feat(voice): match local and macOS voices to the language of the text

Text-to-speech picked one voice regardless of what language a reply was in.
A dependency-free language detector (script, marker letters, function words)
now decides the language of the whole message once; with the new
"Match the voice to the language of the text" setting the local provider
switches to a catalog model for that language (Kokoro zh/en and Piper models
for 12 languages, downloaded on first use like the existing model) and macOS
say switches to an installed voice whose locale matches. The local voice
picker lists voices of every installed model, and the settings show which
language models are on disk.

The Ukrainian Piper medium build is a character-level model that sherpa-onnx
turns into noise, so the espeak-based Lada build is used instead.

Claude-Session: https://claude.ai/code/session_017TK5JAYDfT3Fotc23UEg98

packages/ui/src/hooks/useLocalTTS.ts

@@ -14,10 +14,18 @@ import { useCallback, useEffect, useRef, useState } from 'react';
 import { runtimeFetch } from '@/lib/runtime-fetch';
 
 export interface LocalTTSSpeakOptions {
-    /** Kokoro speaker id (0-10) */
+    /** Catalog id of the local model to use; defaults to the server's default model. */
+    model?: string;
+    /** Speaker id within the model (Kokoro voices; Piper models have one) */
     speakerId?: number;
     /** Playback speed multiplier (1.0 = normal) */
     speed?: number;
+    /**
+     * `'auto'`: the server picks a model and voice for the text's language.
+     * The language is judged on the whole message, not on each chunk sent for
+     * synthesis, so a short chunk cannot flip the voice mid-reply.
+     */
+    language?: 'auto';
     onStart?: () => void;
     onEnd?: () => void;
     onError?: (error: string) => void;
@@ -35,6 +43,8 @@ export interface UseLocalTTSReturn {
 /** Target chunk size: big enough to amortize requests, small enough for low latency. */
 const MIN_CHUNK_CHARS = 60;
 const MAX_CHUNK_CHARS = 400;
+// Enough of the message for language detection to see whole sentences.
+const LANGUAGE_SAMPLE_CHARS = 2000;
 
 /**
  * Split text into sentence-aligned chunks for pipelined synthesis.
@@ -170,6 +180,7 @@ export function useLocalTTS(): UseLocalTTSReturn {
 
         const session: PlaybackSession = { cancelled: false, abort: new AbortController() };
         sessionRef.current = session;
+        const languageSample = options?.language === 'auto' ? text.slice(0, LANGUAGE_SAMPLE_CHARS) : undefined;
 
         const fetchChunk = async (chunk: string): Promise<ArrayBuffer> => {
             const response = await runtimeFetch('/api/dictation/tts/speak', {
@@ -177,8 +188,11 @@ export function useLocalTTS(): UseLocalTTSReturn {
                 headers: { 'Content-Type': 'application/json' },
                 body: JSON.stringify({
                     text: chunk,
+                    model: options?.model,
                     ...(typeof options?.speakerId === 'number' ? { speakerId: options.speakerId } : {}),
                     ...(typeof options?.speed === 'number' ? { speed: options.speed } : {}),
+                    language: options?.language,
+                    languageSample,
                 }),
                 signal: session.abort.signal,
             });