Dictation eats the words you typed to fix dictation

Dictation mishears a word it was never going to get — a brand, a surname, jargon. CrossFit. So you do the obvious thing: stop talking, type the word properly with the keyboard, resume speaking.

Then the live transcript updates, rewrites the field, and your typed text is gone.

The compounding failure

Two bugs stacked, and the second one eats the fix for the first:

1. Speech recognition can’t reliably handle proper nouns and jargon. Forgivable — a genuinely hard problem that will always be somewhat true.
2. The system destroys the one obvious human remedy. Typing the hard word by hand is exactly the right move. The interface punishes it.

A limitation you can work around is a nuisance. A limitation whose workaround is demolished is a trap — and it teaches people to stop trying, the worst outcome an interface can produce.

Why it happens — and why the competitor doesn’t have it

This is a consequence of one design choice: when does the transcript enter the field?

Live streaming is nicer to watch — you see it working. But it means the engine is continuously rewriting the field from its own private buffer, and anything a human typed in the meantime isn’t in that buffer. Classic last-writer-wins clobbering: an asynchronous process assuming nothing else can touch what it’s editing. That assumption is false the entire time the user has a keyboard.

The fix already ships in a competing product. You don’t have to imagine the solution — you can go look at it. That’s the strongest possible case for difficulty: none.

Keeping the live transcript and fixing it

Deferring insertion isn’t the only option; live streaming can be kept. The cheap fix: when a keystroke arrives, commit the current hypothesis and start a new segment. The already-spoken text becomes final and untouchable, the typed text sits safely after it, dictation continues from there. That’s a keystroke listener and a buffer flush.

The robust version is a solved problem — text editors and collaborative editing worked this out decades ago: track the range you inserted, only rewrite within it, reconcile rather than clobber.

The principle: one input method must never destroy another’s input. A keyboard and a microphone aimed at the same field are two hands on the same page. The system’s job is to let both write.

Why this is filed as a micro case

Sized on Claude’s own user base — a fraction of whom dictate, a fraction of those who attempt the pause-and-type workaround — it lands around 0.2 lifetimes a year. Real, small, and the third self-indictment-or-near-it on this site.

Difficulty to fix: none

The machine mishears you, so you reach for the keyboard — the one tool that always works. And the machine deletes what you typed, because it assumed nothing else was allowed to write. Commit on keystroke. Two inputs, one field, nobody’s words erased.