Speaker attribution and refinement improvements
1.3.1·
Improvements to speaker detection, attribution, and note refinement in recordings, plus a fix to how refinement reads the underlying transcript.
New features
- Phantom speaker detection. Recordings now flag speakers that likely don't belong, reducing the time spent cleaning up a transcript.
- Participant context summaries. A summary of confirmed participants is generated once every speaker is confirmed, feeding more accurate notes.
- Automatic title and summary regeneration. Recording titles and summaries regenerate after speaker confirmation so the attribution stays accurate.
- Suggested speaker state for uncertain matches. Recordings now propose a likely speaker for medium-confidence matches, confirmable one at a time or all at once.
Improvements
- Streamlined speaker assignment. Speaker assignment now narrows candidates to the relevant people, fills in solo recordings automatically, and displays participant names.
- Participant context in refinement. Note refinement now factors in who attended, improving the accuracy of refined notes.
- Refinement reads the full transcript. Fixed an issue where refinement drew on a short summary instead of the complete transcript.
- Transcript readiness validation. Transcript cards now indicate when their content is not ready, so refinement always runs against a complete transcript.