ */
+
+import { EventEmitter } from 'events';
+import { existsSync } from 'fs';
+import { randomUUID } from 'crypto';
+
+import { loadSherpaOnnxNode } from './sherpa-loader.js';
+import { pcm16lePeakAbs, pcm16leToFloat32 } from '../audio.js';
+
+function assertFileExists(filePath, label) {
+  if (!existsSync(filePath)) {
+    throw new Error(`Missing ${label}: ${filePath}`);
+  }
+}
+
+export class SherpaOfflineRecognizerEngine {
+  /**
+   * @param {{ type: 'nemo_transducer' | 'whisper',
+   *           encoder: string, decoder: string, joiner?: string, tokens: string,
+   *           numThreads?: number }} config
+   */
+  constructor(config) {
+    assertFileExists(config.encoder, 'offline encoder');
+    assertFileExists(config.decoder, 'offline decoder');
+    if (config.type === 'nemo_transducer') {
+      assertFileExists(config.joiner, 'offline joiner');
+    }
+    assertFileExists(config.tokens, 'tokens');
+
+    const sherpa = loadSherpaOnnxNode();
+
+    const modelConfig =
+      config.type === 'whisper'
+        ? {
+            whisper: {
+              encoder: config.encoder,
+              decoder: config.decoder,
+              // Empty language auto-detects for multilingual Whisper exports.
+              language: '',
+              task: 'transcribe',
+              tailPaddings: -1,
+            },
+            tokens: config.tokens,
+            modelType: 'whisper',
+            numThreads: config.numThreads ?? 2,
+            provider: 'cpu',
+            debug: 0,
+          }
+        : {
+            transducer: {
+              encoder: config.encoder,
+              decoder: config.decoder,
+              joiner: config.joiner,
+            },
+            tokens: config.tokens,
+            modelType: 'nemo_transducer',
+            numThreads: config.numThreads ?? 2,
+            provider: 'cpu',
+            debug: 0,
+          };
+
+    const recognizerConfig = {
+      featConfig: {
+        sampleRate: 16000,
+        featureDim: 80,
+      },
+      modelConfig,
+      decodingMethod: 'greedy_search',
+      maxActivePaths: 4,
+    };
+
+    this.recognizer = new sherpa.OfflineRecognizer(recognizerConfig);
+    const sr = this.recognizer?.config?.featConfig?.sampleRate;
+    this.sampleRate =
+      typeof sr === 'number' && Number.isFinite(sr) && sr > 0
+        ? sr
+        : recognizerConfig.featConfig.sampleRate;
+  }
+
+  createStream() {
+    return this.recognizer.createStream();
+  }
+
+  acceptWaveform(stream, sampleRate, samples) {
+    if (!stream || typeof stream.acceptWaveform !== 'function') {
+      throw new Error('Unexpected sherpa offline stream: missing acceptWaveform()');
+    }
+    // sherpa-onnx-node expects acceptWaveform({ samples, sampleRate });
+    // the WASM build expects acceptWaveform(sampleRate, samples).
+    if (stream.acceptWaveform.length <= 1) {
+      stream.acceptWaveform({ samples, sampleRate });
+    } else {
+      stream.acceptWaveform(sampleRate, samples);
+    }
+  }
+
+  /**
+   * Decode a full PCM16 segment and return its text.
+   * Applies auto-gain when the peak is low so quiet microphones still decode.
+   * @param {Buffer} pcm16
+   * @returns {string}
+   */
+  decodePcm16(pcm16) {
+    if (pcm16.length === 0) {
+      return '';
+    }
+
+    const peak = pcm16lePeakAbs(pcm16);
+    const peakFloat = peak / 32768.0;
+    const targetPeak = 0.6;
+    const maxGain = 50;
+    const gain =
+      peakFloat > 0 && peakFloat < targetPeak ? Math.min(maxGain, targetPeak / peakFloat) : 1;
+
+    const stream = this.createStream();
+    try {
+      const floatSamples = pcm16leToFloat32(pcm16, gain);
+      this.acceptWaveform(stream, this.sampleRate, floatSamples);
+      this.recognizer.decode(stream);
+      const result = this.recognizer.getResult(stream);
+      const text =
+        typeof result === 'object' && result && 'text' in result ? result.text : result;
+      return String(text ?? '').trim();
+    } finally {
+      try {
+        stream.free?.();
+      } catch {
+        // ignore
+      }
+    }
+  }
+
+  free() {
+    try {
+      this.recognizer?.free?.();
+    } catch {
+      // ignore
+    }
+  }
+}
+
+/**
+ * Streaming transcription session backed by the offline recognizer.
+ * Accumulates the current segment's PCM and re-decodes it at most every
+ * `minDecodeIntervalMs` to emit non-final partial transcripts; `commit()`
+ * finalizes the segment and starts a new one.
+ *
+ * Implements the StreamingTranscriptionSession contract used by
+ * DictationStreamManager.
+ */
+export class SherpaRealtimeTranscriptionSession extends EventEmitter {
+  /**
+   * @param {{ engine: SherpaOfflineRecognizerEngine, minDecodeIntervalMs?: number }} params
+   */
+  constructor({ engine, minDecodeIntervalMs }) {
+    super();
+    this.engine = engine;
+    this.requiredSampleRate = engine.sampleRate;
+    this.minDecodeIntervalMs = minDecodeIntervalMs ?? 350;
+    this.connected = false;
+    this.currentSegmentId = null;
+    this.previousSegmentId = null;
+    this.lastPartialText = '';
+    this.pcm16 = Buffer.alloc(0);
+    this.lastDecodeAt = 0;
+    this.decoding = false;
+    this.pendingDecode = false;
+  }
+
+  async connect() {
+    if (this.connected) {
+      return;
+    }
+    this.currentSegmentId = randomUUID();
+    this.connected = true;
+  }
+
+  appendPcm16(chunk) {
+    if (!this.connected || !this.currentSegmentId) {
+      this.emit('error', new Error('Sherpa realtime session not connected'));
+      return;
+    }
+    this.pcm16 = this.pcm16.length === 0 ? chunk : Buffer.concat([this.pcm16, chunk]);
+    this.maybeDecode(false).catch((err) => {
+      this.emit('error', err instanceof Error ? err : new Error(String(err)));
+    });
+  }
+
+  commit() {
+    if (!this.connected || !this.currentSegmentId) {
+      this.emit('error', new Error('Sherpa realtime session not connected'));
+      return;
+    }
+
+    void (async () => {
+      try {
+        await this.maybeDecode(true);
+        const finalText = this.lastPartialText;
+        const segmentId = this.currentSegmentId;
+        const previousSegmentId = this.previousSegmentId;
+
+        this.emit('committed', { segmentId, previousSegmentId });
+        this.emit('transcript', { segmentId, transcript: finalText, isFinal: true });
+
+        this.previousSegmentId = segmentId;
+        this.currentSegmentId = randomUUID();
+        this.lastPartialText = '';
+        this.pcm16 = Buffer.alloc(0);
+      } catch (err) {
+        this.emit('error', err instanceof Error ? err : new Error(String(err)));
+      }
+    })();
+  }
+
+  clear() {
+    if (!this.connected) {
+      return;
+    }
+    this.pcm16 = Buffer.alloc(0);
+    this.currentSegmentId = randomUUID();
+    this.lastPartialText = '';
+  }
+
+  close() {
+    this.connected = false;
+    this.currentSegmentId = null;
+    this.pcm16 = Buffer.alloc(0);
+  }
+
+  async maybeDecode(force) {
+    if (!this.connected || !this.currentSegmentId) {
+      return;
+    }
+
+    const now = Date.now();
+    if (!force && now - this.lastDecodeAt < this.minDecodeIntervalMs) {
+      return;
+    }
+
+    if (this.decoding) {
+      this.pendingDecode = true;
+      return;
+    }
+
+    this.decoding = true;
+    try {
+      const decodeStartedAt = Date.now();
+      const text = this.engine.decodePcm16(this.pcm16);
+      this.lastDecodeAt = Date.now();
+      // Adaptive throttle: on slow hardware (or heavy models) re-decoding the
+      // growing segment every 350ms would monopolize the worker. Space partial
+      // decodes to ~1.5x the observed decode time.
+      this.minDecodeIntervalMs = Math.max(350, (this.lastDecodeAt - decodeStartedAt) * 1.5);
+      if (text !== this.lastPartialText) {
+        this.lastPartialText = text;
+        this.emit('transcript', {
+          segmentId: this.currentSegmentId,
+          transcript: text,
+          isFinal: false,
+        });
+      }
+    } finally {
+      this.decoding = false;
+      if (this.pendingDecode) {
+        this.pendingDecode = false;
+        await this.maybeDecode(true);
+      }
+    }
+  }
+}

