Research2026-08-28
Sesame released TurnBench, an open benchmark that measures how well voice systems judge when to speak, when to stay quiet and when a person is merely acknowledging rather than interrupting. It comes with a recorded and hand-labelled corpus of two-person English conversations, a public leaderboard and scores for 14 existing systems. The data is English-only, studio-recorded and licensed for non-commercial use.
What changed
Turn-taking was mostly evaluated on Switchboard, a 1990s corpus of phone calls between strangers, with backchannels labelled by rule-of-thumb.
What it unlocks
Comparing voice assistants on when they speak, interrupt or stay silent, using a shared dataset and scoring method.
- 30 hours of annotated speech
- 104-hour training set
- 14 systems scored
- best system 0.845 end-of-turn recall
What you need to act on it
- non-commercial licence for the data, voice cloning prohibited
- English-only, studio-recorded two-person conversations
Sources