Speaker diarization explained: how it works and where it fails

Word error rate is the standard accuracy metric, but it understates the problems that matter most for documentary audio: proper nouns, accents, crosstalk, and technical terms. Here is what WER measures and what it does not.

A practical guide to transcribing interviews for documentary post: when AI is good enough, when human review is needed, how to handle speakers, timecode, and overlap.

Running your own pyannote diarization stack makes sense for custom fine-tuning, offline requirements, or research. For production documentary work, the real cost is not the model but the annotation and tuning loop.