Glossary
Speaker diarization
Updated · By the WithTranscript team · How we checked
Speaker diarization: Speaker diarization is the step in transcription that works out "who spoke when": it splits the audio into turns and labels each turn with a speaker (Speaker A, Speaker B), without knowing their names.
Speech recognition turns audio into words; diarization groups those words by voice. Its output is a set of turns, each with a start time, an end time and an anonymous label. Putting real names on the labels is a separate step, done by a person or by matching the labels against names said in the audio.
It matters most for interviews, podcasts and panels, where a wall of unattributed text is hard to quote. It is also where transcripts go wrong most visibly: two similar voices merged into one, or one speaker split in two when the audio changes.
YouTube captions do not carry speaker labels. Auto-generated captions sometimes mark a change of speaker with ">>", which a tool can turn into turn breaks, but not into names.
How WithTranscript handles it
On captions transcripts WithTranscript marks each speaker change (from the ">>" markers) with a new timestamped paragraph and a dash. It does not name speakers.