Why end-of-turn is hard
For an AI agent to hold a natural spoken conversation, it has to know when the other person has finished speaking. Existing systems make this work well in customer support, booking, and other structured flows, where the user completes a request and stops. Ordinary conversation has a different shape: people revise sentences, search for words, and continue after pauses lasting well over a second. A grammatically complete sentence may be the end of a turn, or the setup for what comes next.
Detectors validated on task-oriented traffic lose between 24 and 35 percentage points of reliability when the conversation turns spontaneous. otoTurn is built for that harder case.
The results
The otoSpeech evaluation tests long-form, free-flowing human conversation: 20.96 hours from 84 full-duplex, two-person conversations, with human-reviewed end-of-turn and floor-hold annotations. otoTurn leads every column.
otoSpeech EOT detection (2026)
| Model | AUC | AP | False cutoffs @ 300 ms | False cutoffs @ 600 ms |
|---|---|---|---|---|
| otoTurn | 0.8880 | 0.7226 | 33.0% | 18.0% |
| LiveKit v1 cloud | 0.8240 | 0.5995 | 44.8% | 24.3% |
| ultraVAD base | 0.7928 | 0.5662 | 51.6% | 27.8% |
| LiveKit v1-mini | 0.7282 | 0.4694 | 60.7% | 31.9% |
| SmartTurn 3.2 base | 0.7238 | 0.4665 | 63.0% | 33.1% |
| LiveKit text baseline | 0.5746 | 0.3500 | 78.1% | 38.6% |
| VAD timeout baseline | 0.5000 | 0.2746 | 80.5% | 39.9% |
False-cutoff rate: how often the detector interrupts a speaker who was not done. Lower is better. Each column reports the best policy under the stated mean end-of-turn latency budget.
One change: the training data
otoTurn is a public end-of-turn model, SmartTurn 3.2, fine-tuned on a dedicated full-duplex dataset with turn-taking labels. The architecture was left alone. False interruptions fell by nearly half, about 2,500 pauses survived that the base model cut off, and the same fine-tune improved on a task-oriented benchmark the data was not built for.
The training data comes from the same collection pipeline behind our public otoSpeech releases and the TurnBench training set built with Sesame: channel-separated recordings of real conversation, annotated by hand.
Coming next: otoEnhance
otoEnhance, a speech enhancement model trained on the same two-channel data, is in preparation. It cleans audio without flattening the prosody and turn-taking cues that models like otoTurn depend on. Write to us if you want early access when it ships.
Early access
Put otoTurn in your agent.
Tell us where turn-taking hurts in your product, and we will set you up with the model, the evaluation, or the data behind both.