Nuance Labs raises $50M to fix awkward AI avatars

Nuance Labs' pitch rests on a critique of how most AI avatars are put together: developers stitch a voice-to-text model, a large language model and a text-to-voice model into a chain, and for an avatar they add a fourth step to animate a face. Each handoff adds delay, and Ma argues that is why avatars look frozen or random while a person is speaking. His alternative is a single model that perceives and generates in one full-duplex loop, learning from how people behave during its conversations.

The $50 million Series A, led by Lightspeed Venture Partners with returning backers Accel and South Park Commons and new money from Nvidia Corp. and Define Ventures, is meant to accelerate that model work and add researchers. Lightspeed's Nnamdi Iregbulem framed the bet as a foundational layer for AI products, which suggests the company is being judged less on a finished product than on whether one model can hold a natural conversation.

That question is still open. Nuance has no launched product and describes its demo as a work in progress, with a first research preview targeted for later this year. Whether the single-model approach truly removes the awkwardness Ma describes — and whether it holds up in the sales, service, coaching and training scenarios he names — is likely to be the deciding factor for the next stage of the company's progress.