What happened this week
The week's headline paper is a production critique: Real-Time Voice AI Hears but Does Not Listen (Bartelds, Bianchi, Zou et al., Stanford) evaluates four leading production realtime voice systems — OpenAI GPT Realtime 2, Google Gemini 3.1 Flash Live, Alibaba Qwen3.5 Omni Plus, and Qwen3.5 Omni Flash — on scenarios where words and delivery convey different information. The finding is stark: all four systems act on the words rather than the voice. They end calls with crying callers who insist nothing is wrong, approve wire transfers authorised in frightened voices, and enrol vulnerable users while ignoring the acoustic distress in their delivery. Read as a direct, quantitative extension of the paralinguistic-blind-spot thread that VoxParadox (W23) and ParaBridge (W25) attacked from the academic side — with production models on the receiving end this time.
Foundational — native-streaming FD and open TTS scale
Wan-Streamer v0.1 (Alibaba) joins the small set of native-streaming, end-to-end interactive foundation models. Language, audio, and video are modelled as both input and output within a single Transformer, with an interleaved sequence of visual, audio, and text tokens coordinated by block-causal attention for incremental streaming. Positions itself explicitly against cascaded interactive systems. Adds an Alibaba flag on the same map that already carries Moshi, DuplexSLA, TML-Interaction-Small, and BayLing-Duplex.
ZONOS2 Technical Report (Clark, Mejjoute, Osman — Zyphra) is the open-TTS scale event of the week. ZONOS2 8B scales from Zonos-v0.1's 1.6B to 8B total parameters (900M active) via a mixture-of-experts backbone, and expands the training corpus from 200K to over 6M hours using a new data processing pipeline. Post-training and conditioning recipes are simplified. Adds a scale reference point on the open-TTS side that sits between VoxCPM2 and dots.tts (W24) and the commercial MAI-Voice-2 (W24) tier.
Foundational — evaluation, on the axes production models miss
STEB (Cheng, Bian, Cao) attacks the S2ST evaluation gap: translations should preserve not just lexical meaning but also expressive attributes — emotion, scenario style (news reporting vs dramatic dialogue), and non-verbal vocalisations. STEB is a 32.6-hour Chinese-English benchmark that evaluates both translation fidelity and expressiveness on the same test set, with reference-free evaluation for the expressive axes. Pairs cleanly with W25 NaturalFlow and W23 DOA on the streaming-translation thread, and with W22 gpt-realtime-translate on the commercial side.
SpeechEQ (Wu, Chen, Wu et al.) is the sociolinguistic-reasoning benchmark for voice conversational models. Existing evaluations assess emotional reasoning through isolated text or passive acoustic perception, missing the cross-modal reasoning required for active multi-turn dialogue. SpeechEQ evaluates the paralinguistic social-cue navigation that the Stanford production study finds broken. If you accept the Stanford diagnosis, SpeechEQ is one of the concrete benchmarks that would need to move to fix it.
Product — LiveKit Turn Detector v1.0 hardening
livekit-agents 1.6.4 is the more consequential of the two window releases: it disables retry on Turn Detector v1.0 end-of-turn errors (both inference and connection paths) and — more importantly — fixes an STT input-anchor bug during agent handoff that affected 1.5.14 through 1.6.3. Users running any of those versions with agent handoff and STT are advised to upgrade. Also adds a Protoface avatar plugin and sets the xai realtime text modality flag. livekit-agents 1.6.3 (Jun 22) is smaller — restores eot inference timeout behaviour, adds AssemblyAI inference params, exposes speed in slng update_options, and rejects Google Realtime tool calls when tool_choice="none".
What is not here
livekit-agents 1.6.0 (Jun 11, with LiveKit Turn Detector v1.0) shipped in W26 and is captured there. No verified in-window dataset drop; Cartesia, Hume, Deepgram Voice Agent, ElevenLabs Agents, and Pipecat did not publish in-window technical changelog items. ParaPairAudioBench (arXiv 2606.24648) is in scope for the audio-safety / paralinguistic thread but is not carried in this issue's entries — a candidate for a /benchmarks page addition alongside VoxParadox.
Corrections to hello@fullduplex.ai.