dsh-voice

⚙️ Configuration

The 🎤 mic button needs voice.stt.apiBase (the browser records audio and sends it to the host's Whisper-compatible backend). The 🔊 read-aloud needs nothing (browser speechSynthesis). To customize read-aloud language/rate/pitch, edit the constants at the top of lib/client.js (TTS_LANG, TTS_RATE, TTS_PITCH).

settings.yaml:

voice:
  stt:                        # mic button + voice_transcribe tool
    enabled: true
    apiBase: ""               # REQUIRED for the mic. Examples:
                              #   SiliconFlow: https://api.siliconflow.cn/v1
                              #   local whisper.cpp: http://127.0.0.1:8080/v1
    apiKeyEnv: VOICE_STT_API_KEY
    model: whisper-1
    language: ""              # zh / en / ... ; empty = auto-detect
  tts:                        # voice_speak tool
    enabled: true
    apiBase: ""               # empty = https://api.openai.com/v1
    apiKeyEnv: VOICE_TTS_API_KEY
    model: tts-1
    voice: alloy              # alloy/echo/fable/onyx/nova/shimmer, or a local voice id
    format: mp3

Why the mic needs a backend: Chrome's built-in SpeechRecognition uploads audio to Google, which is unreachable in some regions (you'd see 识别出错：network). dsh-voice records with MediaRecorder and transcribes through your Whisper-compatible backend instead. Two free, keyless options:

• SiliconFlow (China-friendly, free tier) — apiBase: https://api.siliconflow.cn/v1, model FunAudioLLM/SenseVoiceSmall or whisper-1.
• Local whisper.cpp — fully offline, apiBase: http://127.0.0.1:8080/v1 (no key).