What happened this week
Last week's issue said no system passed TurnBench. This week two vendors submitted, one of them now leads the board, and neither of them is open.
Disclosure, since TurnBench runs through this issue. Fullduplex is published by oto, which provided the benchmark's training split and has two co-authors on the paper. The evaluation corpus is Mundo AI's and the scoring is Sesame's. Nothing below was scored by oto.
The leaderboard moved
TurnBench went from 14 systems to 17. The three additions carry a dagger, which the site defines as an external submission, as opposed to the baselines Sesame ran itself. On 31 August the repository also added a rule: submissions must be the submitter's own system's output.
Tavus Sparrow-2 is now first at 0.930 end-of-turn recall, 0.084 false-positive rate and 615 ms median latency. Neosapience's Nunchi v1 dual-channel is second at 0.851, 0.040 and 610 ms. Voice Activity Projection, last week's leader, drops to third at 0.845, 0.055 and 368 ms.
The three numbers only make sense together. Recall is how many real turn ends the system caught; false-positive rate is how often it fired during a pause that was not a turn end, with 0.15 the cut-off for qualifying at all; latency is how long after the event it noticed. Sparrow-2 catches more turn ends than anything else and stays inside the false-positive budget, but it takes 615 ms to do it, against 368 ms for VAP. Nunchi is the cleanest on false positives of the three and gives up recall for it. Last week's finding, that no system is fast, high-recall and low on false positives at once, still holds. The trade has moved.
Two caveats. Tavus's own post reports dev-split numbers, 92.4% end-of-turn recall against a runner-up at 84.1%, and says the private test set was pending; the leaderboard figure is the returned test score. And Sparrow-2 is a hosted API. There are no weights, so the number cannot be reproduced outside Tavus.
Speaker-attributed ASR goes streaming
VibeVoice-ASR-Streaming is Microsoft's answer to a specific gap: models that transcribe who said what have mostly worked offline, which rules them out for live agents. The new 7B and 1.5B models interleave fixed-size audio chunks, a small lookahead and the text so far, so they emit speaker-labelled transcript as speech arrives, with no separate diarization stage. The technical report claims the lowest average error rate across five evaluation sets and the best or tied-best speaker attribution on 12 of 13 settings. Weights are MIT. Ten languages, custom hotwords for names and jargon, and it reached the top of the Hugging Face ASR trending list within days.
Full-duplex synthesis and full-duplex data
Text-Audiobox, from FAIR at Meta, is a 3B-parameter diffusion transformer pretrained on 480,000 hours of speech that generates two-speaker dialogue as a single waveform, with turn-taking, backchannels and emotional dynamics modelled natively rather than stitched together. It is alignment-free, meaning it learns where words land in time through cross-attention instead of a forced aligner, and it operates on a 25 Hz latent over 48 kHz audio. Meta reports it approaches human recordings on short conversations. Comparisons are against Meta's own internal systems, and the paper does not mention a release.
A Kyoto University paper takes a cheaper route to the same behaviour. Neural finite-state-machine dialogue systems serialise turn-taking decisions and words onto one token stream, but they have been trained on synthetic text, which cannot reproduce the timing of real conversation. The fix here is to learn timing from recorded human-to-human dialogue and semantics from separate human-to-agent text, with a rule-based transformation that turns real recordings into training tapes without an LLM annotator. Code and model are released.
Prosody, again
Two issues ago Hear2Act showed a model could have the audio and still not act on it. A CMU study explains why. Across Whisper-large-v2, Qwen2-Audio, Qwen2.5-Omni and Chroma, speaking style is strongly encoded in the top third of the audio encoder and then degraded before the output. The projector reshapes the representation without deleting it; the decoder is where the loss happens. A better encoder would not fix it.
From the other end of the pipeline, GN Group, the company behind Jabra, measured what speech enhancement does to a voice agent's understanding rather than to its audio quality. Their metric, Output Divergence Rate, counts how often enhancement changes an LLM's intent classification relative to clean speech. MetricGAN+ more than doubles it against leaving the noise alone, 0.318 versus 0.135, while improving PESQ. Concretely: the audio sounds better and the agent understands it worse. Unmitigated echo reaches 0.836 by making the model transcribe the wrong speaker, a failure word error rate does not see.
Platforms
Vapi's week of 31 August exposes the Soniox transcriber's endpoint sensitivity and latency adjustment as tunable settings, which is the same recall-versus-latency knob TurnBench measures, handed to the developer. LiveKit Agents 1.8.0 surfaces AssemblyAI's end-of-turn confidence on transcript metadata and adds its universal-3-6-pro streaming model. ESPnet 202609 raises the floor to Python 3.12 and lands OpenBEATs. Deepgram added Kazakh to Nova-3.
And for anyone deploying in care settings, a paper announced this week but submitted just before the window, When Patients Cut In, tests what a cascaded clinical agent loses when the patient interrupts it mid-sentence. Clinical benchmarks almost universally assume the patient waits.