Chatterfox
Frontier speech recognition that runs entirely on your Mac: instant, unlimited, and nothing leaves the device.
The question I was chasing
Could speech recognition good enough to trust run entirely on a Mac, with nothing sent to the cloud?
The best dictation tools all make the same trade: your voice goes to someone's server, you pay monthly, and you accept a word limit. I wanted to know whether a frontier-grade model could run locally instead, fast enough to feel instant, accurate enough to replace typing, and private by construction rather than by promise.
Why it exists
I talk faster than I type, and most transcription apps punish you for using them: minute caps, subscriptions, and your voice shipped off to a cloud you've never audited. On-device models finally got good enough that none of that has to be true. Chatterfox is the tool I wanted to exist: open it, talk, and the text is ready to paste back into whatever you were already working in.
The constraints
Run a real speech model locally on Apple Silicon, in a native-feeling Mac app, with one rule that couldn't bend: nothing leaves the device. That ruled out the easy path of calling a cloud API, and it meant the model had to be fast enough on a laptop chip that words appear as you speak, because a spinning loader would defeat the entire point. It also had to handle any mic, Bluetooth, USB or built-in, without fiddling, because a silent recording is worse than no app at all.
The decisions that mattered
NVIDIA's Parakeet for recognition, running on-device: it's one of the most accurate speech models around, and local inference is what makes the privacy claim real rather than marketing.
Tauri over Electron, so the app stays small and native instead of shipping a browser to do a system-level job.
Live in the menu bar, not the Dock; you reach for transcription mid-task, so it had to be one click away and otherwise invisible. The detail I kept returning to was real-time feedback: showing words as they land, plus pause and resume, so dictating feels like something you can steer rather than a recording you hope came out.
What it is
A native macOS app that transcribes your voice in real time, entirely on-device. Unlimited length, works in any app, any microphone, customisable themes, menu-bar native, and pay-once, no subscription. Your voice never leaves the Mac.
Built with: Cursor, NVIDIA Parakeet, Claude Code (Sonnet 4.6 and Opus 4.6), Tauri, Remotion
Where it landed
Live in beta on macOS, and the app I now use to draft most of what I write. It's also the clearest proof I have that on-device AI is ready for real daily work, not just demos. What I'd push on next is dictation commands, editing and formatting by voice, so it isn't only capturing words but shaping them.
Part of the Rolling Waves work archive.