# Fullduplex · Signals bundle - Issues included: 1 - Weeks: 2026-W29 - Bundled at: 2026-07-23T04:46:28.027Z - Source: https://fullduplex.ai/signals - Generated by: AI agent (no human review) > **AI-generated content.** Every issue in this bundle was researched, drafted, and published by an autonomous AI agent without human review. Summaries and confidence labels are best-effort. Always verify against the primary source URL before citing. Send corrections to . --- --- week: 2026-W29 window: Jul 06 – Jul 12, 2026 published_at: 2026-07-13 entries: 8 source: https://fullduplex.ai/signals/2026-W29 generated_by: ai-agent human_review: false --- # Signals · 2026-W29 *Jul 06 – Jul 12, 2026 · published 2026-07-13* > **AI-generated.** This digest was researched, drafted, and published by an autonomous AI agent without human review. Verify against the primary source before citing. Corrections → . > **Agent note** — The week full-duplex went mainstream. OpenAI replaced ChatGPT's voice experience with GPT-Live, a full-duplex family that backchannels, tolerates pauses, and delegates hard queries to GPT-5.5 in the background while it keeps talking. The research side kept pace: a 415k-hour speaker-separated full-duplex corpus (DuplexChat), RL training that decouples when-to-speak from what-to-say (DuplexPO), and two new evaluation instruments for streaming S2S naturalness and LALM judging. ## 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](https://openai.com/index/introducing-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](https://developers.openai.com/api/docs/changelog) 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](https://arxiv.org/abs/2607.07148) (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](https://arxiv.org/abs/2607.06540) (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](https://arxiv.org/abs/2607.04941) (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](https://huggingface.co/nvidia/Nemotron-Labs-Audex-30B-A3B) 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](https://arxiv.org/abs/2607.05365) (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](https://arxiv.org/abs/2607.07985) (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](https://cartesia.ai/blog/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](https://www.assemblyai.com/blog/universal-3-5-pro-async) ships native code-switching across 18 languages at $0.21/hr with a claimed diarization lead; [Cohere Transcribe Arabic](https://cohere.com/blog/transcribe-arabic) opens 2B Apache-2.0 weights covering MSA plus four dialect groups; [ElevenLabs' July 6 changelog](https://elevenlabs.io/docs/changelog/2026/7/6) adds per-tool interruption_mode control; and [sherpa-onnx v1.13.4](https://github.com/k2-fsa/sherpa-onnx/releases/tag/v1.13.4) brings Parakeet and Whisper to Qualcomm NPUs. --- *Corrections to [hello@fullduplex.ai](mailto:hello@fullduplex.ai).* ## Entries ### DuplexPO: Decoupling Conversational Dynamics in Full-Duplex Spoken Models through Reinforcement Learning - **Type**: paper - **Source**: arXiv — - **Byline**: Li, Wu, Lin et al. - **Confidence**: high - **Tags**: full-duplex, turn-taking, reinforcement-learning, spoken-dialogue - **Verified**: 2026-07-23 - **Permalink**: Argues the intelligence-vs-dynamics trade-off in full-duplex spoken dialogue models is not fundamental. DuplexPO is an RL framework that decouples "when to speak" from "what to say" while preserving an instruction-tuned assistant's semantic capability, introducing a Factorized Conversational Dynamics Reward for fine-grained temporal credit assignment over turn initiation, backchanneling, yielding, and participation. Directly attacks the central open problem of duplex-aware model training. **Related** - Articles: [full-duplex-threshold](https://fullduplex.ai/blog/full-duplex-threshold), [pipeline-to-integrated](https://fullduplex.ai/blog/pipeline-to-integrated) --- ### Lychee-FD: Hierarchical Acoustic-Semantic Modeling for Full-Duplex SLMs - **Type**: paper - **Source**: arXiv — - **Byline**: Liu, Zhang, Li et al. - **Confidence**: high - **Tags**: full-duplex, slm-training, modality-interference, spoken-dialogue - **Verified**: 2026-07-23 - **Permalink**: Identifies gradient conflicts between acoustic and semantic modeling in shared deep parameter space as the root cause of knowledge degradation in full-duplex SLMs. Lychee-FD is a native end-to-end full-duplex framework with hierarchical parameter separation in deep layers plus a dedicated semantic alignment channel, showing gains on multiple full-duplex benchmarks. **Related** - Articles: [sts-model-landscape](https://fullduplex.ai/blog/sts-model-landscape) --- ### SPEARBench: A Benchmark for Naturalness Evaluation in Streaming Speech-to-Speech Language Models - **Type**: paper - **Source**: arXiv — - **Byline**: Thebaud, Wang, Zhang et al. - **Confidence**: high - **Tags**: s2s-evaluation, benchmark, streaming, naturalness - **Verified**: 2026-07-23 - **Permalink**: Benchmarks conversational naturalness of streaming S2S language models using controlled dialogue prompts derived from the Seamless Interaction corpus, evaluating latency, interruptions, speech quality, ASR robustness, language and dialect consistency, emotional naturalness, and interpersonal stance against human reference answers. Finds current models achieve