packages/web/server/lib/dictation/local/sherpa-tts.js

@@ -0,0 +1,110 @@
+/**
+ * Sherpa-onnx offline TTS (Kokoro). Runs inside the dictation worker process
+ * only — never load the native addon in the main server process.
+ */
+
+import { existsSync } from 'fs';
+import path from 'path';
+
+import { loadSherpaOnnxNode } from './sherpa-loader.js';
+
+function assertFileExists(filePath, label) {
+  if (!existsSync(filePath)) {
+    throw new Error(`Missing ${label}: ${filePath}`);
+  }
+}
+
+function float32ToPcm16le(samples) {
+  const out = new Int16Array(samples.length);
+  for (let i = 0; i < samples.length; i += 1) {
+    const clamped = Math.max(-1, Math.min(1, samples[i]));
+    out[i] = Math.round(clamped * 32767);
+  }
+  return Buffer.from(out.buffer, out.byteOffset, out.byteLength);
+}
+
+export class SherpaTtsEngine {
+  /**
+   * @param {{ modelDir: string, files: { model: string, voices: string, tokens: string, espeakData: string }, numThreads?: number }} config
+   */
+  constructor(config) {
+    const modelPath = path.join(config.modelDir, config.files.model);
+    const voicesPath = path.join(config.modelDir, config.files.voices);
+    const tokensPath = path.join(config.modelDir, config.files.tokens);
+    const dataDir = path.join(config.modelDir, config.files.espeakData);
+
+    assertFileExists(modelPath, 'TTS model');
+    assertFileExists(voicesPath, 'TTS voices');
+    assertFileExists(tokensPath, 'TTS tokens');
+    assertFileExists(dataDir, 'TTS espeak-ng dataDir');
+
+    const sherpa = loadSherpaOnnxNode();
+    if (typeof sherpa.OfflineTts !== 'function') {
+      throw new Error('sherpa-onnx-node OfflineTts is unavailable');
+    }
+
+    this.tts = new sherpa.OfflineTts({
+      model: {
+        kokoro: {
+          model: modelPath,
+          voices: voicesPath,
+          tokens: tokensPath,
+          dataDir,
+          lengthScale: 1.0,
+        },
+      },
+      numThreads: config.numThreads ?? 2,
+      provider: 'cpu',
+      maxNumSentences: 1,
+    });
+  }
+
+  /**
+   * Synthesize text to PCM16LE.
+   * @param {string} text
+   * @param {{ speakerId?: number, speed?: number }} [options]
+   * @returns {{ pcm16: Buffer, sampleRate: number }}
+   */
+  synthesize(text, options = {}) {
+    const trimmed = String(text || '').trim();
+    if (!trimmed) {
+      throw new Error('Cannot synthesize empty text');
+    }
+
+    const audio = this.tts.generate({
+      text: trimmed,
+      sid: Number.isInteger(options.speakerId) ? options.speakerId : 0,
+      speed: typeof options.speed === 'number' && options.speed > 0 ? options.speed : 1.0,
+      // Request a copied buffer from sherpa itself: native external-backed
+      // typed arrays are rejected by Electron.
+      enableExternalBuffer: false,
+    });
+
+    let samples = null;
+    if (audio && audio.samples instanceof Float32Array) {
+      samples = Float32Array.from(audio.samples);
+    } else if (audio && Array.isArray(audio.samples)) {
+      samples = Float32Array.from(audio.samples);
+    }
+    if (!samples) {
+      throw new Error('Unexpected sherpa TTS output: missing Float32 samples');
+    }
+
+    const sampleRate =
+      audio && typeof audio.sampleRate === 'number' && audio.sampleRate > 0
+        ? audio.sampleRate
+        : typeof this.tts.sampleRate === 'number' && this.tts.sampleRate > 0
+          ? this.tts.sampleRate
+          : 24000;
+
+    return { pcm16: float32ToPcm16le(samples), sampleRate };
+  }
+
+  free() {
+    try {
+      this.tts?.free?.();
+    } catch {
+      // ignore
+    }
+  }
+}

packages/web/server/lib/dictation/local/worker-client.js

@@ -0,0 +1,352 @@
+/**
+ * Client for the dictation local-speech worker process.
+ *
+ * Lazily forks the worker on first use, correlates request/response messages
+ * by requestId, routes session events to per-session EventEmitters, and
+ * shuts the worker down after an idle TTL so the ONNX runtime does not sit
+ * in memory while dictation is unused.
+ */
+
+import { fork } from 'child_process';
+import { randomUUID } from 'crypto';
+import { EventEmitter } from 'events';
+import { fileURLToPath } from 'url';
+
+import { applySherpaLoaderEnv } from './sherpa-loader.js';
+
+const DEFAULT_REQUEST_TIMEOUT_MS = 30000;
+const DEFAULT_IDLE_TTL_MS = 5 * 60 * 1000;
+const DEFAULT_LOCAL_SAMPLE_RATE = 16000;
+const STDERR_TAIL_MAX_CHARS = 2000;
+
+function forkDictationWorker() {
+  const env = { ...process.env };
+  applySherpaLoaderEnv(env);
+  return fork(fileURLToPath(new URL('./worker-process.js', import.meta.url)), [], {
+    env,
+    serialization: 'advanced',
+    stdio: ['ignore', 'ignore', 'pipe', 'ipc'],
+    windowsHide: true,
+  });
+}
+
+export class DictationWorkerClient {
+  /**
+   * @param {{ requestTimeoutMs?: number, idleTtlMs?: number }} [options]
+   */
+  constructor(options = {}) {
+    this.requestTimeoutMs = options.requestTimeoutMs ?? DEFAULT_REQUEST_TIMEOUT_MS;
+    this.idleTtlMs = options.idleTtlMs ?? DEFAULT_IDLE_TTL_MS;
+    this.pendingRequests = new Map();
+    this.sessionEmitters = new Map();
+    this.worker = null;
+    this.stderrTail = '';
+    this.inFlightRequests = 0;
+    this.idleTimer = null;
+    this.intentionalCloses = new WeakSet();
+  }
+
+  /**
+   * Synthesize speech in the worker. Returns WAV bytes.
+   * @param {{ modelsDir: string, modelId: string, text: string, speakerId?: number, speed?: number }} params
+   * @returns {Promise<{ audio: Buffer, format: string }>}
+   */
+  async synthesizeSpeech(params) {
+    // Long texts on slow hardware can exceed the default request timeout.
+    const result = await this.sendRequest(
+      { type: 'tts.synthesize', ...params },
+      { timeoutMs: 120000 },
+    );
+    return {
+      audio: Buffer.isBuffer(result.audio) ? result.audio : Buffer.from(result.audio),
+      format: result.format || 'audio/wav',
+    };
+  }
+
+  /**
+   * Create a streaming STT session in the worker.
+   * @param {{ modelsDir: string, modelId: string }} params
+   * @param {EventEmitter} emitter receives 'committed' | 'transcript' | 'error'
+   * @returns {Promise<{ sessionId: string, requiredSampleRate: number }>}
+   */
+  async createSession({ modelsDir, modelId }, emitter) {
+    const sessionId = randomUUID();
+    this.sessionEmitters.set(sessionId, emitter);
+    try {
+      const result = await this.sendRequest({
+        type: 'session.create',
+        sessionId,
+        modelsDir,
+        modelId,
+      });
+      return { sessionId, requiredSampleRate: result?.requiredSampleRate ?? DEFAULT_LOCAL_SAMPLE_RATE };
+    } catch (err) {
+      this.sessionEmitters.delete(sessionId);
+      this.scheduleIdleShutdownIfReady();
+      throw err;
+    }
+  }
+
+  appendSessionAudio(sessionId, audio) {
+    void this.sendRequest({ type: 'session.append', sessionId, audio }).catch((err) => {
+      this.emitSessionError(sessionId, err);
+    });
+  }
+
+  commitSession(sessionId) {
+    void this.sendRequest({ type: 'session.commit', sessionId }).catch((err) => {
+      this.emitSessionError(sessionId, err);
+    });
+  }
+
+  clearSession(sessionId) {
+    void this.sendRequest({ type: 'session.clear', sessionId }).catch((err) => {
+      this.emitSessionError(sessionId, err);
+    });
+  }
+
+  closeSession(sessionId) {
+    this.sessionEmitters.delete(sessionId);
+    void this.sendRequest({ type: 'session.close', sessionId }).catch(() => {
+      // Closing is best-effort; the parent already dropped the session.
+    });
+    this.scheduleIdleShutdownIfReady();
+  }
+
+  shutdown() {
+    this.clearIdleTimer();
+    this.rejectAllPending(new Error('Dictation worker shut down'));
+    this.sessionEmitters.clear();
+    const worker = this.worker;
+    this.worker = null;
+    if (worker && !worker.killed) {
+      this.intentionalCloses.add(worker);
+      try {
+        worker.disconnect();
+      } catch {
+        // ignore
+      }
+      try {
+        worker.kill();
+      } catch {
+        // ignore
+      }
+    }
+  }
+
+  sendRequest(input, options = {}) {
+    const worker = this.ensureWorker();
+    const requestId = randomUUID();
+    const message = { ...input, requestId };
+    this.inFlightRequests += 1;
+    this.clearIdleTimer();
+
+    return new Promise((resolve, reject) => {
+      const timeout = setTimeout(() => {
+        this.pendingRequests.delete(requestId);
+        this.inFlightRequests = Math.max(0, this.inFlightRequests - 1);
+        this.scheduleIdleShutdownIfReady();
+        reject(new Error(`Dictation worker request timed out: ${input.type}`));
+      }, options.timeoutMs ?? this.requestTimeoutMs);
+
+      this.pendingRequests.set(requestId, { resolve, reject, timeout });
+
+      worker.send(message, (error) => {
+        if (!error) {
+          return;
+        }
+        const pending = this.pendingRequests.get(requestId);
+        if (!pending) {
+          return;
+        }
+        clearTimeout(pending.timeout);
+        this.pendingRequests.delete(requestId);
+        this.inFlightRequests = Math.max(0, this.inFlightRequests - 1);
+        this.scheduleIdleShutdownIfReady();
+        pending.reject(error);
+      });
+    });
+  }
+
+  ensureWorker() {
+    if (this.worker && !this.worker.killed && this.worker.connected) {
+      return this.worker;
+    }
+    const worker = forkDictationWorker();
+    this.worker = worker;
+    this.stderrTail = '';
+    worker.stderr?.on('data', (chunk) => {
+      const text = Buffer.isBuffer(chunk) ? chunk.toString('utf8') : String(chunk);
+      this.stderrTail = (this.stderrTail + text).slice(-STDERR_TAIL_MAX_CHARS);
+    });
+    worker.on('message', (message) => this.handleWorkerMessage(message));
+    worker.on('close', (code, signal) => this.handleWorkerExit(worker, code, signal));
+    return worker;
+  }
+
+  handleWorkerMessage(message) {
+    if (message?.type === 'response') {
+      const pending = this.pendingRequests.get(message.requestId);
+      if (!pending) {
+        return;
+      }
+      clearTimeout(pending.timeout);
+      this.pendingRequests.delete(message.requestId);
+      this.inFlightRequests = Math.max(0, this.inFlightRequests - 1);
+      this.scheduleIdleShutdownIfReady();
+      if (message.ok) {
+        pending.resolve(message.result);
+      } else {
+        pending.reject(new Error(message.error || 'Dictation worker request failed'));
+      }
+      return;
+    }
+
+    const emitter = this.sessionEmitters.get(message?.sessionId);
+    if (!emitter) {
+      return;
+    }
+    switch (message.type) {
+      case 'session.committed':
+        emitter.emit('committed', message.payload);
+        return;
+      case 'session.transcript':
+        emitter.emit('transcript', message.payload);
+        return;
+      case 'session.error':
+        emitter.emit('error', new Error(message.error));
+        return;
+      default:
+    }
+  }
+
+  handleWorkerExit(worker, code, signal) {
+    const wasCurrentWorker = this.worker === worker;
+    const wasIntentional = this.intentionalCloses.has(worker);
+    this.intentionalCloses.delete(worker);
+    if (!wasCurrentWorker || wasIntentional) {
+      if (wasCurrentWorker) {
+        this.worker = null;
+      }
+      return;
+    }
+
+    const stderr = this.stderrTail.trim();
+    const error = new Error(
+      `Dictation worker exited (code ${code ?? 'null'}${signal ? `, signal ${signal}` : ''}).` +
+        (stderr ? ` Last stderr: ${stderr.slice(-500)}` : ''),
+    );
+
+    this.worker = null;
+    this.clearIdleTimer();
+    this.rejectAllPending(error);
+    for (const emitter of this.sessionEmitters.values()) {
+      if (emitter.listenerCount('error') > 0) {
+        emitter.emit('error', error);
+      }
+    }
+    this.sessionEmitters.clear();
+    this.inFlightRequests = 0;
+  }
+
+  rejectAllPending(error) {
+    for (const [requestId, pending] of this.pendingRequests) {
+      clearTimeout(pending.timeout);
+      pending.reject(error);
+      this.pendingRequests.delete(requestId);
+    }
+  }
+
+  emitSessionError(sessionId, error) {
+    const emitter = this.sessionEmitters.get(sessionId);
+    if (emitter && emitter.listenerCount('error') > 0) {
+      emitter.emit('error', error instanceof Error ? error : new Error(String(error)));
+    }
+  }
+
+  scheduleIdleShutdownIfReady() {
+    if (!this.worker || this.inFlightRequests > 0 || this.sessionEmitters.size > 0) {
+      return;
+    }
+    this.clearIdleTimer();
+    this.idleTimer = setTimeout(() => {
+      if (this.inFlightRequests === 0 && this.sessionEmitters.size === 0) {
+        this.shutdown();
+      }
+    }, this.idleTtlMs);
+  }
+
+  clearIdleTimer() {
+    if (this.idleTimer) {
+      clearTimeout(this.idleTimer);
+      this.idleTimer = null;
+    }
+  }
+}
+
+/**
+ * StreamingTranscriptionSession backed by the worker process.
+ * Matches the session contract consumed by DictationStreamManager.
+ */
+export class WorkerBackedTranscriptionSession extends EventEmitter {
+  /**
+   * @param {DictationWorkerClient} client
+   * @param {{ modelsDir: string, modelId: string }} modelConfig
+   */
+  constructor(client, modelConfig) {
+    super();
+    this.client = client;
+    this.modelConfig = modelConfig;
+    this.requiredSampleRate = DEFAULT_LOCAL_SAMPLE_RATE;
+    this.connectedSessionId = null;
+    this.connecting = null;
+  }
+
+  async connect() {
+    if (this.connectedSessionId) {
+      return;
+    }
+    if (!this.connecting) {
+      this.connecting = (async () => {
+        try {
+          const result = await this.client.createSession(this.modelConfig, this);
+          this.connectedSessionId = result.sessionId;
+          this.requiredSampleRate = result.requiredSampleRate;
+        } finally {
+          this.connecting = null;
+        }
+      })();
+    }
+    await this.connecting;
+  }
+
+  appendPcm16(pcm16le) {
+    if (!this.connectedSessionId) {
+      this.emit('error', new Error('Local STT session not connected'));
+      return;
+    }
+    this.client.appendSessionAudio(this.connectedSessionId, pcm16le);
+  }
+
+  commit() {
+    if (!this.connectedSessionId) {
+      this.emit('error', new Error('Local STT session not connected'));
+      return;
+    }
+    this.client.commitSession(this.connectedSessionId);
+  }
+
+  clear() {
+    if (this.connectedSessionId) {
+      this.client.clearSession(this.connectedSessionId);
+    }
+  }
+
+  close() {
+    const sessionId = this.connectedSessionId;
+    this.connectedSessionId = null;
+    if (sessionId) {
+      this.client.closeSession(sessionId);
+    }
+  }
+}

