Voice and NPCs
Nuplex 1.0: a full-duplex AI face that reacts while you talk
Nuance Labs has launched Nuplex 1.0, a full-duplex audiovisual model, with a video persona called Emma and a beta API. It listens and talks at once, and nods, smiles and looks away while you are still speaking.
On 9 November, Seattle-based Nuance Labs launched Nuplex 1.0, which it calls a full-duplex audiovisual foundation model. “Nuplex” stands for Nuance Full-duplex: a system that listens and speaks at the same time, rather than waiting for you to finish.
The first persona built on it is Emma, a video character you can talk to in a browser. Nuance says she reacts while you are still speaking, with nods, a long blink, a squint or a glance away, and that her face and voice come from a licensed actress.
What it can do
- Face-to-face timing: Nuance measured when a person would reply in 1,785 questions from its own recordings of real conversations. It says Nuplex’s reply timing overlaps the original speaker’s 89% of the time, against 35% for the next closest system.
- Persona: Emma is designed to have opinions and to disagree, while saying plainly that she is an AI.
- Background work: through the API, a persona can hand longer tasks to your own tools while it keeps talking.
- Language models: Nuance says Nuplex works with any language model, but the beta runs one that Nuance serves itself.
Access, price and limits
Emma needs no account. The Nuplex Beta API, with a few personas, is at platform.nuancelabs.ai. The launch post gives no price, parameter count or open weights.
Both are deliberately restricted. Video from the API carries a visible watermark, you cannot upload a face or voice to clone, and custom faces are made by arrangement for one organisation. Nuance bars sexual content, “AI girlfriends or boyfriends”, healthcare and anything involving under-18s. Both products are for adults only. They are available in the US (except Colorado, Connecticut, Illinois and Texas), the EU, the UK and Canada (except Quebec).
Nuance lists its own weak spots: English only, worse turn-taking with background noise, shaky reading of camera input, occasional lip-sync and gaze glitches, and confident mistakes from the language model. The 89% figure is a vendor measurement. Nuance notes that GPT-Live and Gemini Live were reached through public APIs, so their timing includes network delay, and it admits that benchmarks in this area do not yet match the user experience.
Our take
For anyone building interactive worlds, the useful idea is that presence comes from timing and small facial cues, not from a bigger script. A character that nods mid-sentence feels attentive in a way a perfect answer does not.
The restrictions matter too. No cloning, a visible AI watermark and refusal to pose as a human are choices game studios will face as soon as characters get faces. This follows Boson’s Higgs Realtime Duplex. We explain why live timing matters for game characters in the end of the dialogue tree.