What happened this week
OpenAI put a full-duplex voice model on general availability at a flat price, and the research side produced a dozen full-duplex papers in a single window. The two are connected: the platform layer is now selling what the papers are still learning to measure.
GPT-Live 1 is generally available
On 10 September OpenAI made GPT-Live 1 generally available in the API. It is a full-duplex model, meaning it listens and speaks at the same time, and it is priced by the clock: $0.05 per minute of voice session, billed per second, with no rounding up to the minute. Backend reasoning and tools are charged separately.
GPT-Live handles the conversation and delegates anything that needs thinking or tools to a backend. That backend can be an OpenAI Responses model, or, under what the docs call client delegation, any model, agent harness or service you run yourself. Concretely, the part of the system that decides when to talk is OpenAI's, and the part that decides what is true can be yours. The docs are explicit that interrupting the voice does not cancel backend work. The Realtime API stays in the catalogue, priced separately by audio token, so this is a second product rather than a replacement.
For scale, the model page lists rate limits in concurrent sessions rather than tokens, from 25 at Tier 1 to 500 at Tier 5, and its knowledge cutoff is 31 July 2025.
Twelve papers on the same problem
The count of full-duplex and turn-taking papers, which ran 0, 5, 2, 3, 2 across the summer, is twelve this window.
DuplexJail finds a safety hole that only exists because the model can be talked over. A fixed spoken prompt, delivered through the user channel while the model is mid-answer, raises attack success on AdvBench from single digits to 40.3% for PersonaPlex and 48.7% for PersonaPlex-RL, gains of 33.8 and 39.3 points. The prompt does not depend on the request. Safety alignment was tuned on turns that wait; interruption is a different distribution.
Less can be More asks what actually drives end-of-turn detection and answers with an ablation: acoustic plus prosodic features give utterance F1 of 0.93 with 7.8% false alarms at 400 ms median latency, and adding text makes premature detections worse without improving accuracy. Intonation and silence carry the signal; semantic completeness does not. That lines up with last week's TurnBench board, where the semantic systems were the slow, low-recall ones.
Duplex Cue names a third option beyond keep-talking and stop. A human who is corrected mid-sentence usually keeps going and folds the correction in. The paper builds 300 human-confirmed cues from unscripted conversation and compares recorded human responses against PersonaPlex continuations, scoring whether the model adapts within the turn or only yields.
MP-Bench evaluates voice agents as participants in multi-party conversation rather than dyads. SteerDuplex is a Moshi-based model plus SteerBench, 390 prompts and 1,067 human annotations for whether a duplex model can be told to change tone or pace mid-conversation. Self-Listening feeds the model's own played audio back as input, so when it is interrupted it knows what the user actually heard rather than what it intended to say. TASTE2 extends the text-aligned tokenisation line into a streaming dialogue stack with 98.2% of the text model's accuracy retained. DuplexDrama synthesises 2,000 hours of scenario dialogue with interruptions and backchannels and will release 800 hours. A French benchmark finds timing metrics transfer across languages and content metrics do not. TamilEOT fine-tunes Smart Turn v3 on 18,485 Tamil telephone boundaries. And the X2 lab, which shipped a turn detector and a streaming TTS in August, added X2Streaming-ASR, which learns when to commit a partial transcript rather than using a fixed chunk.
TurnBench itself gained one entry, Zero Runtime's Echo Omni at fourth with 0.839 recall, 0.071 false positives and 676 ms, and it is the first external submission with a Hugging Face repo, gated but self-serve. Disclosure: Fullduplex is published by oto, which supplied TurnBench's training split and has co-authors on the paper.
Three foundation-scale reports from China
Tencent's AuK went open source on 9 September under MIT, with a technical report a day earlier. It treats generation and editing as one job driven by natural-language instructions: 3.03 billion instruction-audio pairs and 1.95 million hours of supervision across speech generation, content editing, enhancement and separation, paralinguistic editing and acoustic editing, on a hybrid rectified-flow transformer. It is the most upvoted speech paper on Hugging Face this year and leads the TTS trending list.
Qwen-Audio-3.0-ASR is a mixture-of-experts recogniser trained on tens of millions of hours, covering 30 languages and 16 Chinese dialect varieties, built for production concerns like hotwords and disfluent speech. No weights were located this window. StepAudio 3 Gen goes the other way from AuK architecturally: a discrete autoregressive generator over residual vector quantisation tokens at 12.5 Hz, 16 codebooks of 2,048, covering TTS, voice design, singing, sound effects and music in one model.
Platforms
Deepgram added defer_until_eot on 8 September. Before it, the agent started composing a reply before the recogniser confirms the user finished, and while the audio was held back, function calls were dispatched the moment the model emitted them. So a booking or a hang-up could fire on a turn the user then continued. The flag holds the call until the turn is confirmed and discards it if the user resumes. Pipecat 1.9.0 added a latency breakdown that names the time no service reports, VAD silence and turn detection included, and 1.10.0 lets Gradium's server-side endpointing replace pipeline VAD. LiveKit 1.8.1 exposes Speechmatics end-of-turn configuration. ElevenLabs made Scribe v2 Medical generally available. And Orukeet, a Parakeet variant with half its temporal filters replaced by fixed Gabor kernels, beat the base model on 61 of 74 splits under CC BY-SA.