packages/web/server/lib/dictation/local/worker-process.js

@@ -0,0 +1,197 @@
+/**
+ * Dictation local-speech worker process.
+ *
+ * Hosts the sherpa-onnx native inference (Parakeet STT) in a separate process
+ * so ONNX decoding never blocks the main OpenChamber server. Communicates
+ * with the parent over child_process IPC (advanced serialization, so Buffers
+ * survive the trip as Uint8Array).
+ *
+ * Request/response protocol (parent -> worker):
+ *   { type: 'session.create', requestId, sessionId, modelsDir, modelId }
+ *   { type: 'session.append', requestId, sessionId, audio }
+ *   { type: 'session.commit' | 'session.clear' | 'session.close', requestId, sessionId }
+ * Worker -> parent:
+ *   { type: 'response', requestId, ok, result?, error? }
+ *   { type: 'session.committed' | 'session.transcript' | 'session.error', sessionId, ... }
+ */
+
+import {
+  SherpaOfflineRecognizerEngine,
+  SherpaRealtimeTranscriptionSession,
+} from './sherpa-recognizer.js';
+import { SherpaTtsEngine } from './sherpa-tts.js';
+import { getLocalSttModelDir, getLocalSttModelSpec } from './model-catalog.js';
+import { pcm16ToWav } from '../audio.js';
+import path from 'path';
+
+process.title = 'OpenChamber Dictation';
+
+const engines = new Map();
+const ttsEngines = new Map();
+const sessions = new Map();
+let ipcClosing = false;
+
+function sendToParent(message) {
+  if (ipcClosing || !process.connected || !process.send) {
+    return;
+  }
+  try {
+    process.send(message, (error) => {
+      if (error) {
+        ipcClosing = true;
+      }
+    });
+  } catch {
+    ipcClosing = true;
+  }
+}
+
+function sendOk(requestId, result) {
+  sendToParent({ type: 'response', requestId, ok: true, ...(result !== undefined ? { result } : {}) });
+}
+
+function getEngine(modelsDir, modelId) {
+  const key = `${modelsDir}:${modelId}`;
+  const existing = engines.get(key);
+  if (existing) {
+    return existing;
+  }
+  const modelDir = getLocalSttModelDir(modelsDir, modelId);
+  const spec = getLocalSttModelSpec(modelId);
+  const created = new SherpaOfflineRecognizerEngine({
+    type: spec.type,
+    encoder: path.join(modelDir, spec.files.encoder),
+    decoder: path.join(modelDir, spec.files.decoder),
+    ...(spec.files.joiner ? { joiner: path.join(modelDir, spec.files.joiner) } : {}),
+    tokens: path.join(modelDir, spec.files.tokens),
+    numThreads: 2,
+  });
+  engines.set(key, created);
+  return created;
+}
+
+function cleanupSession(sessionId) {
+  const session = sessions.get(sessionId);
+  sessions.delete(sessionId);
+  try {
+    session?.close();
+  } catch {
+    // ignore
+  }
+}
+
+function toBuffer(audio) {
+  if (Buffer.isBuffer(audio)) {
+    return audio;
+  }
+  if (audio instanceof Uint8Array) {
+    return Buffer.from(audio.buffer, audio.byteOffset, audio.byteLength);
+  }
+  if (audio && typeof audio === 'object' && audio.type === 'Buffer' && Array.isArray(audio.data)) {
+    return Buffer.from(audio.data);
+  }
+  throw new Error('Unsupported audio payload in dictation worker');
+}
+
+function getTtsEngine(modelsDir, modelId) {
+  const key = `${modelsDir}:${modelId}`;
+  const existing = ttsEngines.get(key);
+  if (existing) {
+    return existing;
+  }
+  const spec = getLocalSttModelSpec(modelId);
+  const created = new SherpaTtsEngine({
+    modelDir: getLocalSttModelDir(modelsDir, modelId),
+    files: spec.files,
+    numThreads: 2,
+  });
+  ttsEngines.set(key, created);
+  return created;
+}
+
+async function handleRequest(message) {
+  switch (message.type) {
+    case 'tts.synthesize': {
+      const engine = getTtsEngine(message.modelsDir, message.modelId);
+      const { pcm16, sampleRate } = engine.synthesize(message.text, {
+        speakerId: message.speakerId,
+        speed: message.speed,
+      });
+      sendOk(message.requestId, {
+        audio: pcm16ToWav(pcm16, sampleRate),
+        format: 'audio/wav',
+      });
+      return;
+    }
+    case 'session.create': {
+      cleanupSession(message.sessionId);
+      const engine = getEngine(message.modelsDir, message.modelId);
+      const session = new SherpaRealtimeTranscriptionSession({ engine });
+      session.on('committed', (payload) => {
+        sendToParent({ type: 'session.committed', sessionId: message.sessionId, payload });
+      });
+      session.on('transcript', (payload) => {
+        sendToParent({ type: 'session.transcript', sessionId: message.sessionId, payload });
+      });
+      session.on('error', (err) => {
+        sendToParent({
+          type: 'session.error',
+          sessionId: message.sessionId,
+          error: err instanceof Error ? err.message : String(err),
+        });
+      });
+      await session.connect();
+      sessions.set(message.sessionId, session);
+      sendOk(message.requestId, { requiredSampleRate: session.requiredSampleRate });
+      return;
+    }
+    case 'session.append': {
+      sessions.get(message.sessionId)?.appendPcm16(toBuffer(message.audio));
+      sendOk(message.requestId);
+      return;
+    }
+    case 'session.commit': {
+      sessions.get(message.sessionId)?.commit();
+      sendOk(message.requestId);
+      return;
+    }
+    case 'session.clear': {
+      sessions.get(message.sessionId)?.clear();
+      sendOk(message.requestId);
+      return;
+    }
+    case 'session.close': {
+      cleanupSession(message.sessionId);
+      sendOk(message.requestId);
+      return;
+    }
+    default: {
+      throw new Error(`Unknown dictation worker request: ${message?.type}`);
+    }
+  }
+}
+
+process.on('message', (message) => {
+  void handleRequest(message).catch((error) => {
+    sendToParent({
+      type: 'response',
+      requestId: message?.requestId,
+      ok: false,
+      error: error instanceof Error ? error.message : 'Dictation worker request failed',
+    });
+  });
+});
+
+process.once('disconnect', () => {
+  ipcClosing = true;
+  for (const sessionId of Array.from(sessions.keys())) {
+    cleanupSession(sessionId);
+  }
+  for (const engine of engines.values()) {
+    engine.free();
+  }
+  for (const tts of ttsEngines.values()) {
+    tts.free();
+  }
+  process.exit(0);
+});

