What does AI cleanup actually fix?
It fixes spoken self-corrections and Vocabulary terms heard slightly wrong, and never rewords you; filler words and spoken punctuation are handled without it.
AI cleanup does two things a plain transcript cannot: it resolves a spoken self-correction, and it fixes a name or term from your Vocabulary that the speech model heard slightly wrong. It is optional, off by default, and it never rewords what you said.
#What happens without it
Plenty is tidied in every dictation, instantly and with no AI involved: filler sounds such as um and uh are removed, a word said twice by accident is typed once, spoken punctuation becomes the mark, snippets expand, and a Vocabulary term written with the wrong capitals is put right.
#What cleanup adds
- Spoken self-corrections. “Meet on Monday, no wait, Wednesday” becomes “Meet on Wednesday.” It listens for the ways people correct themselves out loud: sorry, no wait, scratch that, strike that, make that, I mean, I meant, actually, and let me rephrase.
- Your Vocabulary, spelled right. A name or a term from your list that came out close but wrong is corrected to the way you saved it.
#What it will never do
It does not reword, shorten or “improve” you. Anything the model proposes beyond those two jobs is thrown away in code, and your own words stand. The thing you dictated is the thing you meant, including the hedges.
Most dictations have nothing to correct, and those are never sent to the model at all, so they cost nothing. When there is a correction, the model sees only the sentences around it. On a 2018 Intel MacBook Pro that takes a few seconds, and up to about thirteen on a long dictation. See Why is AI cleanup slow?
Turn it on under Settings → Cleanup. It needs either the cleanup model on your own computer or your own cloud key: see running cleanup on your own machine or using your own API key.
Cleanup can only fix a word it has something to fix it against: a correction you spoke, or a term in your Vocabulary. An ordinary word the speech model heard wrong is a recognition problem, which a larger speech model fixes. Cleanup is skipped while Code Mode is on, because it is built for sentences, not symbols.