How long does AI meeting transcription take?
Minutes, not hours. Hosted speech-recognition APIs transcribe far faster than real time, so a one-hour meeting is usually transcribed within a few minutes of ending. Analysis adds a little more, so a full report typically lands within minutes of the bot leaving the call.
What actually determines the wait
Three things. Audio length, because cost and time scale with it. Chunking, because long recordings are split into pieces and transcribed in sequence — Donna caps chunk length so no single upload is rejected for size, a limit that silently destroyed whole transcripts before it was enforced. And the analysis passes: a first pass extracts the facts, commitments and meeting type, and a second pass writes the report through the chosen lens.
Why two passes are worth the extra minute
A single pass over a long transcript produces a summary that drifts and loses quotes. Splitting the work means the facts — commitments, decisions, participants, meeting type — are extracted structurally first and stored, then the writing pass works from a compact factual base plus the transcript. That is what makes re-analysis cheap later: a different lens re-runs only the second pass over the stored facts, so a new read of an old meeting takes well under the original time.
What happens while you wait
Nothing you have to watch. The recording is compressed server-side, the report appears in the dashboard when it is ready, and an email can be sent on completion. During the call itself there is a live speaker-attributed caption feed if you want to follow along without being in the room — and if you need to end things early, an explicit stop leaves the call and still writes the report rather than throwing the meeting away.
Related questions
Is there a live transcript during the meeting?
There is a live caption feed with speaker names in the dashboard while she is in the room. The analysis itself runs after the meeting ends, when the whole conversation is available.
Does a four-hour meeting take four times as long?
Roughly proportionally for transcription, yes, since chunks are processed in sequence. The analysis grows more slowly. Very long recordings are still measured in minutes rather than hours.
What if transcription fails partway?
A rejected chunk is skipped with a warning rather than killing the run, so a single bad segment costs that segment — not the entire meeting’s transcript.
Put Donna in your next meeting
Access is open — sign in with your Google account and she’s in your dashboard in minutes. Rolling out to a whole team, or want a hand with the server setup? We’re glad to help.