Skip to the contentVideo tooling
Ami StudioOleksii Zahlukha

Video toolingEar — meeting transcription that does not quietly lie

Ear — meeting transcription that does not quietly lie

Meeting transcription into text and notes — built because of one finding: a normal recogniser that gets the language wrong quietly translates instead of transcribing.

No screenshots: they would show other people's data

What it is

The tool turns a recording of a call or a meeting into text and notes: who said what, what was agreed, what is left to do. It is a familiar job with plenty of ready-made solutions — but they all share one failure mode, and that is what made me write my own.

The failure is this: when a recogniser gets the language of a recording wrong it does not stop and does not complain — it quietly translates instead of transcribing, and along the way rounds off rare words and names into the ones it knows better. That is the worst kind of breakage, because the result looks flawless. A garbled transcript makes you check it; a smooth wrong translation makes you believe it and send it round to everyone who was in the meeting.

Specifications: Python and faster-whisper, with the models running locally — the recording never goes to the cloud. For this job that is not a detail: meetings contain other people’s speech, and sending it to somebody else’s server is wrong both morally and contractually. For the same reason there is not a single screenshot on the showcase.

Nearby work

Back to Video tooling