Live Transcription¶
Audio captured from your selected applications is transcribed on this Mac while the meeting runs.
Partial and final¶
Partial recognition is an ephemeral guess. It is displayed, replaced by the next one, and discarded — it is never appended to a transcript and never persisted. Only text the recogniser has finalised becomes transcript material, so a sentence heard word by word leaves one entry rather than one per word.
Language¶
The recognition language is chosen explicitly from the locales the recogniser supports, and locales whose on-device model is not installed are listed and disabled. The language is fixed for a run and is never detected automatically: a meeting that changes language is transcribed in the language that was chosen.
Gaps are reported, not closed over¶
When audio is not transcribed, the lost time is written into the transcript as a structural remark rather than silently skipped:
> **Transcription gap:** approximately 0.8 seconds of audio around 10:01:41 AM was not transcribed; recognition fell behind capture.
Two things produce a gap:
- Backpressure. If recognition falls more than about three seconds behind capture, the oldest audio is dropped to keep memory bounded, and the lost time is reported.
- A recogniser restart. A recogniser that stops by itself is restarted at most twice; audio arriving during a restart is not transcribed and is counted as a gap.
A gap is positioned where the audio fell when the pipeline knows where that was, and is honest about the length alone when it does not.
One incident, one marker¶
A recogniser that has fallen behind does not lose audio once. It keeps losing it for as long as it stays behind, and the pipeline sees that as a loss every half second or so. Those observations are summarised into one incident rather than written out one by one, so a meeting that spent four minutes behind capture leaves one marker instead of several hundred:
> **Transcription gap:** approximately 215.0 seconds of audio was not transcribed between 11:38:00 AM and 11:41:23 AM; recognition fell behind capture.
Read that sentence as two separate facts. The range is how long the meeting was in trouble; the seconds are how much audio that cost. They are not the same number, and the range does not mean that everything inside it is missing: recognition goes on transcribing between the losses, so speech from inside the range is usually in the transcript above and below the marker. An incident narrower than a second keeps the concise wording that names the moment instead.
Nothing is merged that ScribeKit cannot show belongs together. A new incident is started whenever the backlog demonstrably caught up in between, and a pause, a recogniser restart, capture ending or the meeting stopping all close the incident that was open — so two separate spells of trouble stay two markers.
The marker is written when the incident ends, which is the first moment its range is known. While one is going on, the session record beside the transcript notes that it started, so a ScribeKit that never gets to finish still leaves evidence of it: see Recovery.
When recognition cannot be brought back¶
A recogniser that has used up its restarts ends the meeting: capture stops, the transcript and the recording are closed and kept, and the session is recorded as failed rather than left capturing audio that nothing transcribes.
Uncertainty¶
Passages the recogniser was unsure about are collected while the meeting runs and surfaced afterwards in History, against the retained audio where there is any. Nothing is corrected or rewritten. See Review & Playback.