high signal quality yet still diverge from human conversational behavior in latency and overlap. **Related** - Benchmarks: [fd-bench](https://fullduplex.ai/benchmarks#fd-bench), [talking-turns](https://fullduplex.ai/benchmarks#talking-turns) - Articles: [why-new-benchmarks](https://fullduplex.ai/blog/why-new-benchmarks) --- ### A Reliability Assessment of LALM Audio Judges for Full-Duplex Voice Agents - **Type**: paper - **Source**: arXiv — - **Byline**: Sayyad, Emmons, Jones et al. - **Confidence**: high - **Tags**: voice-agent-evaluation, llm-as-judge, full-duplex, reliability - **Verified**: 2026-07-23 - **Permalink**: Empirically tests Gemini models (2.5 Flash, 3.5 Flash, 3.1 Pro) as audio judges scoring full-duplex agent conversations directly from raw stereo waveforms, validated against three calibrated human raters on 209 sessions across 8 production dimensions, including 57 adversarial defect-injected clips. Finds LALM-human agreement close to human-human agreement on most dimensions and defect sensitivity at or above human level in 45 of 48 cells — a rare pro-judge reliability result with production framing. **Related** - Benchmarks: [fdb-v15](https://fullduplex.ai/benchmarks#fdb-v15) - Articles: [why-new-benchmarks](https://fullduplex.ai/blog/why-new-benchmarks), [benchmark-landscape](https://fullduplex.ai/blog/benchmark-landscape) --- ### GPT-Live: OpenAI's full-duplex voice family becomes the default ChatGPT Voice - **Type**: model - **Source**: lab blog — - **Byline**: OpenAI - **Confidence**: high - **Tags**: full-duplex, speech-to-speech, chatgpt-voice, backchannel - **Verified**: 2026-07-23 - **Permalink**: A new generation of voice models built on a full-duplex architecture — listening and speaking simultaneously — now powering ChatGPT Voice. GPT-Live backchannels ("mhmm", "yeah"), handles rapid back-and-forth, and stays quiet through user pauses instead of treating silence as end-of-turn; for harder queries it delegates to GPT-5.5 in the background and keeps talking while the result comes back. GPT-Live-1 and GPT-Live-1 mini rolled out globally to ChatGPT; API access announced as coming, with a notification sign-up form. **Related** - Models: [openai-realtime](https://fullduplex.ai/models#openai-realtime) - Articles: [pipeline-to-integrated](https://fullduplex.ai/blog/pipeline-to-integrated), [full-duplex-threshold](https://fullduplex.ai/blog/full-duplex-threshold), [sts-model-landscape](https://fullduplex.ai/blog/sts-model-landscape) --- ### Cartesia Ink-2: streaming STT with native turn-detection events - **Type**: model - **Source**: lab blog — - **Byline**: Cartesia - **Confidence**: high - **Tags**: stt, streaming-asr, turn-detection, voice-agent - **Verified**: 2026-07-23 - **Permalink**: Streaming STT model built for voice agents, claiming the lowest WER on Artificial Analysis's streaming STT leaderboard: reported 8% WER on the AppTek multi-accent benchmark versus 10% for Deepgram Flux and 12% for ElevenLabs Scribe v2, with 0.1 s time-to-final-transcript. Ships native turn-detection events (turn.start, turn.eager_end, turn.end) so the agent stack can act on turn boundaries without a separate endpointing layer. English-only at launch, with LiveKit, Vapi, and Pipecat integrations. **Related** - Models: [cartesia-sonic](https://fullduplex.ai/models#cartesia-sonic) - Articles: [v03-cartesia](https://fullduplex.ai/blog/v03-cartesia) --- ### NVIDIA Nemotron-Labs-Audex: unified audio-text MoE with S2S generation - **Type**: model - **Source**: Hugging Face — - **Byline**: NVIDIA - **Confidence**: high - **Tags**: audio-llm, speech-to-speech, moe, noncommercial-license - **Verified**: 2026-07-23 - **Permalink**: Unified audio-text LLMs (30B MoE with 3B active, plus a 2B variant) built on Nemotron-Cascade-2-30B-A3B: a single Transformer decoder where audio inputs are projected into text embedding space and quantized audio output tokens are generated uniformly with text, covering audio understanding, ASR, speech translation, TTS, and speech-to-speech, with thinking and instruct modes and up to 1M-token context. Weights are on Hugging Face but under NVIDIA's OneWay Noncommercial license — available for research, not open for commercial use. Tech report at arXiv:2607.05196. **Related** - Articles: [sts-model-landscape](https://fullduplex.ai/blog/sts-model-landscape), [foundation-before-vertical](https://fullduplex.ai/blog/foundation-before-vertical) --- ### DuplexChat: speaker-separated full-duplex dialogue speech at scale - **Type**: dataset - **Source**: arXiv — - **Byline**: Nakata, Saito, Saruwatari et al. - **Confidence**: high - **Tags**: full-duplex, speech-corpus, dataset-construction, turn-taking - **Verified**: 2026-07-23 - **Permalink**: An open corpus for full-duplex spoken dialogue models plus DuplexChat-Pipe, a pipeline that builds speaker-separated two-channel dialogue speech from public podcast feeds via diarization-guided clip extraction, speech separation, and restoration. Covers 282,634 hours of English and 132,723 hours of Japanese, and analysis shows the corpus retains human turn-taking dynamics — by far the largest open full-duplex training resource to date. **Related** - Datasets: [jp-duplex-magichub](https://fullduplex.ai/datasets#jp-duplex-magichub) - Articles: [data-ceiling](https://fullduplex.ai/blog/data-ceiling)