Preview Audrey 0.7: speech recognition for real calls
Join the waitlist
Results · 13 Oct 2026

turn-1-mini on Sesame's TurnBench.

Ranked first on end-of-turn and second on interruption on the held-out test set. What the benchmark measures, and the full numbers with latency spread.

Summary

turn-1-mini is ranked first on end-of-turn and second on interruption on Sesame's TurnBench (held-out test set). The only open-weights model in the top five on either task.

What TurnBench measures

TurnBench is Sesame's benchmark for conversational turn-taking. Annotators mark two kinds of event in two-channel recordings of human conversations, and each model outputs a list of event times for every conversation.

  • End-of-turn: the speaker has finished and the floor is open. A false positive is declaring end of turn during a mid-turn pause.
  • Interruption: one speaker takes the floor while the other is still talking. A false positive is declaring an interruption during a backchannel or because of noise.

The leaderboard ranks models by recall on the held-out test set. A model whose false-positive rate is above 0.15 does not qualify. Latency is the delay after the event.

Our result

Sesame TurnBenchHeld-out test set · 116 conversations
TaskRankRecallFalse-positive rateLatency (median)Latency p10Latency p90
End-of-turn1st of 230.9750.075595 ms156 ms1,243 ms
Interruption2nd of 200.9730.048551 ms295 ms1,576 ms
Sesame TurnBench · turn-1-mini · held-out test set. 116 conversations, scored by Sesame with their scorer. The leaderboard ranks by recall. Recall: share of true events detected. False-positive rate: lower is better. Latency: delay after the event.

We quote latency as the median and show the spread. On end-of-turn, one decision in ten came within 156 ms of the event, and one in ten took longer than 1,243 ms.

The model

turn-1-mini is a causal, streaming model with 5.3M parameters, released under Apache-2.0. Plug-ins are available for Pipecat and LiveKit. Sizes and speed are on the technical note.