About
We think faster
than we type.
Kudu is a free, open-source dictation app for people who’d rather say it than type it, and who don’t want their voice on someone else’s server to do it.
Most of us speak at around 150 words a minute and type at 40. Dictation should close that gap, but the built-in kind writes down every “um”, and the polished kind sends your voice to the cloud.
Kudu does both halves on your own computer: it hears you, writes what you meant, and formats it for the app you’re in. Then it gets out of the way.
Principles
Four things we won’t trade
Private by construction
Fast enough to forget
Writes like you
Yours to change
How it works
From breath to text, in five steps
All five run on your computer, usually in well under a second.
- 1
Hear
Audio from your microphone, only while you hold the key, kept in memory.
- 2
Transcribe
An on-device speech model turns it into words, on the Neural Engine or the CPU.
- 3
Clean up
Fillers, repeats and “no wait” corrections are removed by fast, predictable rules.
- 4
Format
Lists, dates, mentions and code are laid out for the app you’re in.
- 5
Type
The finished text is inserted where your cursor is. Nothing is left behind.
Under the hood
Open models, running locally
Kudu stands on excellent open work. Thank you to everyone behind it.
Speech on Mac
NVIDIA Parakeet TDT 0.6B v2, on the Apple Neural Engine through FluidAudio
Formatting on Mac
Apple’s on-device Foundation Models (macOS 26 and later)
Speech on Windows and Linux
Parakeet (Accurate) or Moonshine Base (Compact), on the CPU
Formatting on Windows and Linux
Optional Qwen2.5 0.5B or 1.5B, run locally with llama.cpp
Made in the open, for everyone.
Kudu is an independent project. Issues, ideas and pull requests are all welcome.