dsh-voice-polish · Voice Input with Polish

Talk to your computer like you would send a voice message: transcription runs on your local CPU (SenseVoice), then DeepSeek tidies your loose spoken words into a clean, ready-to-send paragraph and inserts it into the composer draft. The experience matches Feishu/Lark's "smart" voice input.

English | 中文

Why not just the official voice input?

DSH's built-in experimental voice bundle stops at "transcribe → insert raw text". Whatever mess you spoke is what lands in your draft: fillers, repeats, half sentences, homophone typos and all.

This plugin adds the missing layer: after transcription, before insertion, a model tidies what you said.

🎤 speak → local SenseVoice transcript → DeepSeek tidy-up → review panel → draft

Real example (spoken transcript → tidied):

Raw: I want to ask you, so like, is there that voice input feature now? Like I just talk to the computer and then it, um, tidies it into text for me. Ideally like that tidies up my, uh, colloquial expressions before sending. That's it, thanks.

Tidied: I want to ask: is there a voice input feature now? I'd talk to the computer and it would tidy my speech into text — ideally tidying up colloquial phrasing before sending. Thanks.

No expansion, no summarizing, no added content — just tidy. Homophone typos from the recognizer get fixed on the way.

Interface

Before (native composer toolbar):

After (the Voice button appears):

Recording in progress (live level + timer; "Stop & tidy" transcribes and tidies):

Features

• Microphone button in the composer toolbar — click to start, "Stop & tidy" to finish; live level bar + timer; Esc or Cancel discards
• Local recognition — SenseVoiceSmall (ONNX, CPU), auto / zh / en / yue / ja / ko; one-time model download (int8 ≈ 239 MB) with automatic HuggingFace / HF-Mirror fallback, fully offline afterwards
• Review panel — the tidied text is editable; toggle raw transcript, re-tidy, or insert raw — a failed tidy never loses your words
• Three tidy-up styles — light (fillers + punctuation only), standard (Feishu-like, default), strict (further written-form normalization)
• Zero extra setup — tidying reuses the DeepSeek route already configured in your profile; no new API keys
• Light/dark theme safe (all host theme tokens)

Privacy

• Audio is transcribed inside your machine's process and never leaves it
• Only the text transcript is sent to your configured model for tidying (usually a few hundred tokens)
• Nothing is persisted; results become a session message only when you press send

Install & first run

1. Install this plugin (pick one):
   • DSH plugin market (dshmarket / Discover tab): search dsh-voice-polish and one-click install
   • CLI · prebuilt Release tarball (no build-script approval, recommended):

     dsh plugin --profile <profile> add https://github.com/howie79794-sys/dsh-voice-polish/releases/download/v1.0.0/dsh-voice-polish-1.0.0.tgz
     

   • CLI · GitHub source: dsh plugin --profile <profile> add github:howie79794-sys/dsh-voice-polish
2. Refresh the page — a "🎤 Voice" button appears in the composer toolbar
3. First click walks you through the one-time local model download; allow microphone access in the browser and the OS
4. Speak

Configuration

Edit the voice-polish row's config in your profile patch (toggle the plugin once to apply):

Field	Default	Notes
language	auto	Recognition hint: auto / zh / en / yue / ja / ko
style	standard	Tidy-up level: light / standard / strict
polishProvider	deepseek-official	Model route for tidying; falls back to a configured DeepSeek route automatically
polishModel	deepseek-chat	Model for tidying
maxDurationSeconds	180	Per-recording limit
enabled	true	Master switch

Compatibility

• Requires DSH 0.2.2-rc.1+ (which ships the local recognition base).
• Do not enable the official "Voice Input" bundle at the same time — both register the same local recognizer. Disable one first if you want to switch.
• Browser microphone capture requires HTTPS or a loopback address (DSH Desktop / local dsh web satisfy this).

How it works (for developers)

• Host half: loopback route /api/voice-polish (status / prepare / dictate / polish); transcription via ctx.speechToText, tidying via ctx.llm.prepareCall().stream()
• Client half: registers into the conversation.input.activity slot; MediaRecorder capture → Web Audio resample to 16 kHz mono PCM16 WAV → base64 upload
• Recording pipeline adapted from @deepseek-ai/dsh-experimental-client-ui-voice-input (MIT, DeepSeek)

License

MIT — see LICENSE.