dsh-voice

Full-duplex voice mode for DeepSeek Harness: streamed ASR → LLM → TTS with barge-in.

Status

v0.4.0 — barge-in complete. The full voice loop is wired end to end.

Speak into the composer mic: the assistant silences itself (playback stops, host synthesis queue drops), the running turn is cancelled (the stop-button route), and your speech is transcribed locally and submitted. The reply streams back as spoken audio with live captions.

Known limitation: barge-in detection is triggered by the mic's leading speech edge, which relies on browser-level echo cancellation (getUserMedia({ echoCancellation: true })). Loud TTS playback may leak into the mic on some platforms; there is no JS-level AEC.

Demo

The loop: a user prompt streams back as spoken audio sentence-by-sentence, then the user's voice interrupts playback and stops the running turn mid-reply (true barge-in). The mic keeps recording the new speech.

How it works

input:  mic ──RMS endpoint detection──▶ whisper (browser, q8 onnx)
                                           │ text
                                           ▼
        composer draft ──submit──▶ model stream ──llm/stream tap──▶ SentenceSegmenter
                                                                     │
        browser ◀── SSE /dsh-voice-api/stream ── TtsQueue (msedge-tts) ◀──┘
                  (base64 MP3 frames + caption text)

barge-in: speech edge ──▶ engine.skip() + POST /cancel (epoch bump)
                         + session.cancel() when a turn is running

• The llm/stream tap is lossless: every chunk is yielded unchanged, the segmenter only observes. The model stream is never blocked by synthesis.
• ASR runs fully locally in the browser: transformers.js loads from a CDN via a native dynamic import (kept intact by esbuild supported: { 'dynamic-import': true }), the whisper model streams from the configured model host with browser-cache enabled. No API key, no server-side speech processing.
• RMS endpoint detection: 16kHz getUserMedia, 1.3s trailing-silence cutoff, max 30s segment, pre/post padding. Zero dependencies.
• Barge-in is three-layered: local playback queue cleared, host TtsQueue epoch bumped (queued AND in-flight synthesis dropped), and the running turn cancelled when session.running is true. An aborted turn never flushes its trailing half-sentence — exactly what the user interrupted.
• modelHost accepts any HF-compatible mirror (e.g. https://hf-mirror.com for CN networks).

API

Route	Purpose
GET /dsh-voice-api/stream	SSE; event: audio frames {sessionId, seq, text, audio(base64 MP3)}
POST /dsh-voice-api/cancel	{sessionId} drops queued + in-flight synthesis (epoch bump)
GET /dsh-voice-api/config	ASR runtime config {asr: {...}} for the mic button
GET /dsh-voice-api/*	ping: {ok, name, enabled}

Config (bundle patch row):

- id: voice
  name: '@haoku123/dsh-voice'
  config:
    voice: zh-CN-XiaoxiaoNeural
    asr:
      model: onnx-community/whisper-base   # or whisper-tiny / whisper-small
      modelHost: https://huggingface.co    # or https://hf-mirror.com
      cdnBase: https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.2.0
      language: zh                         # zh | en | auto
      autoSend: false
      mode: toggle                         # toggle | hold

Install

dsh plugin --profile web add 
dsh --profile web

Note: needs Node ≥ 22.19 or ≥ 24 (node:zlib zstd APIs).

Tests

npm test                                # segmenter unit tests (pure, no network)
node test/host.integration.test.mjs     # llm/stream tap + real Edge TTS + SSE + /config
node test/bargein.test.mjs              # client inject face wiring (skipPlayback/cancelTurn)
node test/bargein-semantics.test.mjs    # aborted turn no-flush + cancel drops in-flight
node verify-client.mjs                  # client bundle registration/exports/slots/dynamic-import