TurnBench

TurnBench

A multi-domain benchmark for evaluating conversational turn-taking. We hand-annotate end-of-turn and interruption events in dual-channel human conversations, and measure how accurately and how quickly models detect them.

Mid-turn pauseEnd-of-turnInterruptionBackchannelstime →

Leaderboard

The ideal model is capable of high recall while maintaining a low false-positive rate and low latency. The following models are evaluated against our held-out test set and ranked by recall.

#ModelRecall FPR Latency
1Voice Activity Projection0.8450.055368 ms
2ESPnet Turntaking0.8260.078862 ms
3WavLM Large Anchor0.8000.0541076 ms
4Mimi Endpointer0.7820.078645 ms
5Kyutai Semantic VAD0.7730.0591007 ms
6Smart Turn v30.7520.0471017 ms
7ESPnet Turntaking Perchannel0.7110.081730 ms
8Gemini 3.1 Live0.6570.0221234 ms
9WavLM Large Causal0.4080.054683 ms
10WavLM Base Causal0.4030.061701 ms
11OpenAI Realtime (Semantic VAD)0.3030.018793 ms
12Moshi0.2330.044702 ms
13OpenAI Realtime (Server VAD)0.9550.525282 ms
14RMS Energy VAD0.7180.632-117 ms
Recall:
share of true events detected (higher is better).
FPR:
false-positive rate on test (lower is better); above 0.15 does not qualify.
Latency:
median delay after the event, p50 in ms (lower is better).

Click any model to open its full report: dev-set scores per conversation, with its predictions overlaid in the viewer.

Dataset

TurnBench evaluates models on a 30-hour corpus of studio-recorded, dual-channel dyadic speech. The corpus is provided by Mundo AI and features 154 dialogues, 106 actors, and is balanced across 6 conversation types1. Every dialogue is labeled by three independent annotators (Fleiss's κ = 0.78) where the ground truth label is derived from 2/3 consensus.

Additionally, we include a training dataset, otoSpeech, hand-labeled under the same protocol as the evaluation set. All three splits are available on Hugging Face.

train
104 hotoSpeech
collected by Oto
audio + annotations
dev
7.3 hturn-benchmark-dev
collected by Mundo AI
audio + annotations
test
22.9 hturn-benchmark-test
collected by Mundo AI
audio only

We created an interactive data viewer for the public dev set. It is an audio player that displays each conversation with the annotated labels drawn on the waveform. You can compare these labels with a published baseline's predictions or submit your own.

Evaluation

Per conversation, TurnBench requires models to output a discrete list of times2 for the following events:

End-of-turn
When a speaker has finished speaking and the floor is open for the other speaker to take.FP: declaring end-of-turn during a mid-turn pause.
Interruption
When a speaker takes the floor while the other is still talking.FP: declaring an interruption during a backchannel or because of noise, channel bleed, or echo.

We open source the scorer, submission format, and baseline implementations on GitHub.

Results

We find end-of-turn recall is invariant to conversation type1, while interruption false positives are not, tending to concentrate in casual conversations with high backchannel density. As expected, we find that acoustic detectors fire over-eagerly on silences, saturating recall but blowing past the false-positive budget. Semantic systems on the other hand are more selective but slower, staying in budget by waiting for more evidence.

VAP is the strongest model on both the EOT and INT tracks, but no system is simultaneously fast, high-recall, and low on false positives. For reference, humans in the TurnBench corpus begin turn transfers a median 151 ms before the turn ends3. We hope TurnBench is a valuable resource for the community to create more effective spoken interaction systems.

Submissions

We invite the community to improve upon these results, and provide the dev-set scorer as a resource. Please email turnbench@sesame.com with your submission by including your predictions, organization name, and model name.

Collaborators