Repository navigation
Upgrade from 2.x
SpeechSession.stopTranscribing() has been replaced by two operations:
- Call
await session.finishTranscribing()to finish live capture and receive final results for audio already queued. Keep consuming the result stream until it ends. For an active file or media transcription, Finish waits for end of file. - Call
await session.cancelTranscribing()to abandon the run. The result stream ends withCancellationError. Results already yielded may still be buffered by the stream.
Audio and transcription
- Transcribe a selected audio track from a local video or media file on iOS and macOS 27 with
streamTranscription(fromMediaFile:audioTrackID:options:). Results preserve timestamps on the media timeline. (#56) - On OS 27, screen capture can feed compatible sample buffers directly into Speech and convert other audio with
AnalyzerInputConverter. The pipeline preserves timing across capture changes and drains buffered audio on Finish. (#54, #57) - Use the throwing, Sendable microphone tap on OS 27 for AVAudioEngine input. The automatic native microphone provider and OS 26 support remain available. (#55)
- Log native capture state after start, stop, and source changes on OS 27. Fix a crash when the first Speech Recognition permission reply arrives off the main queue. (#52, #53)
This release also includes the Speech session reliability fixes since 2.1.1, including safer capture transitions, shared model reservations, vocabulary isolation, and bounded audio-file reads. Thanks to @aehlke for the original fixes and the Apple Speech sample license notice.