About
Um. Let’s talk about that.
Everyone has a word they lean on. Umbot is a small robot coach that catches yours the moment it slips out, so the next sentence comes out cleaner.
What it does
Umbot listens while you speak and plays a soft cue the instant it hears a filler word: um, uh, and friends. There is no report to read later and no recording to cringe through afterward. Just a gentle nudge, delivered at the one moment it can actually change something.
How it works
Compact acoustic models run inside your browser using WebAssembly. Your microphone feeds them short windows of audio, one model estimates whether speech is present, another scores each moment for filler words, and then that audio is overwritten by the next moment. In the ordinary coach, nothing is recorded or uploaded. A local-development evaluation tool can record a complete session only after it is explicitly enabled.
Why the nudge works
Filler words aren’t a character flaw. They’re a habit your own ears stopped noticing years ago. Feedback that arrives in next week’s review, or even at the end of a practice run, is easy to shrug off. A cue that lands mid-sentence is not. You hear the habit while it’s still warm, and awareness does most of the rest.
Why it stays local
A microphone that’s always on should never phone home. That rule shaped the whole design: the model is small enough to run on your device, so that’s where it runs. It isn’t a privacy toggle we could quietly flip. It’s the architecture. The details live on the privacy page.
The name
It hears “um.” It’s a bot. We considered something cleverer, but a speech coach, of all things, should get to the point.
Where it’s headed
Umbot is test software, built in the open by one person. Expect rough edges, the occasional misfire, and rapid change. If that sounds fun rather than alarming, welcome aboard, and read the terms first.