On-device Speech¶
Speech recognition uses SpeechAnalyzer and SpeechTranscriber, Apple's
on-device speech APIs, against a language model installed on this Mac.
ScribeKit checks that the model for the selected language is installed and
refuses to start when it is not. It never downloads one on your behalf and
never falls back to network recognition. It does not use SFSpeechRecognizer,
the older API that can send audio to Apple's servers. The app ships without the
network client entitlement, so the sandbox does not permit it to open a network
connection at all. See Network Policy.
A run is not the meeting¶
A recognition run counts its offsets from its own first frame, and a meeting
outlives any number of runs. What a run reports is therefore put back on the
meeting's own media timeline before it is displayed or written:
MeetingRuntime holds one accumulator, mediaOffsetBase, set from a captured
media clock and added to every offset a run reports — segments, and the
position of dropped audio alike.
The base is read from captured frames rather than from the wall clock, so nothing a pause spends can leak into it. An offset in the transcript therefore names the same second of the recording however many times the recogniser was rebuilt.
The origin that mapping uses changes only when every event the previous run published has been handled: the boundary a pause reaches by draining, and the one a self-restart has to schedule, because it happens inside the handler for the event that caused it.
Restarts¶
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, and a restarted run's spans keep the meeting's own offsets rather than starting again at zero.
A restart also closes whatever backpressure incident was open. The run that was losing audio is being torn down and the next one counts from a new origin, so folding losses from either side of it into one marker would hide one failure inside another.
A recogniser that cannot be brought back ends the meeting rather than becoming a state to sit in: capture stops, the durable artifacts are closed and kept, the session is recorded as the failure it was, and every hold on the process is released.
Honest reporting¶
A missing model, an unsupported language, a recogniser that stops by itself, and audio that recognition fell too far behind to transcribe are all reported rather than absorbed. Transcription uncertainty is surfaced, not hidden.
Reporting is summarised, never reduced. A backlog that lasts loses audio every half second or so, and those observations become one incident with one marker; the incident's total is a sum over distinct discarded audio, stated beside the range it fell in rather than derived from it. The evidence that separates one incident from the next is the backlog's own behaviour — it has accepted a full capacity of audio without having to evict any — rather than a delay chosen for readability.