What meeting AI still gets wrong (accents, jargon, and who said what)
Modern speech-to-text is genuinely good, which is exactly what makes its failures dangerous. It is right often enough that you stop checking, and the part it fumbles is frequently the part that matters most.
I have run the same meetings through commercial tools and through my own self-hosted Whisper setup, and the failure modes rhyme across all of them.
the four places it reliably trips
My accent is the first. Czech-accented English is not exotic, but it is enough to turn 'deploy' into 'the ploy' and to bend vowels the model expected from somebody else's mouth. The second is jargon: product codenames, our internal acronyms, the name of whatever library happened to come up that week. One tool spent an entire call convinced our service called Vltava was 'well tava', every single time.
Crosstalk is the third. The moment two people on a Brno call talk over each other (which is every genuinely interesting moment), the transcript smears them into one voice. And the fourth is the one nobody has actually solved.
nobody has solved who-said-what
Diarization, attributing each line to the right speaker, is mediocre across the board. Cloud tools do a little better than my garage route because they have more signal to work with, but better here means confusing three speakers instead of five. On a two-person call it is fine. On a six-person workshop in Ostrava it guesses, and it guesses with complete confidence. The confidence is the trap. A hedged guess you would double-check; a clean, certain label you forward without a second thought.
the time it blamed the wrong person
The worst one was not a transcription error.
It was a confident summary.
After a design call in late spring, the tool produced a crisp line stating that a particular architectural decision had been made by my colleague Marek. Except it was my decision. Marek had argued the opposite side the whole time. The diarization crossed the wires, and the summary laundered that mistake into clean, quotable prose that I very nearly forwarded to the client.
A wrong transcript looks wrong; a wrong summary looks exactly as trustworthy as a right one, which is what makes it the more dangerous of the two.
The follow-up was awkward in the small, careful, specifically Czech way these things go: a message to Marek, a correction posted to the thread, and a standing private rule to never again trust an attribution I had not heard with my own ears.
trust the gist, verify the rest
This is the same lesson that turns up wherever these systems meet the real world: they fail softly and plausibly, and they fail hardest at the exact points you most wanted to lean on them.
So my calibration is boring and firm. I trust the gist: the shape of what was discussed, the rough topics, the general drift. I do not trust, without checking, any verbatim quote, any number, or any claim about who owns what. Those are the precise fields I verify by hand before they turn into action items anyone is meant to act on. Where a tool offers a custom vocabulary list, I feed it our codenames, which fixes the Vltava problem and nothing harder than it. It will not teach the model my vowels, and it will not sort out who actually spoke.
The gist is a gift. The specifics are a draft.