Nuance Labs unveils Nuplex 1.0: full-duplex audiovisual model matching human reply timing 89%

rohanpaul_ai · x · 2026-11-10

Nuance Labs released Nuplex 1.0, a full-duplex audiovisual foundation model, plus Emma, a persona built on it (emma.nuancelabs.ai) with a Beta API, after 18 months of research.
Reply timing: instead of latency benchmarks, the team measured when a human would reply — on 1,785 questions from real conversations, Nuplex's timing overlapped the original speaker 89% of the time vs. 35% for the next closest system. Open models like PersonaPlex and Moshi answer maximally fast to game benchmarks; GPT-Live and Gemini Live are slower, likely to avoid interruptions.
The company also emphasizes visual responses and gaze as turn-taking signals, arguing face-to-face naturalness matters more than latency scores.
A technical report is available; the post shares it.

More from Multimodal