What happened this week
Six new benchmarks for voice agents appeared in one week, each measuring something the others miss. One asks whether a pause or an overlap suited what the speaker meant, three whether an agent can tell who in a room is talking and to whom, and two whether any of it holds outside English. All six find a large gap.
Timing that depends on what the speaker meant
Most turn-taking scores use a fixed window: reply soon after a finished sentence and pass, talk over an unfinished one and fail. TACT argues that the right timing depends on intent. A pause after a question and a pause in the middle of a story are different events, and an overlap can be cooperative rather than rude. The benchmark covers 9,728 episodes and 73.2 hours from five two-person corpora, each labelled with a distribution over six intent classes, and scores timing against how people hand over the floor. The best of eleven systems scores 0.47; humans score 0.86. TACT's rankings agree with human judges at a Spearman correlation of 0.81, against 0.46 for the binary metrics it replaces.
The same authors also published an attack-and-defence paper. A full-duplex agent's microphone never closes, so an attacker can hide a perturbation under the masking threshold of ordinary speech, inaudible to a listener. White-box attacks hijack, mute or jailbreak a Moshi-style agent in up to 91.7% of trials; their defence cuts those to 8.3%, 11.2% and 9.1% at no inference cost. DuplexJail showed the audible version two weeks ago.
Rooms with more than two people
MSI-Bench, from Boson AI, puts a voice agent in 1,152 short multi-party scenes, half Mandarin and half English, where it must remember who said what and act for the right person. The strongest configuration passes every check on 66.8% of English scenes and 54.5% of Mandarin ones; the best open-weight setup manages 34.0% and 19.3%. Open models mostly fail at hearing who is speaking. Frontier systems hear it and still make the wrong speaker-specific decision.
Duplex-MPE gives five open full-duplex models a conversation among three or four people, with no transcript and no turn boundaries, and pairs each request in an explicitly addressed and an implicit form. A text model reading the transcript answers explicit requests 64.3 points more often than implicit ones. The speech systems show no significant difference: they do not register whether they were spoken to. VGBench keeps a spoken command word for word but moves it from the device's wearer to a bystander. Concretely, the friend next to you saying "send it" should not make your earbuds send anything. Six audio models mostly pick the right tool and rarely hold back; the best mute rate on switched commands is 14%.
Outside English
τ-Multilingual, from Sierra, Mercor and Princeton, runs 4,500 full-duplex calls through five voice-agent configurations in five more languages. Spanish, Portuguese and Hindi stay within 3.2 task-completion points of English. Korean drops 14.7 points and Mandarin 8.4, and they fail differently: Korean systems more often fail to respond, Mandarin systems interrupt more. Grok completes the most tasks and scores lowest on the quality of what it says, so the authors report task completion, interaction and generation quality separately.
NVIDIA and VoiceArena released IndicFDB, which extends Full-Duplex-Bench to ten Indian languages with 12,350 samples mined from roughly 50,000 hours of channel-separated conversation. Commercial APIs behave consistently across languages, but each is fast or tolerant of pauses, never both.
The delegation architecture ships as an app
The talk-in-front, think-in-back design that OpenAI and Google made generally available this month now has a detailed open write-up. The Qwen-Audio-Agent technical report splits the work three ways: a frontend agent holds the conversation, a backend agent does delegated tasks in its own context, and a runtime decides when results return. Interrupting the voice does not cancel the task, and a finished task is reported when the runtime schedules it rather than the instant it completes. On 134 in-house car-cockpit tasks, letting the frontend choose between acting itself and delegating reaches 91.04% success, against 72.39% when it never delegates and 80.60% when it always does. Version 2.0.0 shipped on 23 September as signed desktop apps for macOS and Windows. The companion Qwen-Audio-3.1-Realtime report cuts how often the model answers background speech on Full-Duplex-Bench v1.5 from 73.0% to 13.0%.
Mindlogic, Sakana AI and Meta added smaller pieces: Context Spanning feeds an outside LLM's answer into a duplex model as raw text, randomized guidance trains Sakana's KAME tandem model, and think-aloud fills reasoning silence with short progress remarks.
Open releases
Audio8 ASR Infinite, from Edge0, the team behind the Audio8 TTS previews, collected 1,896 likes in ten days. It is streaming Chinese and English recognition under Apache-2.0, with a rolling cache that keeps memory flat however long it runs and a semantic VAD meant to tell a thinking pause from a finished turn. It publishes no word error rates yet.
Nemotron 3 Diarization is out of preview. In August we noted that the preview's licence forbade publishing evaluation results without NVIDIA's consent. The full release on 23 September uses OpenMDW 1.1, which permits use without restriction, and community ports to five runtimes followed within a day.
YODAS v3 is about 1.1 million hours of 48 kHz stereo speech in 147 languages, crawled from Creative Commons YouTube uploads and released under CC BY 3.0; its authors call it the largest open speech dataset to date. Google made Gemini 3.8 Flash TTS and Flash-Lite TTS generally available on 22 September, at launch prices that double on 1 January.
TurnBench
Since last week's issue, Ooma's turn detector entered TurnBench in second place: 0.951 recall, 0.082 false-positive rate, 388 ms. That is within 20 ms of VAP, the fastest qualifying system, with 0.106 more recall, though its false-positive rate is half again as high. We found no announcement from Ooma. Disclosure: Fullduplex is published by oto, which supplied TurnBench's training split and has co-authors on the paper.