Meet Boo: Local Speech-to-Text That Feels Native

Press a hotkey, speak, and the text appears at your cursor. Free, open source, and your audio never leaves your machine.

Photo by Josh Hild on Unsplash

The whole loop is three steps

Put your cursor where you want the text: an editor, a chat box, a terminal, a browser field. Press Ctrl+Shift+Space and talk while a small overlay shows a live waveform. Press it again, and the transcript appears right where your cursor was, and lands on your clipboard too.

There is nothing to sign into, no subscription, no telemetry. Transcription runs entirely on your machine. The app makes exactly one kind of outbound request, ever: optionally fetching a tiny voice-activity model on first run, verified against a pinned SHA-256. Your audio and your words stay with you.

Native on every platform, no Electron

Boo ships a real native frontend per OS: Swift and AppKit on macOS, GTK4 and libadwaita on Linux, plain Win32 on Windows. No WebView, no Electron, no lowest-common-denominator UI toolkit pretending to be three apps at once.

The architecture borrows wholesale from Ghostty, one of the most admired open-source projects of the past few years: a portable Zig core exposed through a stable C API, plus a thin native application runtime per platform. The trade-off is more code per OS. The payoff is an app that feels like it belongs on each one. Cross-platform shouldn’t mean foreign.

Fast enough to feel instant

You choose the model. Boo runs any of the 33 GGML models from whisper.cpp, plus NVIDIA’s Parakeet. On an Apple Silicon GPU, Parakeet transcribes at roughly 120x realtime: 11 seconds of audio in under 100 milliseconds, with near large-v3 accuracy and automatic language detection across 25 European languages.

And if you drop in a sub-1 MB Silero VAD model, Boo switches to streaming: it transcribes each phrase while you are still talking, so committed text appears live and long dictations land near-instantly when you stop.

Built for terminal people too

Boo is a natural companion for Ghostty. On macOS it injects text through Ghostty’s own AppleScript API: no clipboard involved, and it keeps working under Secure Input, so you can even dictate where synthesized keystrokes are normally blocked. It also ships 486 Ghostty-format color themes, searchable from the settings.

Honest about where it stands

The project’s status reporting is refreshingly blunt. macOS is fully working. Linux is a preview: recording, transcription, and both desktop-portal grants pass real end-to-end tests, but no human has driven the UI on a real desktop yet. Windows is experimental: it compiles and passes the core test suite on real Windows runners in CI, but nobody has dictated on real hardware. The repo keeps a public ledger of exactly what is verified and what is not, and bug reports genuinely move platforms up the ladder.

Get started

Boo is MIT licensed, like the projects it stands on. Grab a release from GitHub, download a model with one curl command, press Ctrl+Shift+Space, and start talking.

https://github.com/ilovepixelart/boo