Hervé Bredin explains why identifying who spoke and when adds useful context to a transcript, and describes challenges including overlap, short turns and unknown numbers of speakers.
He demonstrates diarization evaluationSpeaker diarization divides recorded speech into time segments and labels which speaker is talking in each segment. using a reference recording and discusses confusion, missed speech and false alarms, noting that performance varies substantially with acoustic conditions.
He then demonstrates the difficulty of aligning word timestamps with speaker turns and describes an approach that combines diarization and transcription.
Watch on YouTube




