Speaker review studio
Experimental shot gallery and reversible interval labels. No face or named-voice matching. Suggestions stay private until an operator reviews exact saved passages.
Source and clock
Use a local recording, or explicitly share the publisher tab with audio. This review source does not start transcription. Capture-relative time = video time + offset. Check speech at several points; drift can invalidate suggestions.
Motion/audio correlation is an experimental activity cue, not proof of speech or identity. Reset the region after a camera change. Wide views remain unverified.
Saved views
Set a view key such as podium-left; assign the occupant for this meeting. Matching background crops suggest the same view locally. Names do not automatically transfer between views or meetings. Enter the saved interval and evidence before publishing a label. All thumbnails below are private. Review suggested intervals carefully: a camera view can show a listener. Reclassify ambiguous views as unknown.
Post-meeting reconciliation
Prepare overlapping batches of 30 passages. The model proposes changes; it never applies them automatically. Run only after capture ends. Protected facts (votes, amounts, negation) require review.