What happened this week
Three papers put full-duplex models, which listen and speak at the same time, into open conversation and measured what they do. The common finding is about timing: the models are the only systems that hold back at a pause, but when two of them talk, the floor changes hands late.
Full-duplex models in conversation
People Making Conversation, from People Make Things and accepted at a NeurIPS 2026 workshop, places six systems at 427 pauses and 482 turn ends from 176 unscripted Seamless Interaction conversations and compares what each generates with what the person did. The hosted APIs (GPT Realtime, Gemini 3.1 Flash Live, Grok Voice) respond at almost every pause. Their latency, about 1 to 4 s with silence-based endpointers (the component that declares a turn finished) and about 8 s with semantic turn detection, decides whether the reply even lands inside the pause. PersonaPlex and Moshi, the two full-duplex models, are the only systems that regularly stay silent and the only ones that interrupt within a human reaction time. Memory goes the other way. Given the history as text, hosted systems recall 41 to 60% of conversation facts under an LLM judge; hearing it as audio, the full-duplex models recall 17 to 22%. The paper lists TurnBench among the detection benchmarks it compares itself with. Fullduplex is published by oto, which supplied TurnBench's training split and has co-authors on the TurnBench paper.
DyaFDB, from KAIST, argues that testing a full-duplex model against pre-recorded audio or a scripted examiner measures only one side of the exchange. Here two models converse directly under assigned roles with cooperative or conflicting goals, and an external judge scores both sides offline. Across four tasks, 140 scenarios and 7,560 recorded conversations with PersonaPlex, MiniCPM-o and Raon-SpeechChat, each model's behaviour keeps reshaping its partner's. When both sides compete for airtime, Raon holds 0.75 of the exclusive speaking time against MiniCPM-o and 0.62 against PersonaPlex. The scenarios and recording protocols are announced but not yet released.
Coupled but Late, from Duke University, runs two PersonaPlex-7B instances on a shared clock and compares them with the Switchboard telephone corpus. Their timing is coupled: re-pairing speakers across conversations destroys it. But the floor changes hands at a median of 400-560 ms against 137 ms for humans, and the last 120 ms of the partner's turn, where humans place a tenth of their transfers, holds 1% of the models'. Delaying one direction of the channel shifts the response one-for-one. The authors read this as a wait after the perceived end of a turn, not the prediction of the end that human timing needs. This matters for self-play data generation, where two models produce training conversations for each other: the late timing goes into the data.
Two training papers work on the same behaviour. HiPLEX, from Qualcomm AI Research and KAIST, splits a full-duplex text policy into a part that decides when to emit content and a part that decides what to emit during reinforcement learning. Across three Moshi seeds on Full-Duplex-Bench v1 it takes over less often at natural user pauses than GRPO. DiffuPlex, from the University of Seoul, predicts several future frames in one backbone call and reports 1.46x and 1.59x wall-clock speedups, with every measured call finishing within the 80 ms interaction interval.
Models and platforms
Amazon made Nova 2.5 Sonic generally available in Bedrock on October 5. AWS claims better reasoning, instruction following and tool-calling accuracy and "lower latency", with no numbers. It keeps voices in seven languages, controllable turn-taking, asynchronous tool calling and a 256K context window, at the same price as Nova 2 Sonic, in N. Virginia, Oregon, Stockholm and Tokyo. LiveKit Agents 1.8.5 added it the next day and fixed a bug on the way: its client matched the model id by substring, so the 2.5 id fell back to a legacy endpointingSensitivity field that Sonic 2 and later reject, and the session never started.
LiveKit Agent Simulations became generally available for LiveKit Cloud projects on October 8. A language-model caller plays out scenarios against an agent, and a judge grades the transcript pass or fail with a reason. A failed scenario returns a non-zero exit code, so it can fail a build. Audio mode adds background noise, interruptions and backchannels (the short "mm-hm" a listener gives without taking the turn) and covers cascaded and full-duplex agents. In practice a voice agent can now get a regression test in the same CI run as its code. The judge reads the conversation, though, and last week's VAmoS paper found that scores built from the conversation alone overstate agents.
Mistral released Voxtral Mini 4B Realtime Arabic on Hugging Face under Apache-2.0, with no launch post. It is a streaming recognizer for Modern Standard Arabic and dialects from the Maghreb, the Gulf, the Nile valley and the Levant, built with Morocco's digital-transition ministry (MTNRA). At a 480 ms transcription delay it averages 8.82% character error rate (CER) over seven benchmarks, two of them new, within 0.91 points of Mistral's offline Voxtral Transcribe Arabic. The figures are self-reported.
Data
SmoothConv and DuplexConv, from ASLP@NPU, Soul AI Lab and QualiaLabs, add 2,100 hours of Mandarin multi-party conversation with a synchronized track per participant: 100 human-verified hours with 3.33% overlap, and 2,000 automatically annotated hours. A turn-detection test set labels four states: complete, incomplete, backchannel and wait. The paper says the data is public; the GitHub repository states no licence.
Missed in W40
Gemini 3.8 Live with Live Avatar, published September 24 and available in Gemini Enterprise, belonged in W40. With the Live API response modality set to VIDEO, the model returns 24 FPS video whose expressions and mouth movements are synchronized to its own synthesized voice. The model card limits sessions to "a few minutes of continuous interaction, rather than extended hours"; behaviour while the user is speaking is not documented, and VideoFDB has not evaluated it.