packages/web/server/lib/dictation/openai-compatible-session.js

@@ -0,0 +1,98 @@
+/**
+ * Pseudo-streaming transcription session for OpenAI-compatible Whisper
+ * endpoints (faster-whisper, whisper.cpp, OpenAI, ...).
+ *
+ * The Whisper HTTP API cannot stream, so audio is buffered per segment and
+ * transcribed on commit(). Live partials therefore only advance at segment
+ * boundaries (the DictationStreamManager auto-commits every ~15s of speech).
+ *
+ * Implements the StreamingTranscriptionSession contract used by
+ * DictationStreamManager.
+ */
+
+import { EventEmitter } from 'events';
+import { randomUUID } from 'crypto';
+
+import { transcribeAudio } from '../tts/stt.js';
+import { pcm16ToWav } from './audio.js';
+
+const OPENAI_COMPATIBLE_SAMPLE_RATE = 16000;
+
+export class OpenAICompatibleTranscriptionSession extends EventEmitter {
+  /**
+   * @param {{ baseURL: string, model: string, apiKey?: string, language?: string, prompt?: string }} config
+   */
+  constructor(config) {
+    super();
+    this.config = config;
+    this.requiredSampleRate = OPENAI_COMPATIBLE_SAMPLE_RATE;
+    this.connected = false;
+    this.segmentId = randomUUID();
+    this.previousSegmentId = null;
+    this.pcm16 = Buffer.alloc(0);
+  }
+
+  async connect() {
+    if (!this.config.baseURL) {
+      throw new Error('Custom STT server URL is not configured');
+    }
+    if (!this.config.model) {
+      throw new Error('STT model is not configured');
+    }
+    this.connected = true;
+  }
+
+  appendPcm16(chunk) {
+    if (!this.connected) {
+      this.emit('error', new Error('STT session not connected'));
+      return;
+    }
+    this.pcm16 = this.pcm16.length === 0 ? chunk : Buffer.concat([this.pcm16, chunk]);
+  }
+
+  commit() {
+    if (!this.connected) {
+      this.emit('error', new Error('STT session not connected'));
+      return;
+    }
+
+    const committedId = this.segmentId;
+    const previousSegmentId = this.previousSegmentId;
+    const committedPcm16 = this.pcm16;
+    this.previousSegmentId = committedId;
+    this.segmentId = randomUUID();
+    this.pcm16 = Buffer.alloc(0);
+    this.emit('committed', { segmentId: committedId, previousSegmentId });
+
+    void (async () => {
+      try {
+        const wav = pcm16ToWav(committedPcm16, OPENAI_COMPATIBLE_SAMPLE_RATE);
+        const text = await transcribeAudio({
+          audioBuffer: wav,
+          mimeType: 'audio/wav',
+          model: this.config.model,
+          baseURL: this.config.baseURL,
+          apiKey: this.config.apiKey,
+          language: this.config.language,
+        });
+        this.emit('transcript', {
+          segmentId: committedId,
+          transcript: (text ?? '').trim(),
+          isFinal: true,
+        });
+      } catch (err) {
+        this.emit('error', err instanceof Error ? err : new Error(String(err)));
+      }
+    })();
+  }
+
+  clear() {
+    this.pcm16 = Buffer.alloc(0);
+    this.segmentId = randomUUID();
+  }
+
+  close() {
+    this.connected = false;
+    this.pcm16 = Buffer.alloc(0);
+  }
+}

packages/web/server/lib/dictation/runtime.js

@@ -0,0 +1,278 @@
+/**
+ * Dictation runtime: registers the streaming dictation WebSocket endpoint and
+ * the HTTP status/model routes.
+ *
+ * WebSocket protocol (JSON text frames) on /api/dictation/ws:
+ *   client -> server:
+ *     { type: 'start',  dictationId, format, options? }
+ *       options: { provider?, language?, localModel?, openaiCompatible? }
+ *     { type: 'chunk',  dictationId, seq, audio }   // audio: base64 PCM16LE
+ *     { type: 'finish', dictationId, finalSeq }
+ *     { type: 'cancel', dictationId }
+ *     { type: 'ping' }
+ *   server -> client:
+ *     { type: 'ready' }
+ *     { type: 'ack',             dictationId, ackSeq }
+ *     { type: 'partial',         dictationId, text }
+ *     { type: 'finish_accepted', dictationId, timeoutMs }
+ *     { type: 'final',           dictationId, text }
+ *     { type: 'error',           dictationId, error, retryable, reasonCode? }
+ *     { type: 'pong' }
+ */
+
+import { WebSocketServer } from 'ws';
+
+import { DictationStreamManager } from './stream-manager.js';
+import { createDictationService } from './service.js';
+
+const DICTATION_WS_PATH = '/api/dictation/ws';
+
+const DICTATION_WS_MAX_PAYLOAD_BYTES = 512 * 1024;
+const DICTATION_WS_HEARTBEAT_INTERVAL_MS = 30000;
+
+const parseRequestPathname = (url) => {
+  try {
+    return new URL(url, 'http://localhost').pathname;
+  } catch {
+    return typeof url === 'string' ? url.split('?')[0] : '';
+  }
+};
+
+export function createDictationRuntime({
+  app,
+  server,
+  express,
+  uiAuthController,
+  isRequestOriginAllowed,
+  rejectWebSocketUpgrade,
+  modelsDir,
+}) {
+  const service = createDictationService({ modelsDir });
+
+  // Local text-to-speech (Kokoro in the dictation worker). Returns WAV bytes;
+  // 503 with a reason code while the model is still downloading.
+  app.post('/api/dictation/tts/speak', express.json({ limit: '1mb' }), async (req, res) => {
+    try {
+      const text = typeof req.body?.text === 'string' ? req.body.text.trim() : '';
+      if (!text) {
+        res.status(400).json({ error: 'Text is required' });
+        return;
+      }
+      const result = await service.synthesizeSpeech({
+        text,
+        model: typeof req.body?.model === 'string' ? req.body.model : undefined,
+        speakerId: Number.isInteger(req.body?.speakerId) ? req.body.speakerId : undefined,
+        speed: typeof req.body?.speed === 'number' ? req.body.speed : undefined,
+      });
+      if (result.error) {
+        res.status(503).json({
+          error: result.error,
+          retryable: result.retryable !== false,
+          ...(result.reasonCode ? { reasonCode: result.reasonCode } : {}),
+        });
+        return;
+      }
+      res.setHeader('Content-Type', result.format || 'audio/wav');
+      res.send(result.audio);
+    } catch (error) {
+      res.status(500).json({ error: error?.message || 'Failed to synthesize speech' });
+    }
+  });
+
+  app.get('/api/dictation/status', async (req, res) => {
+    try {
+      const provider = typeof req.query.provider === 'string' ? req.query.provider : undefined;
+      const localModel = typeof req.query.localModel === 'string' ? req.query.localModel : undefined;
+      const status = await service.getStatus({ provider, localModel });
+      res.json(status);
+    } catch (error) {
+      res.status(500).json({ error: error?.message || 'Failed to read dictation status' });
+    }
+  });
+
+  app.post('/api/dictation/models/:modelId/download', async (req, res) => {
+    try {
+      const result = await service.requestModelDownload(req.params.modelId);
+      if (!result.ok) {
+        res.status(400).json({ error: result.error });
+        return;
+      }
+      res.json(result);
+    } catch (error) {
+      res.status(500).json({ error: error?.message || 'Failed to start model download' });
+    }
+  });
+
+  app.delete('/api/dictation/models/:modelId', async (req, res) => {
+    try {
+      const result = await service.deleteModel(req.params.modelId);
+      if (!result.ok) {
+        res.status(400).json({ error: result.error });
+        return;
+      }
+      res.json(result);
+    } catch (error) {
+      res.status(500).json({ error: error?.message || 'Failed to delete model' });
+    }
+  });
+
+  const wsServer = new WebSocketServer({
+    noServer: true,
+    maxPayload: DICTATION_WS_MAX_PAYLOAD_BYTES,
+  });
+
+  wsServer.on('connection', (socket) => {
+    const send = (msg) => {
+      if (socket.readyState !== 1) {
+        return;
+      }
+      try {
+        socket.send(JSON.stringify(msg));
+      } catch {
+        // socket is going away; the manager cleanup on close handles state
+      }
+    };
+
+    const manager = new DictationStreamManager({
+      emit: ({ type, payload }) => send({ type, ...payload }),
+      createSttSession: (options) => service.createSttSession(options),
+    });
+
+    send({ type: 'ready' });
+
+    const heartbeatInterval = setInterval(() => {
+      if (socket.readyState !== 1) {
+        return;
+      }
+      try {
+        socket.ping();
+      } catch {
+        // ignore
+      }
+    }, DICTATION_WS_HEARTBEAT_INTERVAL_MS);
+
+    socket.on('message', (raw, isBinary) => {
+      if (isBinary) {
+        return;
+      }
+      let message;
+      try {
+        message = JSON.parse(raw.toString('utf8'));
+      } catch {
+        return;
+      }
+      if (!message || typeof message !== 'object') {
+        return;
+      }
+
+      switch (message.type) {
+        case 'start': {
+          if (typeof message.dictationId !== 'string' || typeof message.format !== 'string') {
+            return;
+          }
+          const options =
+            message.options && typeof message.options === 'object' ? message.options : {};
+          void manager.handleStart(message.dictationId, message.format, options);
+          return;
+        }
+        case 'chunk': {
+          if (
+            typeof message.dictationId !== 'string' ||
+            typeof message.seq !== 'number' ||
+            typeof message.audio !== 'string'
+          ) {
+            return;
+          }
+          manager.handleChunk({
+            dictationId: message.dictationId,
+            seq: message.seq,
+            audioBase64: message.audio,
+          });
+          return;
+        }
+        case 'finish': {
+          if (typeof message.dictationId !== 'string' || typeof message.finalSeq !== 'number') {
+            return;
+          }
+          manager.handleFinish(message.dictationId, message.finalSeq);
+          return;
+        }
+        case 'cancel': {
+          if (typeof message.dictationId !== 'string') {
+            return;
+          }
+          manager.handleCancel(message.dictationId);
+          return;
+        }
+        case 'ping': {
+          send({ type: 'pong' });
+          return;
+        }
+        default:
+      }
+    });
+
+    socket.on('close', () => {
+      clearInterval(heartbeatInterval);
+      manager.cleanupAll();
+    });
+
+    socket.on('error', () => {
+      // 'close' follows and performs cleanup.
+    });
+  });
+
+  const upgradeHandler = (req, socket, head) => {
+    const pathname = parseRequestPathname(req.url);
+    if (pathname !== DICTATION_WS_PATH) {
+      return;
+    }
+
+    const handleUpgrade = async () => {
+      try {
+        if (uiAuthController?.enabled) {
+          const sessionToken = await uiAuthController?.ensureSessionToken?.(req, null);
+          if (!sessionToken) {
+            rejectWebSocketUpgrade(socket, 401, 'UI authentication required');
+            return;
+          }
+
+          const originAllowed = await isRequestOriginAllowed(req);
+          if (!originAllowed) {
+            rejectWebSocketUpgrade(socket, 403, 'Invalid origin');
+            return;
+          }
+        }
+
+        wsServer.handleUpgrade(req, socket, head, (ws) => {
+          wsServer.emit('connection', ws, req);
+        });
+      } catch {
+        rejectWebSocketUpgrade(socket, 500, 'Upgrade failed');
+      }
+    };
+
+    void handleUpgrade();
+  };
+
+  server.on('upgrade', upgradeHandler);
+
+  const stop = () => {
+    server.off('upgrade', upgradeHandler);
+    for (const client of wsServer.clients) {
+      try {
+        client.close(1001, 'server shutting down');
+      } catch {
+        // ignore
+      }
+    }
+    try {
+      wsServer.close();
+    } catch {
+      // ignore
+    }
+    service.shutdown();
+  };
+
+  return { stop };
+}

