Diarization & Real Audio

ASR gives you what was said; diarization gives you who. They're separate pipelines glued afterward — the transcript says 'let's ship Friday', the diarizer says speaker 2 at 14:03.

The classic pipeline:

1. VAD (voice activity detection) — speech vs. silence/noise/music, frame by frame 2. Segmentation — split at speaker-change points 3. Speaker embeddings — encode each segment to a vector (x-vectors, d-vectors): the 'voice fingerprint' 4. Clustering — group segments by embedding proximity; each cluster = one speaker label 5. Re-segmentation — clean up boundaries, resolve ambiguities

Then ASR transcripts get attributed by timestamp. Every stage can fail independently — and each failure looks like a different kind of wrongness.