maoyuching/dsh-voice-chat

A Doubao-style voice chat plugin for the DeepSeek Harness Web GUI: click 🎤 and talk — the AI's reply is read back to you aloud.

📖 Full user manual: MANUAL.md (install / usage / configuration / FAQ / internals).

Features

• Voice input: click 🎤 to start listening (ding) → speak → auto-stops after 2.5s of silence (dong) → transcribed and sent;
• Voice output: AI replies are synthesized with TTS and read aloud at +10% speed; a condensed reading option (off by default) can rewrite longer replies into a short spoken summary before speaking — the rewrite follows the LLM actually selected in the current conversation;
• ASR engines: supports SiliconFlow (SenseVoice, free tier), Groq (Whisper), Xiaomi MiMo (chat/completions protocol), and custom OpenAI-compatible endpoints — switchable in settings;
• TTS engines: supports Edge TTS (Microsoft, free), Xiaomi MiMo TTS (chat/completions protocol with built-in Chinese voices), and custom TTS (OpenAI-compatible API);
• Isolated engine configs: every ASR/TTS engine stores its own Base URL / model / API key / voice, so switching engines never overwrites another one (legacy single-slot configs migrate automatically);
• Single-channel playback: a new reply preempts the previous one; interrupt anytime via hotkey or button — only one voice at a time;
• No duplicate playback: which replies were already spoken is remembered per session, so re-entering a session does not replay them;
• Mute toggle 🔊: while playing, click to mute immediately; click again to resume;
• Settings UI: all settings live inside DSH's built-in settings dialog (gear icon → "voice chat" section), applied immediately — no restart needed;
• Hotkeys: Ctrl+Shift+Space toggles the mic (alternates Ctrl+M / Ctrl+Shift+M).

Requirements

• DeepSeek Harness (dsh): installed and running the Web GUI (dsh web);
• Node.js ≥ 22 (host and browser half; the edge-tts client runs on ws, no extra runtime needed);
• Browser: Chrome / Edge (recording needs MediaRecorder support);
• ASR key: required for speech-to-text (SiliconFlow offers a free tier; MiMo charges per usage).

Install

# Option 1 (recommended — published to npm):
dsh plugin --profile web add dsh-voice-chat

# Then restart dsh web

Note: the plugin ships an inline edge-tts client (Microsoft Edge's free read-aloud service) — no TTS API key is needed for Edge TTS; MiMo TTS / custom TTS require their own keys.

Configuration

⚙️ Settings panel (recommended, highest priority)

Open DSH's settings dialog (gear icon, bottom-left) → "voice chat" on the left → edit:

🎤 Voice Recognition Settings

• ASR engine (SiliconFlow / Groq / MiMo / custom)
• Base URL / model name / API key
• Auto-send after transcription
• Silence auto-stop duration (seconds)

🔊 Readout Settings

• TTS engine (Edge TTS / MiMo TTS / custom TTS)
• Per-engine Base URL / model name / API key / voice
• Condensed reading toggle (off by default)

🔒 Per-engine isolation: each of the 4 ASR engines and 3 TTS engines keeps its own Base URL / model / API key / voice. Switching engines only switches which slot you are looking at — they never overwrite each other, and saving only writes the slot being edited. Legacy (≤0.3.x) single-slot configs are migrated automatically on first start (MiMo voice → MiMo TTS, custom URL/key → custom TTS).

Saving writes to settings.local.json and takes effect immediately.

📄 Config file (lower priority)

Override config by id in the profile's ~/.dsh/profiles/web/cordis.patch.yml (all keys optional; defaults apply when absent):

- id: dsh-voice-chat
  name: 'dsh-voice-chat'
  config:
    asrEngine: siliconflow          # siliconflow | groq | mimo | custom
    asrApiKey: sk-xxxx              # ASR key (legacy single slot: applies to the active engine; or env var DSH_VOICE_ASR_KEY)
    asrBaseUrl: https://api.siliconflow.cn/v1
    asrModel: FunAudioLLM/SenseVoiceSmall
    asr:                            # per-engine ASR config (takes precedence)
      custom: { baseUrl: http://127.0.0.1:8000/v1, model: whisper-v3, apiKey: sk-xxxx }
    llmModel: deepseek-v4-flash     # rewrite model (fallback, follows current conversation)
    silenceMs: 2500
    rewrite: false                  # condensed reading (default off; toggle in settings)
    ttsEngine: edge                 # edge | mimo | custom
    voice: zh-CN-XiaoxiaoNeural     # Edge voice (same as tts.edge.voice)
    ttsBaseUrl: https://api.openai.com/v1   # legacy single slot: applies to the active engine
    ttsModel: tts-1
    ttsApiKey: sk-xxxx
    tts:                            # per-engine TTS config (takes precedence)
      mimo:   { baseUrl: https://api.xiaomimimo.com/v1, model: mimo-v2.5-tts, apiKey: sk-xxxx, voice: 冰糖 }
      custom: { baseUrl: https://api.openai.com/v1, model: tts-1, apiKey: sk-xxxx, voice: alloy }
    rate: '+10%'
    shortTextChars: 50

Restart dsh web after editing. Priority: settings panel > cordis.patch.yml (asr.<engine> / tts.<engine> > legacy flat keys) > env vars > defaults.

Structure

• lib/index.js — host half: routes /stt (ASR, dual-protocol: OpenAI multipart + MiMo chat/completions), /tts (Edge/MiMo/custom TTS), /speak (rewrite + synthesize), /settings (settings panel persistence);
• lib/client.js — browser half: mic/mute buttons, silence detection, single-channel playback, hotkeys (Ctrl+Shift+Space), auto WAV conversion for chat-protocol ASR; injects the "voice chat" settings section into DSH's settings dialog;
• lib/edge-tts.js — inline edge-tts protocol client (Microsoft Edge free read-aloud); the only runtime dependency is ws;
• test/ — settings-layer tests (pnpm test): per-engine isolation + legacy config migration;
• cordis.patch.yml — inserts the dsh-voice-chat line + config example;
• settings.local.json — overrides saved from the settings dialog (generated at runtime, never committed); since v0.4 it stores per-engine slots under asr.<engine> / tts.<engine>.