packages/web/server/lib/dictation/service.js

@@ -0,0 +1,302 @@
+/**
+ * Dictation service: resolves STT providers, tracks local model download
+ * state, and exposes a readiness snapshot for the status route.
+ *
+ * Providers:
+ * - 'local' (default): sherpa-onnx Parakeet running in a worker process.
+ *   Models auto-download in the background on first use.
+ * - 'openai-compatible': any OpenAI-compatible /v1/audio/transcriptions
+ *   endpoint (faster-whisper, whisper.cpp, OpenAI).
+ */
+
+import { rm } from 'fs/promises';
+
+import { DictationWorkerClient, WorkerBackedTranscriptionSession } from './local/worker-client.js';
+import { OpenAICompatibleTranscriptionSession } from './openai-compatible-session.js';
+import {
+  DEFAULT_LOCAL_STT_MODEL,
+  DEFAULT_LOCAL_TTS_MODEL,
+  LOCAL_STT_MODEL_CATALOG,
+  LOCAL_STT_MODEL_IDS,
+  LOCAL_TTS_MODEL_CATALOG,
+  LOCAL_TTS_MODEL_IDS,
+  getLocalSttModelDir,
+  isLocalModelId,
+  isLocalSttModelId,
+  isLocalTtsModelId,
+} from './local/model-catalog.js';
+import { ensureLocalSttModel, isLocalSttModelInstalled } from './local/model-downloader.js';
+
+export function createDictationService({ modelsDir }) {
+  const workerClient = new DictationWorkerClient();
+  /** modelId -> 'downloading' | 'error' */
+  const downloadStates = new Map();
+  /** modelId -> last download error message */
+  const downloadErrors = new Map();
+  /** modelId -> in-flight ensure promise */
+  const downloadPromises = new Map();
+  /** modelId -> 0..100 download percent (null while size unknown) */
+  const downloadProgress = new Map();
+
+  const startModelDownload = (modelId) => {
+    const existing = downloadPromises.get(modelId);
+    if (existing) {
+      return existing;
+    }
+    downloadStates.set(modelId, 'downloading');
+    downloadErrors.delete(modelId);
+    downloadProgress.set(modelId, 0);
+    const promise = ensureLocalSttModel({
+      modelsDir,
+      modelId,
+      onProgress: (downloadedBytes, totalBytes) => {
+        downloadProgress.set(
+          modelId,
+          totalBytes ? Math.min(100, Math.round((downloadedBytes / totalBytes) * 100)) : null,
+        );
+      },
+    })
+      .then(() => {
+        downloadStates.delete(modelId);
+        downloadPromises.delete(modelId);
+        downloadProgress.delete(modelId);
+      })
+      .catch((error) => {
+        downloadStates.set(modelId, 'error');
+        downloadErrors.set(modelId, error?.message || String(error));
+        downloadPromises.delete(modelId);
+        downloadProgress.delete(modelId);
+      });
+    downloadPromises.set(modelId, promise);
+    return promise;
+  };
+
+  const resolveLocalModelId = (requested) => {
+    return isLocalSttModelId(requested) ? requested : DEFAULT_LOCAL_STT_MODEL;
+  };
+
+  /**
+   * Create a connected StreamingTranscriptionSession for one dictation.
+   * Returns { session } on success or { error, retryable, reasonCode } when
+   * the provider is not ready.
+   *
+   * @param {{ provider?: string, language?: string, localModel?: string,
+   *           openaiCompatible?: { baseUrl?: string, model?: string, apiKey?: string } }} options
+   */
+  const createSttSession = async (options = {}) => {
+    const provider = options.provider === 'openai-compatible' ? 'openai-compatible' : 'local';
+
+    if (provider === 'openai-compatible') {
+      const config = options.openaiCompatible || {};
+      const session = new OpenAICompatibleTranscriptionSession({
+        baseURL: config.baseUrl,
+        model: config.model,
+        apiKey: config.apiKey || undefined,
+        language: options.language || undefined,
+      });
+      try {
+        await session.connect();
+      } catch (error) {
+        return {
+          error: error?.message || String(error),
+          retryable: false,
+          reasonCode: 'stt_not_configured',
+        };
+      }
+      return { session };
+    }
+
+    const modelId = resolveLocalModelId(options.localModel);
+    const installed = await isLocalSttModelInstalled(modelsDir, modelId);
+    if (!installed) {
+      const state = downloadStates.get(modelId);
+      if (state === 'error') {
+        const message = downloadErrors.get(modelId) || 'Model download failed';
+        // Allow a retry on the next attempt.
+        downloadStates.delete(modelId);
+        return {
+          error: `Failed to download dictation model: ${message}`,
+          retryable: true,
+          reasonCode: 'model_download_failed',
+        };
+      }
+      void startModelDownload(modelId);
+      return {
+        error: 'Dictation model is downloading',
+        retryable: true,
+        reasonCode: 'model_download_in_progress',
+      };
+    }
+
+    const session = new WorkerBackedTranscriptionSession(workerClient, { modelsDir, modelId });
+    try {
+      await session.connect();
+    } catch (error) {
+      const message = error?.message || String(error);
+      // A model that passes the file-presence check but fails to load is
+      // corrupt on disk (e.g. truncated by an interrupted extraction). Remove
+      // it so the next attempt re-downloads instead of crashing forever.
+      if (/Load model|Protobuf parsing failed/i.test(message)) {
+        await rm(getLocalSttModelDir(modelsDir, modelId), { recursive: true, force: true })
+          .catch(() => undefined);
+        return {
+          error: 'Dictation model files were corrupt and have been removed; retry to re-download',
+          retryable: true,
+          reasonCode: 'model_corrupt',
+        };
+      }
+      return {
+        error: message,
+        retryable: true,
+        reasonCode: 'stt_unavailable',
+      };
+    }
+    return { session };
+  };
+
+  /**
+   * Readiness snapshot for the status route and UI gating.
+   * @param {{ provider?: string, localModel?: string }} [options]
+   */
+  const getStatus = async (options = {}) => {
+    const provider = options.provider === 'openai-compatible' ? 'openai-compatible' : 'local';
+    const modelId = resolveLocalModelId(options.localModel);
+
+    const describeModel = async (id, catalog) => ({
+      id,
+      description: catalog[id].description,
+      installed: await isLocalSttModelInstalled(modelsDir, id),
+      downloading: downloadStates.get(id) === 'downloading',
+      downloadProgress: downloadProgress.get(id) ?? null,
+      downloadError: downloadErrors.get(id) || null,
+    });
+
+    const models = await Promise.all(
+      LOCAL_STT_MODEL_IDS.map((id) => describeModel(id, LOCAL_STT_MODEL_CATALOG)),
+    );
+    const ttsModels = await Promise.all(
+      LOCAL_TTS_MODEL_IDS.map((id) => describeModel(id, LOCAL_TTS_MODEL_CATALOG)),
+    );
+
+    if (provider === 'openai-compatible') {
+      return { provider, available: true, models, ttsModels };
+    }
+
+    const model = models.find((entry) => entry.id === modelId) || null;
+    if (model?.installed) {
+      return { provider, available: true, activeModel: modelId, models, ttsModels };
+    }
+    if (model?.downloading) {
+      return {
+        provider,
+        available: false,
+        reasonCode: 'model_download_in_progress',
+        activeModel: modelId,
+        models,
+        ttsModels,
+      };
+    }
+    if (model?.downloadError) {
+      return {
+        provider,
+        available: false,
+        reasonCode: 'model_download_failed',
+        error: model.downloadError,
+        activeModel: modelId,
+        models,
+        ttsModels,
+      };
+    }
+    return {
+      provider,
+      available: false,
+      reasonCode: 'models_missing',
+      activeModel: modelId,
+      models,
+      ttsModels,
+    };
+  };
+
+  /**
+   * Synthesize speech with the local TTS model. Returns WAV bytes, or a
+   * readiness error while the model is missing/downloading.
+   * @param {{ text: string, model?: string, speakerId?: number, speed?: number }} options
+   */
+  const synthesizeSpeech = async ({ text, model, speakerId, speed }) => {
+    const modelId = isLocalTtsModelId(model) ? model : DEFAULT_LOCAL_TTS_MODEL;
+    const installed = await isLocalSttModelInstalled(modelsDir, modelId);
+    if (!installed) {
+      const state = downloadStates.get(modelId);
+      if (state === 'error') {
+        const message = downloadErrors.get(modelId) || 'Model download failed';
+        downloadStates.delete(modelId);
+        return {
+          error: `Failed to download TTS model: ${message}`,
+          retryable: true,
+          reasonCode: 'model_download_failed',
+        };
+      }
+      void startModelDownload(modelId);
+      return {
+        error: 'TTS model is downloading',
+        retryable: true,
+        reasonCode: 'model_download_in_progress',
+      };
+    }
+
+    const result = await workerClient.synthesizeSpeech({
+      modelsDir,
+      modelId,
+      text,
+      speakerId,
+      speed,
+    });
+    return { audio: result.audio, format: result.format };
+  };
+
+  /**
+   * Kick off a background download for a model (used by the status route's
+   * download action so Settings can pre-download models).
+   */
+  const requestModelDownload = async (modelId) => {
+    if (!isLocalModelId(modelId)) {
+      return { ok: false, error: 'Unknown model id' };
+    }
+    if (await isLocalSttModelInstalled(modelsDir, modelId)) {
+      return { ok: true, installed: true };
+    }
+    void startModelDownload(modelId);
+    return { ok: true, installed: false };
+  };
+
+  /**
+   * Delete an installed model from disk. A model that is mid-download cannot
+   * be deleted. An engine already loaded in the worker keeps its in-memory
+   * copy until the worker's idle shutdown; the files are simply re-downloaded
+   * on the next use if the model is selected again.
+   */
+  const deleteModel = async (modelId) => {
+    if (!isLocalModelId(modelId)) {
+      return { ok: false, error: 'Unknown model id' };
+    }
+    if (downloadStates.get(modelId) === 'downloading') {
+      return { ok: false, error: 'Model is downloading' };
+    }
+    await rm(getLocalSttModelDir(modelsDir, modelId), { recursive: true, force: true });
+    downloadErrors.delete(modelId);
+    return { ok: true };
+  };
+
+  const shutdown = () => {
+    workerClient.shutdown();
+  };
+
+  return {
+    createSttSession,
+    synthesizeSpeech,
+    getStatus,
+    requestModelDownload,
+    deleteModel,
+    shutdown,
+  };
+}

