What happened this week
Two weeks ago OpenAI shipped a full-duplex model that talks and hands the thinking to a backend. This week Google, Ant Group and NVIDIA all did the same thing, and the shape now has a name in three separate papers: a duplex frontend, a delegation token, and a backend that reasons and calls tools while the voice keeps going.
Google prices full-duplex below OpenAI
On 15 September Google made Gemini 3.8 Live generally available in the Live API, alongside an Extended Thinking variant that reasons in the background while narrating progress. The headline feature is asynchronous function calling: tool calls run while audio keeps streaming to the user. Pricing is $0.005 per minute of audio in and $0.018 per minute out. GPT-Live 1, which went GA the week before, charges a flat $0.05 per session minute with backend usage billed on top. Google says Extended Thinking ranks first on Artificial Analysis' speech-to-speech leaderboard; that is Google's claim, not ours.
Ant Group open-sources the same shape
Realtime-Venus, from Ant Group with Tsinghua, is two 9B models under Apache-2.0: an audio-visual one that watches and listens continuously, and an audio-only one for spoken interaction. Both are adapted from MiniCPM-o 4.5 and both do what the closed systems do: a shared causal timeline carries user input, model output and delegation events, and a harness executes tasks asynchronously while the foreground keeps talking. On Full-Duplex-Bench v1.5 the audio model responds to 75% of user interruptions and keeps talking through 97% of backchannels, 88% of speech directed at someone else, and 86% of background speech, which the paper says beats Gemini 3.1 Live and GPT-4o on all three. The weights landed on 16 September. It is the first open release we have seen with the delegation architecture built in.
NVIDIA writes the architecture down
Two NVIDIA papers describe the pattern as a design choice. The frontend-backend paper trains a duplex speech-to-text frontend to emit a delegation token and forward the streaming transcript to a text LLM, which calls tools; results come back through a prefill-and-repeat step and out via streaming TTS. Tool-call recall is 92 to 97%, and the frontend rejects irrelevant calls 81.2% of the time. With Qwen3-235B as the backend it is competitive on Full-Duplex-Bench v3 and beats GPT-realtime-mini and Qwen3-Omni-30B on EVA-Bench. Concretely: the model that decides when to speak is small and fixed, and you can swap the model that decides what to do.
The NemotronLabs VoiceChat paper is the system report for the 11B model NVIDIA put on Hugging Face in late July, whose TTS component we covered in August. It runs parallel output streams for agent text and structured function calls, with an RNN-T branch for user transcription. On Full-Duplex-Bench 1.0 it has the lowest pause-handling takeover rate among open-weight systems and takes over after 100% of user interruptions; on FDB 1.5 it resumes after backchannels 93% of the time. Tool-selection F1 on FDB 3.0 is 82.5%, and the authors say argument accuracy and end-to-end execution still need work. A third paper from the same group adds a lightweight ASR head to a duplex model and gets 10.21% streaming WER on the Open ASR sets without touching turn-taking.
StepAudio 3 Realtime, a 90-author report from StepFun, has the same loop under different names: Deep Perception, Seamless Duplex, Think-While-Speaking, and an integrated voice agent for asynchronous tools. It reports 98.9 on Artificial Analysis' Full-Duplex Bench and 56.0% task success on tau-Voice. The v1 was submitted on 12 September, inside last week's window, and missed there; no weights were located this week either.
Two ways full-duplex models fail the room
Full-Duplex Speech Models Take the Floor When Asked, Not When Needed builds context-matched monologues where only the trigger sentence varies, then asks whether five model families speak up when someone states a false fact or describes a hazard. They do not. Being addressed and silence are reliable triggers; false facts and hazards are not. Even when handed the floor, only 14 to 15% of replies to a false claim challenge it, and 4 to 7% of hazard replies warn. The models can hold a turn but cannot yet decide that one is needed.
ECHO, a Chinese benchmark from the group behind DuplexDrama, pairs examples with the same overlap transcript but different preceding context, so that one needs the model to yield and the other needs it to keep talking. Scoring requires both halves right, so a model that always yields scores zero. Most systems tested are biased toward yielding and do well on interruptions and badly on backchannels. Interruption-only evaluation overstates how reliable turn-taking is.
TurnBench, week four
The board is at 20 entries. Two external submissions arrived: Vox Maru v1 is first at 0.960 recall, 0.072 false-positive rate and 548 ms, which is both higher recall and faster than Sparrow-2, and SPX-AI MicroCF is third at 0.906, 0.079 and 678 ms. Neither company has a public page we could find. The repository also added GPT-Live 1 as a baseline on 14 September, with test recordings still in progress, so OpenAI's full-duplex model will appear on the board once they finish. Disclosure: Fullduplex is published by oto, which supplied TurnBench's training split and has co-authors on the paper.
Smaller, and a correction
Parakeet Redux from Moondream is NVIDIA's Parakeet 0.6B with every encoder weight forced to -1, 0 or +1. It shrinks from 1.2 GB to 178 MB, runs at 113 times real time on eight CPU cores, and stays within 0.3 WER of the original on English while beating it on 25-language FLEURS. It is worse under background noise, 9.04 against 6.72. CC BY 4.0. Also missed last week: Confucius4-R2T2 from NetEase Youdao, an append-only streaming recogniser on Qwen3-ASR that never revises committed words, now first on the ASR trending list.
Deepgram shipped ForceEndTurn in its SDKs and a Nova-3 model for drug names, then on 18 September published a correction: the Browser Agent SDK never had client-side voice activity detection, the documentation that said it did was wrong, and barge-in is server-side. ElevenLabs added call queueing with hold audio for agents at their concurrency limit.