VidXP's transcript records what was said and when, but not who said it. Speaker diarization divides the audio into turns from different speakers.
Start by making anonymous speakers such as “Speaker 1” searchable inside each video. Linking a speaker to a named person should happen later through user review and the people list in #85.
Expected behavior:
- Record anonymous speaker turns with start/end times and overlapping speech where detected.
- Attach transcript words to the appropriate speaker turn without changing the original transcript text or timing.
- Show and search “who said what” in transcript and evidence views.
- Keep uncertain or overlapping assignments visibly unresolved instead of guessing.
- Use visible person appearances as supporting evidence for a suggested link, never as automatic proof.
- Do not identify the same voice across different videos unless a separate, consented voice-reference workflow is added.
- Measure speaker separation and audiovisual linking separately, including clips with overlapping speech.
A useful first milestone is anonymous, searchable speaker turns with no identity linking. Pyannote is an initial provider to evaluate, not a required implementation.
VidXP's transcript records what was said and when, but not who said it. Speaker diarization divides the audio into turns from different speakers.
Start by making anonymous speakers such as “Speaker 1” searchable inside each video. Linking a speaker to a named person should happen later through user review and the people list in #85.
Expected behavior:
A useful first milestone is anonymous, searchable speaker turns with no identity linking. Pyannote is an initial provider to evaluate, not a required implementation.