@nn12138/dsh-voice
DeepSeek Harness 的语音输入插件：麦克风 → 本地/浏览器语音识别 → 作为普通聊天消息提交文本（仅输入，不依赖预设）
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @nn12138/dsh-voice@0.2.6Voice 以 @nn12138/dsh-voice 发布，当前版本为 0.2.6。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。
English | 中文
A voice input plugin for DeepSeek Harness: click 🎤 in the web UI (or press a hotkey), speak, and the recognized text is submitted as a normal chat message. Input only — it never touches the agent preset/persona, so it behaves like "another input method" in every mode.
dsh-voice-models when you want native offline recognition# Plugin
dsh plugin --profile web add @nn12138/dsh-voice
# Optional: offline native recognition (the plugin does not auto-install this runtime)
dsh plugin --profile web add sherpa-onnx-node
dsh-voice-models # one-shot model download (~100MB) → ./dsh-voice-models
# Optional: configuration (edit ~/.dsh/profiles/web/cordis.patch.yml)
- id: voice
config:
modelDir: './dsh-voice-models' # native ASR model directory
hotkey: 'ctrl+space' # global hotkey
vadThreshold: 0.3 # lower = less clipping at sentence boundaries
tailPadSeconds: 0.6 # tail-padding duration
engine: auto # auto (default) | native | browser
The row-level config is received by the host half. engine and hotkey are synced to the browser half over the /voice.config loopback RPC, so there is no separate client config to write. auto probes host native capability: with a model it uses native; without one it falls back to Web Speech, so zero-config users keep working. Restart dsh web after changing the config.
dsh web # 🎤 button appears on the left of the composer, or press Ctrl+Space
See USAGE.md and INSTALL.md (Chinese) for details.
Browser captures mic audio (auto-resampled to 16 kHz)
→ PCM base64 chunks (256 ms) → /voice RPC channel (loopback)
→ host: silero VAD + zipformer2 streaming decode
→ partials returned per chunk (live echo) / finals committed
(VAD segmentation + 0.6s tail padding)
→ conversation service submits the text (same path as typing)
Engine selection: the host resolves the effective engine (config + model-load result) and the client consumes it via /voice.ping — native unavailable falls back to browser Web Speech. /voice.config carries the row-level engine/hotkey from host to client.
pnpm install --ignore-workspace # standalone deps (no DSH monorepo needed); no install-time scripts
pnpm --ignore-workspace test # unit tests (including real-model smoke tests)
pnpm --ignore-workspace typecheck # type check
pnpm --ignore-workspace build # build (tsc host half + tsdown client half)
pnpm --ignore-workspace verify:package # pack-level checks: scripts-free install, complete runtime files
The build only ever runs on the publisher's side (prepack — at npm pack/npm publish time) and in CI before release; consumers installing @nn12138/dsh-voice from the registry never execute lifecycle scripts, so --ignore-scripts installs are complete and usable (see issue #2).
Real-model smoke tests look for the local voxelf assets and skip when absent; override with:
DSH_VOICE_MODEL_DIR (model directory) / DSH_VOICE_TEST_WAV (test wav) / DSH_VOICE_DOWNLOADED_MODELS (downloaded model directory).
Layout: src/index.ts (host half) / src/client/ (browser half) / src/core/ (recognition core) / tools/ (wire-protocol smoke tools + model downloader).
MIT