What happened this week
OpenAI opened outbound phone calling for GPT-Live-1 to every organization, and Retell put the same model on its phone agents the same day. Two open full-duplex models (models that can speak and take in speech at the same time) appeared on Hugging Face, and a new benchmark scores seven live voice systems from the same recorded user audio.
Benchmarks
InteractionBench, from Avera Labs, runs 221 items per system across thirteen scenarios. It grades GPT-Live, GPT-Realtime, Gemini Live, Moshi, PersonaPlex, Freeze-Omni and a turn-based cascade (separate speech recognition, text model and speech synthesis) in 1,519 runs. Among the six full-duplex systems, timing and spoken-task accuracy rise together (rank correlation ρ = 0.83). The cascade has the highest spoken-task score (38/50) and the second-lowest timing score. Groundedness probes check whether a model still uses facts from earlier in a conversation of more than twenty turns: hosted systems pass 67 to 88% of them, open full-duplex models 14 to 35%. Only Gemini Live whispers on most requests (8 of 10).
Timing is measured where the client hears each system, so hosted results include network delay. For GPT-Live, whose server reports its own clock, removing about 1.5 s of delay lowers its timing score from 86% to 56%. A timing score is only comparable when the measurement point is stated.
Voice-Native LLMs Hear Frustration but Do Not Listen to It, from New York University Abu Dhabi, crosses frustrated wording with a frustrated voice over 353 tasks and releases the 1,412 clips as VoxAffect. Seven of eight models show no detectable effect of the voice on accuracy. With neutral wording, all eight label the frustrated voice as frustrated in at most 3.1% of clips, although seven of them detect acted anger in 73 to 100% of clips and human listeners classify the voice correctly in 65.3% of judgments. Linear probes on the audio encoders of the two open-weight models decode the frustration, with paired accuracy 0.67 and 0.69. For a support agent, that means a caller whose tone sharpens while the words stay polite is treated like a calm caller.
Latency
Speculative Listening, from The Chinese University of Hong Kong, Shenzhen and Amphion, targets chunk-interleaved full-duplex models, which take in audio and produce output in fixed one-second chunks. While the model waits for audio, its compute sits idle; the method spends that time decoding the next chunk as if the incoming audio were silence, then checks the drafts when the real audio arrives. Across four models, the next audio chunk did not change the output in 95.3% of chunks, and 88% of changes began at the first output token. On the authors' eight-task TickTock-Bench, speak chunks run 2.6 to 2.8 times faster and tool-call chunks up to 23.4 times faster. The average delay before a response starts falls 73%, to 74 to 241 ms, which the paper sets against the 200 ms gap between human turns. It needs no training, and the output matches ordinary decoding.
Open full-duplex models
Sophon-Realtime v1, from adventists-ai, is a 32B open full-duplex model under Apache-2.0 with three voice back-ends (Tag, Flash and E2E). Self-reported: VoiceBench 82.8, Big Bench Audio 81.8%, and Full-Duplex-Bench v3 tool calling Pass@1 48% with a 100% response rate. The card says its own tool-calling figure is the best of several runs, while the other rows are official results. The Flash build reports 0.31 s to first audio and up to 40 concurrent users on one 80 GB GPU. A gateway in the repository speaks the OpenAI GPT-Live protocol, so a GPT-Live client can switch by changing the URL and key. No external validation found.
Awaken Full Duplex appeared under openmartdev, a namespace new to Hugging Face, with its code and project page under Avera Labs, the group behind InteractionBench. It converts the half-duplex Fun-Audio-Chat-8B into a full-duplex model without a voice activity detector (VAD), the separate component that usually decides when the user has stopped talking. Two supervised fine-tuning stages are followed by a short GRPO (group relative policy optimization) stage for backchannels, the "uh-huh" said without taking the turn. Self-reported: the backchannel pass rate rises from 1% to 63%, the model passes 61% of groundedness probes against 66% for its base, and it solves 90% of spoken logic puzzles. English only. Its comparison rows come from InteractionBench, so benchmark and model share an author.
Platforms
OpenAI's API changelog entry for November 2 makes outbound SIP calling through the Live API available to all organizations for gpt-live-1. SIP (Session Initiation Protocol) is the standard that sets up phone calls between networks, so a developer can now connect GPT-Live to a phone number through their own SIP provider. The same day, Retell made GPT-Live-1 available on its phone agents. GPT-Live-1 talks to the caller while a backend text model runs the prompt, tools and knowledge base. Retell restates OpenAI-reported figures: 86.2% of Tau3 voice tasks passed on the first attempt with GPT-6 Astra as the backend, against 45.7% for GPT-Realtime-2.1, and 0.798 s turn-taking latency on Full Duplex Bench against 1.41 s.
ElevenLabs' November 2 changelog makes Eleven Realtime available to every workspace, completing the rollout announced October 19. It also changes the Realtime API's turn events. speech_started and speech_stopped now fire once per committed user turn instead of on every pause, each manual commit transcribes only the speech since the previous commit, and idle sessions close after 3 minutes instead of 60 seconds. Code that drives barge-in from those events will receive fewer of them.
Also in the window
LiveKit Agents 1.8.11 (November 5) adds Rime's Prism speech-to-speech and gemini-3.5-live-translate-preview. It also changes turn detection so that late speech recognition finals arriving after the user resumes speaking join the resumed turn.
Coverage
When this issue was compiled, arXiv had listed submissions only through November 5, so papers from November 6 to 8 move to next week. No full-duplex video item was found in the window.