OpenRouter Voice

System-wide push-to-talk dictation, powered by the OpenRouter API. Hold a hotkey anywhere on your desktop, speak, release — your words appear at your cursor. Built with Rust + Tauri 2 as a reference app for OpenRouter's speech-to-text endpoint.

Where to start reading

You want…	Read
The API call, minimally	src-tauri/examples/minimal.rs — mic → WAV → transcript in ~100 lines
The full integration, explained	docs/INTEGRATION.md — transports, biasing, OAuth, model choice
The integration code itself	src-tauri/src/stt/ — client, model registry, vocabulary
The desktop-app machinery	src-tauri/src/platform/ — hotkeys, tray, overlay (no API code)

# smallest possible taste — record 5s, print the transcript
export OPENROUTER_API_KEY=sk-or-...
cd src-tauri && cargo run --example minimal

Features

• Push-to-talk anywhere — device-wide hotkey (keyboard or mouse side button); hold to dictate, release to type the transcript into whatever app has focus, optional second hotkey that presses Enter to send
• Rolling live preview — while you hold, the whole utterance is re-transcribed every ~1.2s into a floating overlay pill, so the text self-corrects as context grows
• 16 models, benchmarked — every dedicated STT model on OpenRouter plus streaming-capable audio chat models, with real WER/RTFx/price data and an in-app benchmark that ranks them by measured latency on your connection and voice
• Accuracy stack — vocabulary biasing (Deepgram keyterms / Whisper prompts), mis-hearing suggestions, filler-word removal, and an opt-in LLM correction pass
• Background service — menu-bar tray, close-to-tray, launch at login, starts hidden; the main window is optional
• OpenRouter OAuth — "Log in with OpenRouter" (PKCE, loopback callback), or paste an API key; credentials never reach the webview

Architecture

src-tauri/src/
├── stt/             ★ the OpenRouter integration (start here)
│   ├── client.rs       one-shot + SSE streaming transports, LLM cleanup, warm pool
│   ├── models.rs       registry: WER/RTFx/price per model, transport routing
│   └── vocab.rs        per-provider vocabulary biasing + suggestion mining
├── audio/             capture (cpal) → resample (rubato) → VAD gate (Silero)
├── pipeline/          session orchestration
│   ├── mod.rs            hold-buffer, rolling re-transcription, dispatch
│   ├── text.rs           filler cleanup, continuity, type-into-app
│   └── commands.rs       Tauri command surface for the UI
├── platform/          OS machinery — zero API code (skippable for API readers)
│   ├── hotkey.rs         device-wide push-to-talk (poll + raw HID)
│   ├── hid.rs            macOS IOHID mouse-button listener
│   ├── tray.rs           menu-bar presence, background lifecycle
│   ├── overlay.rs        floating dictation pill window
│   └── traffic.rs        macOS traffic-light positioning
├── auth.rs            OpenRouter OAuth PKCE (reusable in any native app)
├── settings.rs / history.rs / benchmark.rs   persistence
└── lib.rs             wiring
ui/                    vanilla JS/CSS, no framework, no build step

Audio flows one direction: audio → pipeline → stt → OpenRouter. The only network calls in the codebase live in stt/client.rs and auth.rs.

Run the app

Prereqs: Rust stable, Node 18+, Tauri prerequisites.

npm install
npm run dev

First run: grant Microphone access, plus Input Monitoring and Accessibility (System Settings → Privacy & Security) for global hotkeys and type-into-app. The app starts in the menu bar — open the window from the tray icon, log in (or export OPENROUTER_API_KEY), bind a hotkey, dictate.

Platform support

	Status
macOS	✅ fully supported (primary target)
Windows / Linux	⚠️ core pipeline is portable; platform/hid.rs (side buttons), platform/traffic.rs, and overlay behavior need porting — see module docs

License & attribution

MIT — see LICENSE. Voice activity detection uses the Silero VAD v5 model (MIT), bundled via the voice_activity_detector crate (MIT); release builds redistribute the model weights.

Privacy

The global input listener is a matcher, not a recorder: key/button state is compared against your two bindings and discarded — no keystroke history exists, in memory or on disk, and nothing but microphone audio (plus model id and optional vocabulary hints) is ever sent to OpenRouter. Logging deliberately excludes key identities and transcript contents. Local files: settings.json (preferences), history.jsonl (your transcripts, local only), benchmark.wav (your recorded reference clip).