Review WhisperX Speaker Diarization on a Two-Speaker Recording
WhisperX adds alignment and optional diarization. A speaker ID groups segments by a predicted voice, so map labels to verified participants only after reviewing the recording.
Worked example
Illustrative correction after listening
0.30–2.10 SPEAKER_00: “Mine is the blue folder.” [confirmed Mina] 2.15–3.40 SPEAKER_01: “I'll bring the green one.” [confirmed Lee] 3.15–3.60 [overlap: review before assigning]
Steps
- Follow the WhisperX setup and model access instructions. Start with two authorized speakers, each saying a line you can clearly attribute.
- Run diarization and inspect the output near speaker changes. Preserve the original labels when a voice cannot be safely attributed.
- Review interruptions and crosstalk against the recording. Export a speaker-labeled transcript and flag any lines that need human review.
What to check
- Check multiple turns before replacing a label with a name.
- Keep the timestamped output alongside the cleaned transcript so reviewers can revisit crosstalk.