Chatterfox
Building voice-to-text transcription without the subscriptions.
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 send the voice to someone's server, renew a subscription monthly, and put a word limit on it. A strong model running locally would have to be fast enough to feel instant, accurate enough to replace typing, and private by design.
I talk faster than I type. On-device models finally got good enough that none of that has to hold. Chatterfox is the tool I wanted to exist: open it, talk, and the text is ready to paste.
The constraints
The real limits were the current state of open-source voice models, and what would actually run well on a laptop. There was no engineering team to retrain models, and no server infrastructure to hide behind.
That is why it is a native Mac app on Apple Silicon, and why the recognition model is NVIDIA Parakeet. One focused build that runs on the devices I could support properly, rather than a thinner tool stretched across every platform.
The decisions that mattered
I started from Handy, an open-source voice transcription project that looked well maintained and solid enough to build on rather than starting from nothing.
I customised the design and limited the models. Narrowing to high performance for English speakers was the point, not offering every option. The result is now comparable with paid alternatives and industry leaders such as Wispr Flow.
Built with: Cursor, NVIDIA Parakeet, Claude Code (Sonnet 4.6 and Opus 4.6), Tauri, Remotion
Where it landed
Chatterfox is in beta on macOS, and it is the app I now use to draft most of what I write.
The rule that made it worth building is also the limit. Nothing leaves the device, so there is no larger cloud model to fall back on. What the local model hears is what gets typed.
Part of the Rolling Waves work archive.