Ever notice when a transcript accidentally labels your agent as the customer (or vice versa)? Yeah, we noticed too.
We've implemented a few changes that improves the accuracy:
- Continued improvement through our Diarization capabilities. Our latest training round produced scored 22.8% on the latestest CALLHOME benchmark. This is down from 28.5%.
- While segmentation has always been a core ability, this update brings a significant improvement to speaker assignment and counting — marking reductions in speaker confusion and overlap
- Conversation turn speech modeling to catch moments where there was quick replies between speakers.
Improvements
- scorecards Deleted scorecard criteria no longer appear when filtering results—keeping your data clean and your views clutter-free.
Fixes
- ui-ux Fixed drilling down by scorecard criteria from the explorer view
- analysis Summarizing email threads with few exchanges