packages/web/server/lib/dictation/stream-manager.js

@@ -0,0 +1,461 @@
+/**
+ * DictationStreamManager
+ *
+ * Server-authoritative streaming dictation state machine. One manager owns
+ * all dictation streams for a single WebSocket connection.
+ *
+ * Responsibilities:
+ * - Reorders inbound chunks by `seq` and acks the highest contiguous seq.
+ * - Resamples client PCM (16 kHz by default) to the provider's required rate.
+ * - Auto-commits a segment every `autoCommitSeconds` of audio, but clears
+ *   silence-only segments instead of committing them.
+ * - Concatenates per-segment transcripts into live partials and emits the
+ *   final text once every committed segment has a final transcript.
+ * - Applies an adaptive finalization timeout budget based on pending work.
+ */
+
+import { Pcm16MonoResampler, parsePcmRateFromFormat, pcm16lePeakAbs } from './audio.js';
+
+const DEFAULT_FINAL_TIMEOUT_MS = 10000;
+const DEFAULT_AUTO_COMMIT_SECONDS = 15;
+const FINAL_TIMEOUT_MAX_MS = 5 * 60 * 1000;
+const FINAL_TIMEOUT_PER_PENDING_SEGMENT_MS = 15 * 1000;
+const FINAL_TIMEOUT_PER_PENDING_AUDIO_SECOND_MS = 1500;
+const FINAL_TIMEOUT_PER_MISSING_SEQ_MS = 250;
+const SILENCE_PEAK_THRESHOLD = 300;
+
+export class DictationStreamManager {
+  /**
+   * @param {object} params
+   * @param {(msg: { type: string, payload: object }) => void} params.emit
+   * @param {(startOptions: object) => Promise<{ session: object } | { error: string, retryable: boolean, reasonCode?: string }>} params.createSttSession
+   *   Resolves a connected streaming transcription session for one dictation.
+   *   The streaming transcription session contract:
+   *   { requiredSampleRate, appendPcm16(buf), commit(), clear(), close(), on(event, handler) }
+   * @param {number} [params.finalTimeoutMs]
+   * @param {number} [params.autoCommitSeconds]
+   */
+  constructor({ emit, createSttSession, finalTimeoutMs, autoCommitSeconds }) {
+    this.emit = emit;
+    this.createSttSession = createSttSession;
+    this.finalTimeoutMs = finalTimeoutMs ?? DEFAULT_FINAL_TIMEOUT_MS;
+    this.autoCommitSeconds = autoCommitSeconds ?? DEFAULT_AUTO_COMMIT_SECONDS;
+    this.streams = new Map();
+  }
+
+  cleanupAll() {
+    for (const dictationId of Array.from(this.streams.keys())) {
+      this.cleanupStream(dictationId);
+    }
+  }
+
+  /**
+   * @param {string} dictationId
+   * @param {string} format e.g. "audio/pcm;rate=16000;bits=16"
+   * @param {object} startOptions provider/config options forwarded to createSttSession
+   */
+  async handleStart(dictationId, format, startOptions = {}) {
+    this.cleanupStream(dictationId);
+
+    const inputRate = parsePcmRateFromFormat(format, 16000) ?? 16000;
+    if (!Number.isFinite(inputRate) || inputRate <= 0) {
+      this.failStream(dictationId, `Invalid dictation input rate in format: ${format}`, false);
+      return;
+    }
+
+    let resolved;
+    try {
+      resolved = await this.createSttSession(startOptions);
+    } catch (error) {
+      this.failStream(dictationId, error?.message || String(error), true);
+      return;
+    }
+    if (!resolved || resolved.error) {
+      this.failStream(
+        dictationId,
+        resolved?.error || 'Dictation STT not configured',
+        Boolean(resolved?.retryable),
+        resolved?.reasonCode,
+      );
+      return;
+    }
+
+    const stt = resolved.session;
+
+    stt.on('committed', ({ segmentId }) => {
+      const state = this.streams.get(dictationId);
+      if (!state) {
+        return;
+      }
+      state.committedSegmentIds.push(segmentId);
+      state.bytesSinceCommit = 0;
+      state.peakSinceCommit = 0;
+
+      if (state.finishRequested && state.awaitingFinalCommit) {
+        state.awaitingFinalCommit = false;
+      }
+
+      this.maybeFinalizeStream(dictationId);
+    });
+
+    stt.on('transcript', ({ segmentId, transcript, isFinal }) => {
+      const state = this.streams.get(dictationId);
+      if (!state) {
+        return;
+      }
+      state.transcriptsBySegmentId.set(segmentId, transcript);
+      if (isFinal) {
+        state.finalTranscriptSegmentIds.add(segmentId);
+      }
+
+      if (state.finishRequested && state.awaitingFinalCommit && isFinal) {
+        state.awaitingFinalCommit = false;
+      }
+
+      const orderedIds = state.committedSegmentIds.includes(segmentId)
+        ? state.committedSegmentIds
+        : [...state.committedSegmentIds, segmentId];
+      const partialText = orderedIds
+        .map((id) => state.transcriptsBySegmentId.get(id) ?? '')
+        .join(' ')
+        .trim();
+      this.emit({ type: 'partial', payload: { dictationId, text: partialText } });
+
+      this.maybeSealStreamFinish(dictationId);
+      this.maybeFinalizeStream(dictationId);
+    });
+
+    stt.on('error', (err) => {
+      const message = err?.message || String(err);
+      this.failAndCleanupStream(dictationId, message, true);
+    });
+
+    this.streams.set(dictationId, {
+      dictationId,
+      inputFormat: format,
+      stt,
+      inputRate,
+      outputRate: stt.requiredSampleRate,
+      resampler:
+        inputRate === stt.requiredSampleRate
+          ? null
+          : new Pcm16MonoResampler({ inputRate, outputRate: stt.requiredSampleRate }),
+      receivedChunks: new Map(),
+      nextSeqToForward: 0,
+      ackSeq: -1,
+      autoCommitBytes:
+        this.autoCommitSeconds > 0
+          ? Math.max(1, Math.round(this.autoCommitSeconds * stt.requiredSampleRate * 2))
+          : 0,
+      bytesSinceCommit: 0,
+      peakSinceCommit: 0,
+      committedSegmentIds: [],
+      transcriptsBySegmentId: new Map(),
+      finalTranscriptSegmentIds: new Set(),
+      awaitingFinalCommit: false,
+      finishRequested: false,
+      finishSealed: false,
+      finalSeq: null,
+      finalTimeout: null,
+    });
+
+    this.emitAck(dictationId, -1);
+  }
+
+  /**
+   * @param {{ dictationId: string, seq: number, audioBase64: string }} params
+   */
+  handleChunk({ dictationId, seq, audioBase64 }) {
+    const state = this.streams.get(dictationId);
+    if (!state) {
+      this.failStream(dictationId, 'Dictation stream not started', true);
+      return;
+    }
+
+    if (!Number.isInteger(seq) || seq < 0) {
+      return;
+    }
+
+    if (seq < state.nextSeqToForward) {
+      this.emitAck(dictationId, state.ackSeq);
+      return;
+    }
+
+    if (!state.receivedChunks.has(seq)) {
+      let chunk;
+      try {
+        chunk = Buffer.from(audioBase64, 'base64');
+      } catch {
+        return;
+      }
+      if (chunk.length % 2 !== 0) {
+        chunk = chunk.subarray(0, chunk.length - 1);
+      }
+      state.receivedChunks.set(seq, chunk);
+    }
+
+    while (state.receivedChunks.has(state.nextSeqToForward)) {
+      const nextSeq = state.nextSeqToForward;
+      const pcm16 = state.receivedChunks.get(nextSeq);
+      state.receivedChunks.delete(nextSeq);
+
+      const resampled = state.resampler ? state.resampler.processChunk(pcm16) : pcm16;
+      if (resampled.length > 0) {
+        state.stt.appendPcm16(resampled);
+        state.bytesSinceCommit += resampled.length;
+        state.peakSinceCommit = Math.max(state.peakSinceCommit, pcm16lePeakAbs(resampled));
+        try {
+          this.maybeAutoCommitSegment(state);
+        } catch (error) {
+          this.failAndCleanupStream(dictationId, error?.message || String(error), true);
+          return;
+        }
+      }
+
+      state.nextSeqToForward += 1;
+      state.ackSeq = state.nextSeqToForward - 1;
+    }
+
+    this.emitAck(dictationId, state.ackSeq);
+    this.maybeSealStreamFinish(dictationId);
+    this.maybeFinalizeStream(dictationId);
+  }
+
+  /**
+   * @param {string} dictationId
+   * @param {number} finalSeq highest seq the client sent (or -1 if none)
+   */
+  handleFinish(dictationId, finalSeq) {
+    const state = this.streams.get(dictationId);
+    if (!state) {
+      this.failStream(dictationId, 'Dictation stream not started', true);
+      return;
+    }
+
+    state.finishRequested = true;
+    state.finalSeq = finalSeq;
+
+    if (
+      finalSeq >= 0 &&
+      state.ackSeq < 0 &&
+      state.nextSeqToForward === 0 &&
+      state.receivedChunks.size === 0
+    ) {
+      this.failStream(
+        dictationId,
+        'Dictation finished but no audio chunks were received',
+        true,
+      );
+      this.cleanupStream(dictationId);
+      return;
+    }
+
+    this.maybeSealStreamFinish(dictationId);
+    this.maybeFinalizeStream(dictationId);
+
+    const updatedState = this.streams.get(dictationId);
+    if (!updatedState) {
+      return;
+    }
+
+    const timeoutMs = this.estimateFinalizationTimeout(updatedState);
+    if (updatedState.finalTimeout) {
+      clearTimeout(updatedState.finalTimeout);
+    }
+    updatedState.finalTimeout = setTimeout(() => {
+      this.failAndCleanupStream(dictationId, 'Timed out waiting for final transcription', true);
+    }, timeoutMs);
+
+    this.emit({ type: 'finish_accepted', payload: { dictationId, timeoutMs } });
+  }
+
+  handleCancel(dictationId) {
+    this.cleanupStream(dictationId);
+  }
+
+  emitAck(dictationId, ackSeq) {
+    this.emit({ type: 'ack', payload: { dictationId, ackSeq } });
+  }
+
+  failStream(dictationId, error, retryable, reasonCode) {
+    this.emit({
+      type: 'error',
+      payload: {
+        dictationId,
+        error,
+        retryable,
+        ...(reasonCode ? { reasonCode } : {}),
+      },
+    });
+  }
+
+  failAndCleanupStream(dictationId, error, retryable) {
+    this.failStream(dictationId, error, retryable);
+    this.cleanupStream(dictationId);
+  }
+
+  cleanupStream(dictationId) {
+    const state = this.streams.get(dictationId);
+    if (!state) {
+      return;
+    }
+    if (state.finalTimeout) {
+      clearTimeout(state.finalTimeout);
+    }
+    try {
+      state.stt.close();
+    } catch {
+      // no-op
+    }
+    this.streams.delete(dictationId);
+  }
+
+  estimateFinalizationTimeout(state) {
+    const bytesPerSecond = Math.max(1, state.outputRate * 2);
+    const pendingCommittedSegments = state.committedSegmentIds.reduce((count, segmentId) => {
+      return state.finalTranscriptSegmentIds.has(segmentId) ? count : count + 1;
+    }, 0);
+    const committedSet = new Set(state.committedSegmentIds);
+    const pendingUncommittedTranscriptSegments = Array.from(
+      state.transcriptsBySegmentId.keys(),
+    ).reduce((count, segmentId) => {
+      if (committedSet.has(segmentId)) {
+        return count;
+      }
+      return state.finalTranscriptSegmentIds.has(segmentId) ? count : count + 1;
+    }, 0);
+    const pendingSegments =
+      pendingCommittedSegments +
+      pendingUncommittedTranscriptSegments +
+      (state.awaitingFinalCommit ? 1 : 0);
+    const pendingAudioSeconds = Math.ceil(Math.max(0, state.bytesSinceCommit) / bytesPerSecond);
+    const missingSeqCount =
+      state.finalSeq === null ? 0 : Math.max(0, state.finalSeq - state.ackSeq);
+
+    const extraMs =
+      pendingSegments * FINAL_TIMEOUT_PER_PENDING_SEGMENT_MS +
+      pendingAudioSeconds * FINAL_TIMEOUT_PER_PENDING_AUDIO_SECOND_MS +
+      missingSeqCount * FINAL_TIMEOUT_PER_MISSING_SEQ_MS;
+
+    return Math.max(
+      this.finalTimeoutMs,
+      Math.min(FINAL_TIMEOUT_MAX_MS, this.finalTimeoutMs + extraMs),
+    );
+  }
+
+  maybeAutoCommitSegment(state) {
+    if (state.finishRequested) {
+      return;
+    }
+    if (state.autoCommitBytes <= 0 || state.bytesSinceCommit < state.autoCommitBytes) {
+      return;
+    }
+    if (state.peakSinceCommit < SILENCE_PEAK_THRESHOLD) {
+      state.stt.clear();
+      state.bytesSinceCommit = 0;
+      state.peakSinceCommit = 0;
+      return;
+    }
+
+    state.bytesSinceCommit = 0;
+    state.peakSinceCommit = 0;
+    state.stt.commit();
+  }
+
+  maybeSealStreamFinish(dictationId) {
+    const state = this.streams.get(dictationId);
+    if (!state) {
+      return;
+    }
+    if (!state.finishRequested || state.finalSeq === null) {
+      return;
+    }
+    if (state.ackSeq < state.finalSeq) {
+      return;
+    }
+    if (state.finishSealed) {
+      return;
+    }
+
+    if (state.bytesSinceCommit > 0) {
+      if (state.peakSinceCommit < SILENCE_PEAK_THRESHOLD) {
+        state.stt.clear();
+        state.bytesSinceCommit = 0;
+        state.peakSinceCommit = 0;
+        state.awaitingFinalCommit = false;
+        this.dropUncommittedNonFinalTranscripts(state);
+      } else {
+        state.awaitingFinalCommit = true;
+        try {
+          state.stt.commit();
+        } catch (error) {
+          this.failAndCleanupStream(dictationId, error?.message || String(error), true);
+          return;
+        }
+      }
+    } else {
+      state.awaitingFinalCommit = false;
+    }
+
+    state.finishSealed = true;
+  }
+
+  dropUncommittedNonFinalTranscripts(state) {
+    const committedSet = new Set(state.committedSegmentIds);
+    for (const segmentId of Array.from(state.transcriptsBySegmentId.keys())) {
+      if (committedSet.has(segmentId)) {
+        continue;
+      }
+      if (state.finalTranscriptSegmentIds.has(segmentId)) {
+        continue;
+      }
+      state.transcriptsBySegmentId.delete(segmentId);
+    }
+  }
+
+  maybeFinalizeStream(dictationId) {
+    const state = this.streams.get(dictationId);
+    if (!state) {
+      return;
+    }
+
+    if (!state.finishRequested || state.finalSeq === null) {
+      return;
+    }
+    if (state.ackSeq < state.finalSeq) {
+      return;
+    }
+    if (state.awaitingFinalCommit) {
+      return;
+    }
+
+    const committedSet = new Set(state.committedSegmentIds);
+    const orderedSegmentIds = [...state.committedSegmentIds];
+    for (const segmentId of state.transcriptsBySegmentId.keys()) {
+      if (!committedSet.has(segmentId)) {
+        orderedSegmentIds.push(segmentId);
+      }
+    }
+
+    if (orderedSegmentIds.length === 0) {
+      this.emit({ type: 'final', payload: { dictationId, text: '' } });
+      this.cleanupStream(dictationId);
+      return;
+    }
+
+    const allTranscriptsReady = orderedSegmentIds.every((segmentId) =>
+      state.finalTranscriptSegmentIds.has(segmentId),
+    );
+    if (!allTranscriptsReady) {
+      return;
+    }
+
+    const orderedText = orderedSegmentIds
+      .map((segmentId) => state.transcriptsBySegmentId.get(segmentId) ?? '')
+      .join(' ')
+      .trim();
+
+    this.emit({ type: 'final', payload: { dictationId, text: orderedText } });
+    this.cleanupStream(dictationId);
+  }
+}

