Nuance Labs, Founded by Ex-Apple Researchers, Launches Nuplex 1.0 Full-Duplex Audiovisual Model and Emma Persona: Reacts With Her Face While You Talk, Matches Human Reply Timing 89% of the Time per the Company, Beta API Opens

Background and Pain Points

Most voice assistants and digital humans are half-duplex: they wait for the user to finish, then answer, and while you talk the on-screen face either stares blankly or loops a canned animation. Real people nod, frown and use their eyes to signal it is your turn. Nuance Labs, co-founded by former Apple researcher Fangchang Ma (MIT PhD), argues that what digital humans lack is not a more realistic face but the right timing.

Architecture Highlights

Nuplex is a full-duplex audiovisual foundation model: it takes in the user's voice and camera feed while generating its own voice and facial video, without waiting for the user to stop. Gaze is treated as a turn-taking signal, so Emma visibly reacts while listening. Her face and voice are licensed from a real actress who helped record how she actually listens and reacts, and her personality is designed to hold opinions and disagree rather than flatter. Reasoning still comes from a language model; Nuplex can work with any LLM, but Nuance serves one itself during the beta. Parameter counts and training data size were not disclosed.

Benchmarks

These are company figures. Nuance took 1,785 questions from its own recordings of real conversations and compared when each system started replying with when the original human answered. Nuplex overlapped 89% of the time; the next closest system reached 35%. Nuance says open models such as PersonaPlex and Moshi are tuned to answer fast, while GPT-Live and Gemini Live answer later, and those two were measured through public APIs that include network time. The company also admits there is no benchmark yet that captures the real experience of audiovisual full-duplex chat. No independent evaluation has been published.

Getting Started

Visit emma.nuancelabs.ai to video chat with no account; headphones help because background noise hurts turn-taking. Developers can request the Nuplex Beta API on Nuance's platform, brief a persona with a prompt, and let it hand longer tasks to their own tools mid-conversation. Known gaps: English only, occasional interruptions or long pauses, lip sync and gaze glitches on poor connections, and early-stage visual understanding. It suits interview practice, customer support and coaching; healthcare and education are not available yet.