Voco transcribes speech locally and inserts text at the cursor in your current app. It runs on Apple Silicon Macs.
Hold Fn or Globe, speak, and release. Voco transcribes locally and inserts your words at the cursor. Fn + Space starts and stops continuous dictation. Escape cancels recording or pending transcription.
- Replaced an Electron interface and Swift helper with one Rust executable. GPUI and Metal render the dashboard and recording indicator without Chromium, Node.js, or a separate helper process.
- Integrated Moonshine, Parakeet, and Whisper behind one inference worker. All three speech engines run offline after model download.
- Protected text insertion with session generation IDs and destination revalidation. Cancelled results cannot commit text, and changed fields, selections, or password fields block insertion.
- Transcribed a 17.369-second synthetic clip in a median 2.818 seconds with Whisper and 1.384 seconds with Parakeet. Five trials per engine included model verification and loading. See the file transcription method and limits.
The file timings cover the current build, including model verification and loading. The native insertion checks use disposable editors. Live microphone dictation and compatibility with every app need separate checks. See the architecture, current verification record, and insertion report.
You need an Apple Silicon Mac with macOS 26 or newer, Rust 1.92 or newer, and CMake. Apple's Command Line Tools must include Swift 6.2 and the macOS 26 SDK or newer. Cargo compiles or fetches the native C and C++ speech libraries. The first build needs internet access.
git clone https://github.com/darshmahadevia/Voco.git
cd Voco
cargo run -- --showVoco is a Rust application. GPUI renders its dashboard and floating recording pill with Metal. Rust bindings call macOS frameworks for the menu bar, permissions, shortcuts, nonactivating windows, and text insertion. The app runs as one executable, with no browser renderer or separate helper.
I moved Voco to Rust to reduce the runtime dependencies and maintenance of an app that stays in the menu bar. The earlier version used TypeScript and Electron for the interface, plus a Swift helper for macOS integration.
The goals were to:
- Remove Chromium and Node.js from the app's runtime. Voco now renders its dashboard and recording pill through GPUI and Metal.
- Keep recording, shortcuts, cancellation, and text insertion in one application so I can follow and debug the dictation flow.
- Use Rust's ownership and type checks to catch memory, concurrency, and state-handling mistakes before release. macOS integration and native dependencies still need runtime testing.
- Simplify packaging and updates. The app has one executable with its resources, without a separate helper process to start and coordinate.
Lower memory use and faster startup were goals of the port, but I have not measured those gains yet. The speech engines already ran native code, so the language change alone does not establish faster or more accurate transcription. The tradeoff is more platform-specific UI work if Voco expands beyond macOS.
To install or update the macOS app from this checkout:
git pull --ff-only
./scripts/install.sh
open ~/Applications/Voco.app --args --showThe installer builds the current source, verifies the bundle, and replaces ~/Applications/Voco.app. It quits a running copy and preserves settings and downloaded models. Pulling source alone does not update the installed app.
Local builds use your configured signing identity, or ad hoc signing. They are not notarized. See development for packaging and signing.
Normal launch starts in the menu bar with the dashboard hidden. Choose Open Dashboard to open it. Closing the dashboard keeps Voco running. Choose Quit Voco from the menu bar to exit.
- Download a speech model in the dashboard.
- Allow Microphone and Accessibility access for Voco. There is no separate helper to approve.
- Set Press Globe key to to Do Nothing in macOS Keyboard settings. Use a different shortcut for Apple Dictation.
- Click the practice field and try holding Fn. Then dictate into a text field in another app.
Voco supports Moonshine Medium Streaming, Parakeet TDT 0.6B v2, and Whisper Large v3 Turbo in English. Model downloads use the internet. Recognition works offline after installation.
Existing Moonshine and Whisper downloads are reused. The Rust Parakeet runtime needs a separate ONNX download. Existing CoreML Parakeet files stay untouched.
Voco inserts the final transcript once. It refuses insertion if the destination app, text field, selection, or readable text changes. It does not write into password fields, press Return, or send messages. If insertion fails, copy the complete transcript from recovery. Review your words before sending them.
Fast keeps the speech model's transcript.
Balanced removes fillers and spoken restarts with local rules while protecting literal text.
Accurate adds punctuation and capitalization to Balanced output with Apple's on-device FoundationModels. It requires available Apple Intelligence. If formatting is unavailable, times out, or changes protected content, Voco keeps the Balanced output and shows a notice.
Audio and transcript history stay in memory. Voco never saves recordings or transcripts to disk. Turning off Keep recent transcripts clears retained history immediately. Failed insertion text remains available for recovery. Continuous recording stops at ten minutes and transcribes the captured audio.
Settings and usage totals live in ~/Library/Application Support/Voco/settings.json. Model files live in the adjacent models directory. The app migrates previous local preferences and usage totals on its first run.
./scripts/verify.sh
VOCO_VERIFY_UI=1 ./scripts/verify.sh
cargo run -- --diagnosticsThe GPUI verification record distinguishes tested behavior from hardware checks. See architecture for module ownership and cancellation.
Documents under docs/migration, the dated project audits, and the Rust migration records describe earlier builds. Use the guides above for the current app.
Voco application code uses the MIT License. Speech engines, model weights, and Manrope have their own licenses. The app bundles third-party notices.
