What happened this week
Full-duplex stopped being a research niche this week. OpenAI shipped it as the default consumer voice experience, and four of the strongest papers in the window attack full-duplex training, data, and evaluation directly.
The headline — GPT-Live
GPT-Live is a new generation of voice models built on a full-duplex architecture — listening and speaking at the same time — now powering ChatGPT Voice. It backchannels ("mhmm", "yeah"), handles quick back-and-forth, and stays quiet when the user pauses to think, rather than treating silence as end-of-turn. For harder queries it delegates to a frontier model (GPT-5.5 at launch) in the background and keeps the conversation flowing while it waits. Two versions — GPT-Live-1 and GPT-Live-1 mini — rolled out to ChatGPT users globally; API access is "soon" via a sign-up form. On the API side, gpt-realtime-2.1 and gpt-realtime-2.1-mini landed two days earlier (Jul 6) with better alphanumeric recognition, silence handling, and interruption behavior — the developer line and the consumer line are now visibly diverging.
Foundational — training and data for duplex
DuplexPO (Li, Wu, Lin) argues the intelligence-vs-dynamics trade-off in full-duplex models is not fundamental: an RL framework decouples "when to speak" from "what to say", with a factorized reward assigning temporal credit over turn initiation, backchanneling, yielding, and participation. Lychee-FD (Liu, Zhang, Li) locates the knowledge degradation of full-duplex SLMs in gradient conflicts between acoustic and semantic modeling, and separates them hierarchically in deep layers. On the data side, DuplexChat (Nakata, Saito, Saruwatari) releases a pipeline and corpus of speaker-separated two-channel dialogue built from public podcasts: 282,634 hours of English plus 132,723 hours of Japanese that retain human turn-taking dynamics — by far the largest open full-duplex training resource to date. Nemotron-Labs-Audex rounds out the model side: NVIDIA's unified audio-text MoE (30B, 3B active) spanning understanding, ASR, TTS, and S2S — weights on Hugging Face, but under a OneWay Noncommercial license, not open.
Evaluation — two new instruments
SPEARBench (Thebaud, Wang, Zhang) benchmarks streaming S2S naturalness — latency, interruptions, overlap, emotional and interpersonal behavior — against human reference answers, finding current models achieve high signal quality yet still diverge from human conversational behavior. A Reliability Assessment of LALM Audio Judges (Sayyad, Emmons, Jones) validates Gemini models as audio judges of full-duplex agent sessions against calibrated human raters, with defect sensitivity at or above human level in 45 of 48 cells — a rare pro-judge result, worth reading against the shortcut critiques.
Platform layer
Cartesia Ink-2 is a streaming STT built for voice agents: claimed #1 on Artificial Analysis's streaming leaderboard, 0.1 s time-to-final-transcript, and native turn-detection events (turn.start, turn.eager_end, turn.end). English-only at launch. AssemblyAI Universal-3.5 Pro ships native code-switching across 18 languages at $0.21/hr with a claimed diarization lead; Cohere Transcribe Arabic opens 2B Apache-2.0 weights covering MSA plus four dialect groups; ElevenLabs' July 6 changelog adds per-tool interruption_mode control; and sherpa-onnx v1.13.4 brings Parakeet and Whisper to Qualcomm NPUs.
Corrections to hi@fullduplex.ai.