dsh-dictate

dsh-plugin-dictate

Live speech-to-text dictation for the DeepSeek Harness web UI. Click the microphone button in the composer tool row, speak, and your words appear in the prompt as you talk — mixable with normal typing, at your cursor, without touching what you already wrote.

Transcription runs on OpenRouter's STT endpoint (/api/v1/audio/transcriptions). The API key stays server-side: the browser only POSTs audio segments to a local route this plugin registers inside dsh.

Install

Requires dsh ≥ 0.1.1-rc.1 and a running dsh web profile.

# from npm
dsh plugin --profile web add dsh-plugin-dictate

# or straight from this repo
dsh plugin --profile web add https://github.com/navneetset/dsh-dictate.git

# or a local checkout (link: — edits + rebuild show up on next restart)
dsh plugin --profile web add /path/to/dsh-dictate

dsh plugin add installs the package into the profile and appends it to the profile's bundle list, so one command activates both halves (host route + browser mic button). If pnpm complains about the workspace root (first-time setup on some profiles), add -w right after add. Restart dsh web and reload the page. Updating later: re-run the same add command to resolve the latest version.

Features

• Hands-free mode — tap the mic and talk continuously. Segments are cut on natural pauses (voice-activity detection) and inserted with ~1–2s lag; long continuous speech is force-split well before the 60s upstream ceiling.
• Push-to-talk mode — hold the mic button, release to transcribe. Better in noisy rooms.
• Ctrl+M toggles dictation (configurable off in settings).
• Right-click the mic for settings: mode, STT model, dictation language, hotkey. Stored per browser in localStorage, sent per request — nothing else to configure.
• Default model: openai/gpt-4o-mini-transcribe (cheap, fast). Also good: openai/gpt-4o-transcribe (best accuracy), openai/whisper-large-v3 (strongest multilingual, notably Dutch).
• /dictate slash command prints a status/usage summary.

API key

The host half proxies transcription to OpenRouter and needs one of:

1. OPENROUTER_API_KEY environment variable (preferred — visible to the dsh process), or

2. A plain value in the profile patch (~/.dsh/profiles/web/cordis.patch.yml):

   - id: dictate
     config:
       apiKey: sk-or-v1-...
   

The env var wins when both are set. Without either, the mic button reports Missing OpenRouter API key on first use.

Config reference (host row config)

Key	Default	Meaning
apiKeyEnv	OPENROUTER_API_KEY	Env var name holding the OpenRouter key
apiKey	—	Plain-value key fallback (env var preferred)
endpoint	https://openrouter.ai/api/v1/audio/transcriptions	STT endpoint
routePath	/dictate/transcribe	Local POST route the browser calls
maxAudioBytes	25000000	Segment size cap (matches OpenRouter's limit)

Privacy & security notes

• Audio is captured only while the button is active (recording or push-to-talk held); segments are transcribed in-memory and never written to disk.
• Every segment is sent to OpenRouter (a third party) for transcription — that's the point of the plugin; don't dictate secrets you wouldn't put through your LLM provider.
• The OpenRouter key never reaches the browser.
• This route is unauthenticated like the rest of the dsh webserver — if you bind dsh to your LAN, anyone on it can also reach this route (and transcribe on your OpenRouter account). Keep the firewall rule tight.

Troubleshooting

• No mic button — the plugin must be in the profile's bundle list (dsh --profile web --dump-config | grep dictate) and dsh restarted; the page reloaded. Check the browser console for [dsh-dictate] warnings.
• Missing OpenRouter API key — set the env var or config value above, restart dsh web.
• Microphone permission denied — allow mic access for the page (browser address bar → site settings).
• Wrong language / garbled text — set Language in the mic settings; auto-detect needs a clear first sentence.
• Text lands in the wrong spot — insertion goes to the composer's current draft end; dictation and typing interleave best by speaking first, then editing.
• Debug — window.__dshDictate in the browser console: .state (controller phase), .prefs (active settings), and .inserter('text') to test insertion without the mic.

Scope / limitations

• Web UI only (dsh web). The terminal (TUI) surface has no client-plugin slots for this; not supported.
• OpenRouter's transcription endpoint is not streaming: text arrives in segment-sized chunks (~1–2s behind speech), not word-by-word.
• No TUI, no audio history, no voice commands — it types into the prompt, that's all.

Developing

npm install
npm run build   # builds lib/index.js (host) + dist/client.js (browser)

The profile install is a pnpm link: to this repo when installed from a local path — rebuild here, restart dsh web, reload the page. The client bundle is a window.__ModuleLoader__.load({ id, factory }) envelope whose module exports inject/apply (a Cordis client plugin); the host module exports the standard Cordis name/inject/Config/apply.

License

MIT