What happened this week
Turn-taking got a public benchmark, and the first thing it showed is that nothing on it works well yet.
Turn-taking is the part of conversation where you decide when to start talking. People do it constantly and mostly without error. Voice agents do it badly, which is why they cut you off mid-sentence, or sit in silence after you have clearly finished. Until now there was no shared test for it, so every vendor graded itself.
The benchmark
TurnBench, from Sesame AI Labs with Mundo AI, CMU and oto, pairs a 30-hour corpus of studio-recorded two-person conversation with a fixed scoring protocol. The corpus runs 154 dialogues across 106 actors, balanced over six conversation types, and every dialogue is labelled by three independent annotators with agreement measured at Fleiss's kappa 0.78. Ground truth is whatever two of the three agree on. Systems are scored on two events: end-of-turn, meaning the speaker has finished and the floor is open, and interruption, meaning someone took the floor while the other was still talking.
Disclosure. Fullduplex is published by oto. The TurnBench training split is oto's data, and two oto researchers are co-authors on the paper. What oto did not provide is the evaluation corpus or the scores, both of which come from Mundo AI and Sesame. Three of the eight entries below are oto's work or carry oto's name. Weight them accordingly.
What the leaderboard shows
Fourteen systems were measured. The result Sesame reports is that no system is simultaneously fast, high-recall, and low on false positives.
The failure modes split cleanly. Acoustic detectors, the kind that listen for silence, fire on every pause. OpenAI Realtime's server VAD catches 95.5% of real events, which is the best recall on the board, but its false-positive rate is 0.525 against a budget of 0.15, so it does not qualify at all. Concretely, that setting is a system that will interrupt you every time you stop to think. Semantic systems, which wait to understand what was said, are far more selective and much slower: OpenAI Realtime's semantic VAD holds false positives to 0.018 but recall collapses to 0.303, and Gemini 3.1 Live takes a median 1234ms to notice a turn ended.
The best overall is Voice Activity Projection, an older academic model, at 0.845 recall, 0.055 false positives and 368ms. For scale, the humans in the corpus begin their turn transfers a median 151ms before the previous turn actually ends. They are predicting where the turn ends rather than reacting once it has.
Specialising the data closes part of the gap
otoTurn is oto's own result, and it isolates one variable: training data. The team took the public SmartTurn 3.2 detector, fine-tuned it on full-duplex conversation carrying turn-taking labels, and changed nothing about the architecture.
On oto's spontaneous-conversation test set, false interruptions at a 300ms latency budget fell from 63.0% to 33.0%, a 47.6% relative reduction, with AUC moving 0.7238 to 0.8880. Across the 8,391 hold decisions in that test set, roughly 2,500 pauses that the base model cut off were preserved. The same fine-tune also improved the model on a task-oriented benchmark it was not built for, 35.2% to 21.6%.
Five detectors were measurable on both a task-oriented benchmark and spontaneous conversation, and every one of them lost between 23.9 and 34.9 percentage points moving from the first to the second. A detector validated on people booking appointments degrades sharply on people thinking out loud. The cause is the training distribution rather than any one vendor's implementation.
The training data itself is 104.94 hours across 420 conversations, hand-annotated per speaker channel with 17 event labels, which is finer than the two events TurnBench scores. It is gated and released under a non-commercial research licence.
Elsewhere
SpeechGym attacks a related gap in how voice agents are built. Today they are trained on text and then wrapped in speech recognition and synthesis, so nothing learns from what the audio actually sounded like. SpeechGym has two audio-native models talk to each other directly, over the unmodified tasks and tools of an existing text benchmark, so the only variable is the modality. Its finding: understanding audio does not give a model the ability to act on it.
A multi-party study from Aalborg goes the other direction, adding people and eyes. Using four-person conversations, it predicts whether a floor change will land as a clean gap or an overlap from gaze, speech and how close the speakers feel to each other, with deliberately interpretable models rather than a neural detector.
Two models shipped with licences worth reading before the weights. Breeze-TTS-2 reached 215 likes in its first week; its source is Apache-2.0 but the weights, anything derived from them, and self-hosted output are research and non-commercial only. NVIDIA's streaming diarization preview is stricter still: internal evaluation only, NVIDIA GPUs only, and you may not publish evaluation or test results without NVIDIA's written consent. In a week whose main event was an open leaderboard, that clause stands out.
Finally, OpenAI announced the deprecation of whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, all shutting down on 26 February 2027. If you are running transcription on any of them, gpt-live-transcribe and gpt-transcribe are the named replacements and you have six months.