How Avette works

A real input method

Avette isn’t a floating window you copy-paste out of, and it doesn’t fake keystrokes or poke at Accessibility APIs. It installs as a macOS input source (a Palette input method), so the text it produces goes straight to wherever you’re typing — your editor, browser, chat app, or terminal — the same path your keyboard uses. That’s why it works even in places clipboard- or Accessibility-based dictation tools can’t, like the terminal.

See what you say, as you say it

While you speak, your words appear as marked text — underlined and highlighted — so you can watch the transcription form and fix it before it commits. Only what you keep is inserted for real.

On-device, always

Your microphone audio is transcribed locally on your Mac. No audio is uploaded; there’s no cloud service and no account. The only times Avette touches the network are to download the model (once) and to validate your license. See the Privacy Policy.

The model

Avette is based on Qwen3-ASR 1.7B, run on your Mac's GPU by a streaming engine we wrote in Swift on Apple's MLX: most dictation apps wrap Whisper or Parakeet; we didn't. It downloads once, about 2 GB.

Languages

Qwen3-ASR supports 30 languages and 22 Chinese dialects. Pick yours in Settings → Languages: with one, Avette always writes in it; with several, it recognizes which one you are speaking.