dsh-voice-input-qwen-asr

September 5, 2026 · View on GitHub

简体中文 · English

A DeepSeek Harness (DSH) dual-face plugin that adds voice input to the chat composer: a mic button next to the send action, a live recording bubble, and a local Qwen3-ASR inference service managed entirely by the host.

Demo

Voice input demo (15s)

▶ Click the frame to play on GitHub (opens the built-in video player) · or download the clip.

Features

  • Mic button — registered into the conversation.input.right slot, rendered on the left of the composer's send button. Click to start, click again (or press "Done, insert") to stop.
  • Recording bubble — a portal-anchored bubble with a pulsing indicator, RMS-driven wave bars, elapsed time, and the live partial transcript streamed back from ASR.
  • Real-time streaming — the browser captures 16 kHz PCM16 mono via AudioWorklet and pushes chunks over a host-relayed WebSocket (/api/dsh-voice-input-qwen-asr.ws) to the local ASR service; partial transcripts stream back in real time.
  • Local ASR service — the host half spawns python/asr_server.py with the runtime virtualenv's Python. The server loads Qwen3-ASR-0.6B (transformers backend, CUDA when available) and serves a WebSocket transcription endpoint.
  • Environment setup page — a settings.section page ("Voice") that clones the runtime repo and model repo, creates a Python virtualenv inside the runtime directory, installs dependencies, starts/stops the service, and streams install/server logs.

Requirements

  • DSH host with webServer + connection services (web profile)
  • git (and git-lfs — model weights are distributed via LFS)
  • Python 3.12+ (on PATH, or an absolute path configured in settings)
  • Optional: NVIDIA GPU + CUDA for fast inference; CPU fallback works but is slow

Install

dsh plugin --profile web add ./dsh-voice-input-qwen-asr   # or npm name / .tgz / github:you/repo#sha
dsh --profile web --dump-config                  # verify the layer
dsh --profile web                                # restart the host

Usage

  1. Open Settings → Voice, press Install all (or run the four steps individually): clone runtime → clone model → create virtualenv → install deps. Watch the install log; the deps step downloads torch and may take a while. pip mirror can be configured on the same page.
  2. Press Start service and wait until the state chip turns "Running" (device shows cuda:0 or cpu).
  3. In any chat, click the mic button left of the send button, speak, then press Done, insert — the recognized text is appended to the input draft. "Cancel" discards the recording.

Architecture

browser                          host (Node)                      local python
┌─────────────────┐   POST /dsh-voice-input-qwen-asr/api   ┌──────────────┐
│ mic button      │ ────────────────────────────▶ │ ops: status  │
│ bubble (portal) │                               │ install / sv │
│ settings page   │   WS /api/dsh-voice-input-qwen-asr.ws  │ installer    │
│  AudioWorklet   │ ────────────────────────────▶ │ relay (ws)   │──▶ ws://127.0.0.1:18787/ws
│  PCM16 @16k     │ ◀──────────────────────────── │              │    asr_server.py
└─────────────────┘    partial / final JSON        └──────────────┘    Qwen3-ASR-0.6B
  • src/host/ — cordis plugin: HTTP API route (trust-fenced), WebSocket upgrade route (connection-service auth), installer (git clone / venv / pip), ASR process manager, relay hub, JSON store at $DSH_HOME/dsh-voice-input-qwen-asr.json.
  • src/client/ — browser plugin: conversation.input.right + settings.section slots, i18n (zh/en), styles, mic capture, WS client.
  • python/asr_server.py — asyncio + websockets server; buffers incoming PCM, periodically transcribes a tail window for partials, transcribes the full buffer on stop.

Build

pnpm install
pnpm run check    # tsc -b (host + client programs)
pnpm run build    # → lib/index.js (node) + lib/client.js (browser factory)
pnpm run verify   # simulate host module loader, assert factory shape

License

MIT