BuzzbeeSCD/openchamber
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

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

Browse Source 
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
This commit is contained in:
Bohdan Triapitsyn
2026-08-30 02:24:22 +03:00
parent 49f0a9e62f
commit 391f938334
29 changed files with 949 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/web/server/lib/dictation/DOCUMENTATION.md
+19 -6
View File
@@ -11,12 +11,25 @@ live transcript costs O(n^2) work for a result the final decode replaces. The
composer shows no text while recording and inserts the full transcript on
stop.
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.
Local TTS (Kokoro and Piper/VITS via sherpa-onnx OfflineTts) runs in the same
worker process and is exposed as `POST /api/dictation/tts/speak` (JSON
`{text, speakerId?, speed?, model?, language?, languageSample?}` → 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.
Each TTS catalog entry declares the `languages` it speaks. With
`language: 'auto'` the service detects the language of `languageSample` — the
whole message the chunk belongs to, sent by the client with every chunk — or
of `text` when no sample is given
(`../tts/language-detect.js`, script plus function-word scoring, no
dependencies) and keeps the caller's model when it speaks that language;
otherwise it switches to the catalog model for the language, downloading it on
first use like any other model, and starts from that model's default speaker
(`defaultSpeakerByLanguage`) instead of the caller's speaker id. A language no
catalog model covers keeps the caller's model, so text is always spoken. The
response carries `X-Speech-Model` and `X-Speech-Language`.
## Ownership