Why Codictate exists
My Wispr Flow trial was running out, so I built my own.
I wanted to learn a framework called Electrobun, and a dictation tool was the most useful thing I could think of to build with it. To make it harder I set one rule: no AI assistance until I could dictate my prompts to it using my own software. That rule is why Codictate works the way it does: it had to be good enough to type with before it was allowed to help me build itself.
Who builds it
What on device means here
Press a shortcut, speak, and your words appear wherever your cursor already is: your editor, your browser, chat, email, a terminal. The speech model runs on your machine. There is no account, no telemetry, no analytics and no crash reporting; your audio is held in memory during transcription and discarded.
The only network request the app makes is downloading the model file the first time you use it. After that it works with the Wi-Fi off. It runs on macOS (Apple Silicon) and Windows, ships with Whisper Large V3 Turbo, and lets you swap models when you want smaller or more accurate.
How it differs from a cloud tool
A cloud dictation product records you, sends the audio to a server and sends text back. That buys it a very large model, and sometimes that model is better than anything that fits on your laptop. We measure that instead of arguing about it, and we publish the languages where we lose.
Every figure behind that claim, including the head to head against Wispr Flow, is on the benchmarks page.