Summary
A single speaker can be assigned multiple speaker IDs when their vocal tone changes during a turn.
Steps to reproduce
- Record a session with one live speaker.
- Vary vocal tone naturally during the session.
- Inspect the generated transcript around those changes.
Actual behavior
Cues from the same person are written with different speaker IDs, requiring manual correction. This was reproduced during an hour-long single-speaker session with other audio paused.
Expected behavior
Natural tonal variation from one person should retain the same speaker identity.
Technical context
R4 currently treats six dominant 80 ms frames (0.48 seconds) as a stable speaker change in CutConstants.speakerStableFrames. The split happens before cues are written, so the transcript fixer receives it as structural speaker attribution rather than something it can correct.
Reported by Joey Virrueta.
Summary
A single speaker can be assigned multiple speaker IDs when their vocal tone changes during a turn.
Steps to reproduce
Actual behavior
Cues from the same person are written with different speaker IDs, requiring manual correction. This was reproduced during an hour-long single-speaker session with other audio paused.
Expected behavior
Natural tonal variation from one person should retain the same speaker identity.
Technical context
R4 currently treats six dominant 80 ms frames (0.48 seconds) as a stable speaker change in
CutConstants.speakerStableFrames. The split happens before cues are written, so the transcript fixer receives it as structural speaker attribution rather than something it can correct.Reported by Joey Virrueta.