packages/web/server/lib/dictation/stream-manager.test.js

@@ -0,0 +1,194 @@
+import { describe, it, expect } from 'bun:test';
+import { EventEmitter } from 'events';
+
+import { DictationStreamManager } from './stream-manager.js';
+
+const FORMAT = 'audio/pcm;rate=16000;bits=16';
+
+class FakeSttSession extends EventEmitter {
+  constructor({ transcriptBySegment = () => 'hello world' } = {}) {
+    super();
+    this.requiredSampleRate = 16000;
+    this.appended = [];
+    this.commits = 0;
+    this.clears = 0;
+    this.closed = false;
+    this.segmentCounter = 0;
+    this.transcriptBySegment = transcriptBySegment;
+  }
+
+  async connect() {}
+
+  appendPcm16(buf) {
+    this.appended.push(buf);
+  }
+
+  commit() {
+    this.commits += 1;
+    const segmentId = `seg-${this.segmentCounter}`;
+    this.segmentCounter += 1;
+    this.emit('committed', { segmentId, previousSegmentId: null });
+    setTimeout(() => {
+      this.emit('transcript', {
+        segmentId,
+        transcript: this.transcriptBySegment(segmentId),
+        isFinal: true,
+      });
+    }, 0);
+  }
+
+  clear() {
+    this.clears += 1;
+  }
+
+  close() {
+    this.closed = true;
+  }
+}
+
+function loudChunkBase64(samples = 1600, amplitude = 8000) {
+  const arr = new Int16Array(samples);
+  for (let i = 0; i < samples; i += 1) {
+    arr[i] = i % 2 === 0 ? amplitude : -amplitude;
+  }
+  return Buffer.from(arr.buffer).toString('base64');
+}
+
+function silentChunkBase64(samples = 1600) {
+  return Buffer.from(new Int16Array(samples).buffer).toString('base64');
+}
+
+function createManager(session) {
+  const messages = [];
+  const manager = new DictationStreamManager({
+    emit: (msg) => messages.push(msg),
+    createSttSession: async () => ({ session }),
+  });
+  return { manager, messages };
+}
+
+function waitFor(predicate, timeoutMs = 1000) {
+  return new Promise((resolve, reject) => {
+    const startedAt = Date.now();
+    const tick = () => {
+      if (predicate()) {
+        resolve(undefined);
+        return;
+      }
+      if (Date.now() - startedAt > timeoutMs) {
+        reject(new Error('waitFor timed out'));
+        return;
+      }
+      setTimeout(tick, 5);
+    };
+    tick();
+  });
+}
+
+describe('DictationStreamManager', () => {
+  it('transcribes ordered chunks and emits final text', async () => {
+    const session = new FakeSttSession();
+    const { manager, messages } = createManager(session);
+
+    await manager.handleStart('d1', FORMAT, {});
+    manager.handleChunk({ dictationId: 'd1', seq: 0, audioBase64: loudChunkBase64() });
+    manager.handleChunk({ dictationId: 'd1', seq: 1, audioBase64: loudChunkBase64() });
+    manager.handleFinish('d1', 1);
+
+    await waitFor(() => messages.some((m) => m.type === 'final'));
+
+    const final = messages.find((m) => m.type === 'final');
+    expect(final.payload.text).toBe('hello world');
+    expect(session.commits).toBe(1);
+    expect(session.closed).toBe(true);
+
+    const acks = messages.filter((m) => m.type === 'ack');
+    expect(acks[acks.length - 1].payload.ackSeq).toBe(1);
+  });
+
+  it('reorders out-of-order chunks before appending', async () => {
+    const session = new FakeSttSession();
+    const { manager, messages } = createManager(session);
+
+    await manager.handleStart('d1', FORMAT, {});
+    manager.handleChunk({ dictationId: 'd1', seq: 1, audioBase64: loudChunkBase64() });
+    expect(session.appended.length).toBe(0);
+    manager.handleChunk({ dictationId: 'd1', seq: 0, audioBase64: loudChunkBase64() });
+    expect(session.appended.length).toBe(2);
+    manager.handleFinish('d1', 1);
+
+    await waitFor(() => messages.some((m) => m.type === 'final'));
+  });
+
+  it('clears silence-only tails instead of committing', async () => {
+    const session = new FakeSttSession();
+    const { manager, messages } = createManager(session);
+
+    await manager.handleStart('d1', FORMAT, {});
+    manager.handleChunk({ dictationId: 'd1', seq: 0, audioBase64: silentChunkBase64() });
+    manager.handleFinish('d1', 0);
+
+    await waitFor(() => messages.some((m) => m.type === 'final'));
+
+    const final = messages.find((m) => m.type === 'final');
+    expect(final.payload.text).toBe('');
+    expect(session.commits).toBe(0);
+    expect(session.clears).toBe(1);
+  });
+
+  it('fails fast when finish arrives with no chunks', async () => {
+    const session = new FakeSttSession();
+    const { manager, messages } = createManager(session);
+
+    await manager.handleStart('d1', FORMAT, {});
+    manager.handleFinish('d1', 3);
+
+    const error = messages.find((m) => m.type === 'error');
+    expect(error).toBeDefined();
+    expect(error.payload.retryable).toBe(true);
+    expect(session.closed).toBe(true);
+  });
+
+  it('reports provider readiness errors from createSttSession', async () => {
+    const messages = [];
+    const manager = new DictationStreamManager({
+      emit: (msg) => messages.push(msg),
+      createSttSession: async () => ({
+        error: 'Dictation model is downloading',
+        retryable: true,
+        reasonCode: 'model_download_in_progress',
+      }),
+    });
+
+    await manager.handleStart('d1', FORMAT, {});
+    const error = messages.find((m) => m.type === 'error');
+    expect(error.payload.reasonCode).toBe('model_download_in_progress');
+    expect(error.payload.retryable).toBe(true);
+  });
+
+  it('emits partials as segment transcripts arrive', async () => {
+    let segment = 0;
+    const session = new FakeSttSession({
+      transcriptBySegment: () => {
+        segment += 1;
+        return segment === 1 ? 'first part' : 'second part';
+      },
+    });
+    const { manager, messages } = createManager(session);
+    // Force auto-commit after ~0.05s of audio so two segments form.
+    manager.autoCommitSeconds = 0.05;
+
+    await manager.handleStart('d1', FORMAT, {});
+    manager.handleChunk({ dictationId: 'd1', seq: 0, audioBase64: loudChunkBase64(1600) });
+    await waitFor(() => session.commits >= 1);
+    manager.handleChunk({ dictationId: 'd1', seq: 1, audioBase64: loudChunkBase64(1600) });
+    manager.handleFinish('d1', 1);
+
+    await waitFor(() => messages.some((m) => m.type === 'final'));
+
+    const final = messages.find((m) => m.type === 'final');
+    expect(final.payload.text).toBe('first part second part');
+    const partials = messages.filter((m) => m.type === 'partial');
+    expect(partials.length).toBeGreaterThan(0);
+  });
+});

