STARDUSTLC666/dsh-voice

中文

dsh-voice

DSH (DeepSeek Harness) voice plugin pair: let the agent speak and listen.

• voice_tts: text-to-speech over the edge-tts protocol (Microsoft Edge read-aloud service — free, unlimited, 22+ curated voices)
• voice_stt: speech-to-text over any OpenAI-compatible ASR endpoint (Groq / OpenAI / custom)
• voice_list: voice catalog

Installation

dsh plugin --profile web add dsh-voice

Configuration

voice_tts works with zero config; voice_stt needs an ASR key:

- id: voice
  name: 'dsh-voice'
  config:
    asrEngine: groq                       # groq | openai | custom
    asrModel: whisper-large-v3-turbo      # Groq whisper model
    # asrApiKey: gsk_...                  # prefer env var DSH_VOICE_ASR_KEY
    ttsVoice: zh-CN-XiaoxiaoNeural        # default voice
    # proxyUrl: http://127.0.0.1:7890     # enable when the ASR endpoint needs a special proxy

Tools

Tool	Purpose	Key parameters
voice_tts	Synthesize MP3 from text (free)	text required; voice/rate/pitch/output optional
voice_stt	Transcribe audio to text	audio required; engine/model/language/prompt/output optional
voice_list	Curated voice catalog	none

Examples

voice_tts { text: hello world }                              # outputs voice_output.mp3
voice_tts { text: hello, voice: en-US-AriaNeural }           # English female voice
voice_stt { audio: E:\audio\meeting.mp3, language: zh }     # transcribe a recording
voice_list {}

Under the hood

• Direct edge-tts protocol: the Sec-MS-GEC token is generated locally with the official DRM algorithm (SHA256 of Windows file time + trusted client token, 5-minute windows); transport uses the ws library with permessage-deflate and an optional HTTP CONNECT proxy tunnel
• Zero API cost: TTS is completely free; STT costs only whatever your ASR provider charges
• Up-front validation: text ≤ 5000 chars, audio ≤ 25MB; same-name outputs auto-suffixed
• Protocol aligned with current open-source edge-tts (7.x) — no reliance on the outdated token endpoint

Development

pnpm install
pnpm test       # build + 31 tests, including a real-synthesis integration test

License

MIT