dsh-voice-live

Live streaming voice input for the DeepSeek Harness Web composer.

Speak, and words appear while you talk; press Done and the transcript is corrected once, then inserted into your draft. Recognition runs on your machine — audio never leaves the computer.

中文说明 · Releases

The live caption card sits above the composer: settled sentences in normal color, the sentence being recognized in grey.

Features

• Streaming captions — the sentence in progress is re-decoded on a sliding window every ~0.5–0.8 s.
• Sentence-level final decode — when Silero VAD detects an endpoint, the complete sentence is decoded again, so final quality equals whole-clip decoding.
• Clean audio path — 16 kHz PCM straight from the microphone with browser echo cancellation, noise suppression and AGC disabled; SenseVoice ITN (digits) and punctuation enabled.
• Optional model-assisted correction — a minimal-edit pass fixes homophones (查件 → 插件), product names (拼森 → Python), digits and punctuation. It keeps your wording, and silently falls back to the local transcript on timeout or implausible output.
• No manual model setup — the first use asks DSH's own speech service to prepare the SenseVoice model, shows progress, then starts recording automatically.

Requirements

• DSH Desktop 0.2.2-rc.1 or newer. Tested on Windows x64; macOS and Linux are expected to work because the engine is resolved from the installation, but are untested.
• No npm dependencies: sherpa-onnx comes from the DSH installation and the model from DSH's speech service.
• About 240 MB of model on disk and ~400 MB RAM while the engine is warm (released after idleTimeoutMs).

Install

From npm

Install the bundle by its package name — in the Plugins page, or by asking your agent to install it:

dsh-voice-live

From this repository

git clone https://github.com/18301301961/dsh-voice-live.git

Then install that directory as a DSH bundle — either through the plugin manager with the absolute path, or by asking your agent to install it.

From a git source

github:18301301961/dsh-voice-live#v1.1.0

Reload the page once after installing; the microphone button appears to the left of the send button. See PUBLISH.md for the full distribution guide.

Usage

1. Hover the microphone to warm the engine (~3–4 s, once per idle period).
2. Click it and speak: settled sentences render in normal color, the sentence in progress in grey with a caret.
3. Done inserts the text at the cursor; ✕ or Esc discards it.
4. If the composer changed during dictation, the transcript stays on the card and Insert into composer writes it at the current cursor.

Configuration

The bundle row lives in cordis.patch.yml:

Field	Default	Meaning
language	zh	auto, zh, en, yue, ja, ko
threads	4	ONNX Runtime threads
partialIntervalMs	480	Minimum live-caption refresh interval; widens automatically with decode cost
partialWindowSeconds	6	Audio window used for the sentence in progress
vadThreshold	0.5	Silero VAD threshold; raise it in noisy rooms
minSilenceSeconds	0.35	Silence that ends a sentence
maxSpeechSeconds	12	Longest continuous sentence before a forced break
idleTimeoutMs	300000	Release the engine after this idle time
polish	final	final corrects the transcript once at the end; off keeps the local text
vocabulary	[]	Terms the recognizer tends to misspell, for example [DSH, Cordis]
polishInstructions	empty	One extra sentence appended to the correction prompt
polishMaxLatencyMs	8000	Correction timeout; on timeout the local text is used

How it works

• engine.mjs — Silero VAD segmentation plus SenseVoice decoding; the in-progress sentence is re-decoded on a sliding window, the finished sentence once on its complete audio.
• worker.mjs — a child process (Electron Node mode) hosting the engine on a random loopback port with a per-process token, so model loading never blocks the Host.
• index.js — the /voice-live/* routes (same-origin and x-voice-live checks), sessions, warm-up, model preparation, audio forwarding and finish handling.
• polish.mjs — the correction pass: cheapest reasoning effort, temperature 0, hard timeout, and two guards (length ratio and character overlap) that reject rewrites.
• client.js — AudioWorklet capture (linear resampling to 16 kHz when needed), the caption card and draft insertion.

Privacy

Audio is decoded locally and never uploaded. With polish: final, only the finished text is sent to the model configured in DSH — which is where that text was headed anyway.

Limitations

• The recognition model is shared with DSH's built-in voice input; this plugin ships no model of its own.
• Uninterrupted speech is split at maxSpeechSeconds; the correction pass usually repairs the punctuation seam.
• The composer integration depends on DSH client slots and is untested against future releases.

License

MIT — see LICENSE.