I type quickly and make a lot of mechanical mistakes. Missing letters, doubled keys, letters in the wrong order. I don’t always need something to rewrite my sentence. I want it to fix the word and let me keep typing.
That’s what I’m building TypeWake for. It’s a local-first typing corrector, and the current prototype is written in Rust. It started with a dictionary-backed candidate generator and a benchmark. Now there’s a terminal window where I can type normally and have completed words corrected when I press Space or punctuation.
The correction logic weighs likely typing slips and the surrounding words. Neighboring keys, swapped letters, and accidental double presses aren’t equally likely mistakes. A word that is already in the dictionary needs a much stronger reason to change than an obvious misspelling. If no correction is clearly ahead, the word stays as it is.
I also added a typing test that learns my own slips. It presents sentences to type without stopping to fix mistakes, records the mistyped words, and compares corrections with and without that personal profile. The point is to measure whether learning my mistakes actually helps, not just add personalization as a feature label.
Corrections in the test window turn green. Backspace immediately after a fix undoes it and leaves that word alone for the rest of the session. Getting a correction wrong is annoying enough without fighting the same correction again.
The newest work is an IBus engine for Linux, reusing the same word-boundary editor outside the terminal test window. That integration is still on a feature branch. I’m not calling it reliable across applications yet, and there isn’t a public release or a finished Windows or macOS integration.
The early synthetic benchmark was useful for checking the implementation, but its perfect candidate recall wasn’t proof of a useful product. The harder test is whether this fixes my typing without changing words I meant to write.