dsh-voice-kit 🎙️

Voice input (Web Speech API) and read-aloud (speechSynthesis) for the
DeepSeek Harness web GUI.

中文说明

Status: v0.3.1, published on npm — pnpm typecheck passes, 35 unit
tests pass, pnpm build emits the ecosystem-standard closure-factory bundle
(host half + browser half). Voice input and read-aloud are wired against
the real conversation API (session kit's inputActions/useInput/
useSession), and read-aloud supports Microsoft Edge neural voices
(host-side synthesis via msedge-tts, cached, with automatic fallback to
the system voice). Verified inside a real DSH Desktop profile. Install:
dsh plugin add dsh-voice-kit.

Features

• 🎤 Voice input — mic button in the composer's left rail; speech is
  transcribed (Chrome/Edge webkitSpeechRecognition) and appended to the
  draft. Auto-restarts across Chrome's ~60s session limit; Esc cancels.
• 🔊 Read aloud with neural voices — per-message button at each assistant
  message tail. Default engine is Microsoft Edge neural TTS (晓晓/云希/
  云健/云扬…), synthesized host-side and played back with an <audio>
  element; falls back to speechSynthesis when offline or on failure.
  Markdown and emoji are stripped before speaking; long replies are chunked
  at sentence boundaries (decimals/URLs never split); only one voice at a
  time; the playing message is scrolled into view with an on-screen bubble
  showing what is being read.
• ⚙️ Settings — a first-level settings section: voice engine
  (neural/system), voice, rate, pitch, recognition language, toggles —
  persisted in the voice settings namespace (registered host-side via
  @deepseek-ai/dsh-settings).

Install

# from npm (after publish)
dsh plugin add dsh-voice-kit

# or from a local checkout (development)
dsh plugin --profile web add link:/path/to/dsh-voice-kit

Restart the harness, refresh the web GUI.

⚠️ Voice input is experimental. webkitSpeechRecognition is a
Chromium feature and does not work inside Electron (DSH Desktop): the
mic button enters the listening state but no transcript ever lands. It may
work in Chrome/Edge when DSH is served in a regular browser. Read-aloud
works everywhere (system voice fallback included).

Development

pnpm install
pnpm typecheck   # tsc --noEmit
pnpm test        # vitest (pure logic: markdown stripping / chunking)
pnpm build       # tsdown → lib/index.js (host) + lib/client.js (browser)

Build pipeline is the ecosystem-standard closure-factory bundle
(window.__ModuleLoader__.load) driven by shared/tsdown.client.ts
(adapted from the official DeepSeek Harness packages/client/tsdown.client.ts,
MIT; libExternal option from the dsh-web-ui family bucket, Apache-2.0).

Wiring status

• [x] Locale registration — 3-arg ctx.locale.register(ns, locale, dict) overload.
• [x] Draft insertion — session kit's inputActions.setDraft(appendDraft(current, text))
  with useInput((s) => s.draft) reads (src/client/draft.ts).
• [x] Message text resolution — conversation.chat.assistant-actions owner's
  messageId + useSession snapshot lookup, text blocks only
  (src/client/message.ts).
• [ ] Runtime verification — install into a real DSH profile
  (dsh plugin --profile desktop add link:...) and confirm the three slots
  render in the GUI; adjust composed-prop details if the renderer disagrees.
• [ ] Settings persistence — if the voice namespace does not survive a
  restart, register it host-side via @deepseek-ai/dsh-settings
  (installSettingsSection) in src/index.ts.

Roadmap

• v0.1 — read-aloud (safe, no permissions) + settings scaffold
• v0.2 — voice input with draft insertion (needs the draft API + mic permission)
• v0.3 — auto-send on voice stop, keyboard shortcut, bilingual polish, CI
  compat checks against new DSH rc releases

License

MIT. The bundled shared/tsdown.client.ts adapts official DSH build tooling
(MIT) plus the dsh-web-ui libExternal option (Apache-2.0); see the file header.