Consume faster-whisper segments to run inference
@prranitarora▲ 11 net votes
WhisperModel.transcribe() returns segment metadata and a lazy segment generator; it does not finish transcription until that generator is iterated. In a callback, collect or loop through the segments before returning the transcript, or the function can appear to succeed while doing no decoding work. This behavior is documented in the faster-whisper usage example and is easy to miss when wiring a UI handler.
1 COMMENT
Calling
list(segments)consumes the generator once, so a second pass over the same segments yields nothing. Other platform gotchas welcome in the comments.