packages/web/server/lib/opencode/settings-helpers.js

@@ -721,10 +721,18 @@ export const createSettingsHelpers = (dependencies) => {
       }
     }
 
+    if (typeof candidate.dictationEnabled === 'boolean') {
+      result.dictationEnabled = candidate.dictationEnabled;
+    }
     if (typeof candidate.sttProvider === 'string') {
       const provider = candidate.sttProvider.trim();
-      if (provider === 'browser' || provider === 'server' || provider === 'wasm') {
+      if (provider === 'local' || provider === 'openai-compatible') {
         result.sttProvider = provider;
+      } else if (provider === 'server') {
+        // Legacy provider migration: 'server' was the OpenAI-compatible endpoint.
+        result.sttProvider = 'openai-compatible';
+      } else if (provider === 'browser' || provider === 'wasm') {
+        result.sttProvider = 'local';
       }
     }
     if (typeof candidate.sttServerUrl === 'string') {
@@ -739,10 +747,10 @@ export const createSettingsHelpers = (dependencies) => {
         result.sttModel = trimmed;
       }
     }
-    if (typeof candidate.wasmSttModel === 'string') {
+    if (typeof candidate.sttLocalModel === 'string') {
-      const trimmed = candidate.wasmSttModel.trim();
+      const trimmed = candidate.sttLocalModel.trim();
-      if (trimmed.length <= 256) {
+      if (trimmed.length <= STT_MODEL_MAX_LENGTH) {
-        result.wasmSttModel = trimmed;
+        result.sttLocalModel = trimmed;
       }
     }
     if (typeof candidate.sttLanguage === 'string') {
@@ -751,15 +759,6 @@ export const createSettingsHelpers = (dependencies) => {
         result.sttLanguage = trimmed;
       }
     }
-    if (typeof candidate.sttSilenceThresholdDb === 'number' && Number.isFinite(candidate.sttSilenceThresholdDb)) {
-      result.sttSilenceThresholdDb = Math.max(-100, Math.min(0, candidate.sttSilenceThresholdDb));
-    }
-    if (typeof candidate.sttSilenceHoldMs === 'number' && Number.isFinite(candidate.sttSilenceHoldMs)) {
-      result.sttSilenceHoldMs = Math.max(250, Math.min(10000, Math.round(candidate.sttSilenceHoldMs)));
-    }
-    if (typeof candidate.sttTranscribeOnStop === 'boolean') {
-      result.sttTranscribeOnStop = candidate.sttTranscribeOnStop;
-    }
 
     return result;
   };

packages/web/server/lib/opencode/startup-pipeline-runtime.js

@@ -1,6 +1,7 @@
 export const createStartupPipelineRuntime = (dependencies) => {
   const {
     createTerminalRuntime,
+    createDictationRuntime,
     createMessageStreamWsRuntime,
     createServerStartupRuntime,
   } = dependencies;
@@ -52,6 +53,7 @@ export const createStartupPipelineRuntime = (dependencies) => {
       tunnelRuntimeContext,
       attachSignals,
       apiOnly,
+      dictationModelsDir,
     } = options;
 
     const terminalRuntime = createTerminalRuntime({
@@ -71,6 +73,16 @@ export const createStartupPipelineRuntime = (dependencies) => {
       TERMINAL_INPUT_WS_MAX_REBINDS_PER_WINDOW: terminalMaxRebindsPerWindow,
     });
 
+    const dictationRuntime = createDictationRuntime({
+      app,
+      server,
+      express,
+      uiAuthController,
+      isRequestOriginAllowed,
+      rejectWebSocketUpgrade,
+      modelsDir: dictationModelsDir,
+    });
+
     const messageStreamRuntime = createMessageStreamWsRuntime({
       server,
       uiAuthController,
@@ -125,6 +137,7 @@ export const createStartupPipelineRuntime = (dependencies) => {
 
     return {
       terminalRuntime,
+      dictationRuntime,
       messageStreamRuntime,
     };
   };

packages/web/server/lib/security/request-security.js

@@ -1,6 +1,16 @@
 export const createRequestSecurityRuntime = (deps) => {
   const { readSettingsFromDiskMigrated } = deps;
-  const packagedClientOrigins = new Set(['openchamber-ui://app', 'capacitor://localhost']);
+  // Origins of packaged (non-browser) clients whose WebView origin never
+  // matches the server host: the desktop shell, the iOS Capacitor WebView
+  // (capacitor://localhost), and the Android Capacitor WebView, which uses
+  // androidScheme 'https' and therefore reports 'https://localhost'. Missing
+  // the Android origin 403'd every WebSocket upgrade from the Android app
+  // (message stream, terminal, dictation) while SSE kept working.
+  const packagedClientOrigins = new Set([
+    'openchamber-ui://app',
+    'capacitor://localhost',
+    'https://localhost',
+  ]);
 
   const getUiSessionTokenFromRequest = (req) => {
     const cookieHeader = req?.headers?.cookie;

packages/web/server/lib/security/request-security.test.js

@@ -24,5 +24,26 @@ describe('request security runtime', () => {
       },
       socket: {},
     })).resolves.toBe(true);
+
+    // Android Capacitor WebView (androidScheme 'https') reports this origin.
+    await expect(runtime.isRequestOriginAllowed({
+      headers: {
+        origin: 'https://localhost',
+        host: '192.168.1.130:1202',
+      },
+      socket: {},
+    })).resolves.toBe(true);
+  });
+
+  test('rejects unknown origins', async () => {
+    const runtime = createRuntime();
+
+    await expect(runtime.isRequestOriginAllowed({
+      headers: {
+        origin: 'https://evil.example.com',
+        host: '192.168.1.130:1202',
+      },
+      socket: {},
+    })).resolves.toBe(false);
   });
 });

packages/web/server/lib/ui-auth/ui-auth.js

@@ -310,6 +310,7 @@ const isUrlAuthWebSocketPath = (pathname) => {
     || pathname === '/api/global/event/ws'
     || pathname === '/api/openchamber/realtime-proxy/ws'
     || pathname === '/api/terminal/ws'
+    || pathname === '/api/dictation/ws'
     || pathname.startsWith('/api/preview/proxy/');
 };
 

packages/web/server/lib/ui-auth/ui-auth.test.js

@@ -213,6 +213,28 @@ describe('ui auth client credential seam', () => {
     });
     expect(mountedServeCalled).toBe(true);
 
+    const dictationWsReq = {
+      method: 'GET',
+      path: '/api/dictation/ws',
+      url: `/api/dictation/ws?oc_url_token=${encodeURIComponent(urlToken)}`,
+      headers: { upgrade: 'websocket' },
+    };
+    expect(await auth.ensureSessionToken(dictationWsReq, null)).toBe('client:device-1');
+
+    const dictationHttpReq = {
+      method: 'GET',
+      path: '/api/dictation/ws',
+      url: `/api/dictation/ws?oc_url_token=${encodeURIComponent(urlToken)}`,
+      headers: { accept: 'application/json' },
+    };
+    const dictationHttpRes = createResponse();
+    let dictationHttpCalled = false;
+    await auth.requireAuth(dictationHttpReq, dictationHttpRes, () => {
+      dictationHttpCalled = true;
+    });
+    expect(dictationHttpCalled).toBe(false);
+    expect(dictationHttpRes.statusCode).toBe(401);
+
     const arbitraryGetReq = { method: 'GET', path: '/api/config/settings', url: `/api/config/settings?oc_url_token=${encodeURIComponent(urlToken)}`, headers: { accept: 'application/json' } };
     const arbitraryGetRes = createResponse();
     let arbitraryGetCalled = false;