Nuplex

Nuplex is the full-duplex audiovisual foundation model from Seattle-based Nuance Labs: it takes audio and video in and produces audio and video out, instead of chaining speech-to-text, a language model, and text-to-speech. The company was founded by former Apple researchers Fangchang Ma (CEO), Edward Zhang, and Karren Yang, whose earlier work covered machine perception and human reconstruction. The company site is https://www.nuancelabs.ai/ .

In this radar window Nuance Labs also launched Emma, a demo persona running on Nuplex 1.0 that listens and reacts while you are still speaking and uses visual cues to show when it is her turn. The try-it page is https://app.nuancelabs.ai/try . Public write-ups say Emma’s face, voice, and conversational reactions come from recordings made with an actress and then fed into Nuplex. The radar excerpt also noted that the Nuplex API is in beta.

Highlights:

• Company site: https://www.nuancelabs.ai/ ; live demo: https://app.nuancelabs.ai/try ;
• The architectural bet is one end-to-end full-duplex audiovisual model rather than an ASR → LLM → TTS → lip-sync pipeline;
• On 14 September 2026 Nuance Labs announced a $50 million Series A led by Lightspeed, with Accel and South Park Commons returning and NVIDIA and Define Ventures among the new investors;
• Founder @fangchangma’s launch post drew about 757k impressions in this window. Rohan Paul reported that Emma’s reply timing matched real human speakers 89% of the time—that figure comes from secondary coverage, not an evaluation we reproduced.

Why it matters

Most conversational avatars still listen, generate text, then speak and lip-sync. The seams show up as stiff turn-taking, missed interruptions, and dead air. Nuplex puts perception and generation in one full-duplex model so a face can react—and start speaking—while you are still talking. Emma turns that claim into a public demo rather than only a launch video, which is the practical way to judge whether a full-duplex audiovisual model actually feels usable.

The company context matters too: the founders came from Apple work on perception and reconstruction, and in September 2026 they raised a $50 million Series A led by Lightspeed with NVIDIA among the participants. For anyone building digital humans, realtime companions, or support avatars, this is a concrete example of the single audiovisual foundation-model bet, not just a manifesto.

Limits

Public pages do not publish latency numbers, a full benchmark table, or a price list that we could verify, so this entry does not invent Nuplex pricing or performance claims. The API is described as beta; availability, quotas, and commercial terms follow whatever Nuance Labs says at the time you check. Emma is one character built from an actress’s recordings, not proof that any avatar will match the same timing and mannerisms. The 89% timing figure is from secondary reporting. The September funding note described a later-in-2026 research preview and did not yet include a public demo, so that article and this launch are different stages. Read the product’s privacy and likeness/voice-data terms before trying the demo.

1. Read the company page at https://www.nuancelabs.ai/ , then try Emma at https://app.nuancelabs.ai/try ;
2. If you want to integrate it, watch Nuance Labs’ channels for Nuplex API access (the radar excerpt described the API as beta). Do not assume a public, stable price list already exists;
3. When you evaluate the demo, watch full-duplex timing—whether it keeps listening while you speak, whether expressions track the conversation, and whether turn-taking feels natural—not just still-frame fidelity;
4. Before any commercial or likeness/voice-capture use, check licensing, privacy, and regional availability. Treat unpublished latency, concurrency, or pricing figures as unknown until the company states them.