dsh-voice-input-qwen-asr
语音输入插件（双界面）：麦克风按钮位于编写器发送操作旁，实时录音气泡将 PCM 流式传输至由主机管理的本地 Qwen3-ASR Python 服务；另提供 ASR 环境设置页面（克隆运行时/模型仓库、创建 venv、运行服务器）。双语界面（zh/en）
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-voice-input-qwen-asr@0.1.1
Voice Input Qwen Asr 以 dsh-voice-input-qwen-asr 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
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.
▶ Click the frame to play on GitHub (opens the built-in video player) · or download the clip.
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./api/dsh-voice-input-qwen-asr.ws) to the local ASR service; partial transcripts stream back in real time.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.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.webServer + connection services (web profile)git (and git-lfs — model weights are distributed via LFS)PATH, or an absolute path configured in settings)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
pip mirror can be configured on the same page.cuda:0 or cpu).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.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
MIT