← Back to feed

Nuance Labs, a Seattle-based AI lab founded by former Apple PhD researchers, announced a $50 million Series A on September 14. Lightspeed Venture Partners led the round, returning from the seed stage. NVIDIA and Define Ventures are new investors; Accel and South Park Commons also participated, both returning from the company’s $10 million seed in 2025. Total announced funding is $60 million.

The company is building a foundation model for real-time human conversation and expression. Its technical target is full-duplex audiovisual: a single model that can simultaneously interpret and generate speech, facial expression, and body language without requiring the conversation to pause for each response.

The Technical Problem

Full-duplex audiovisual inference is harder than it looks. Current voice AI systems operate in half-duplex mode — they wait for the speaker to finish, process, and then respond. Real human conversation happens in parallel: both parties continuously read facial cues, adjust tone, and use backchannel signals (a nod, a brief “mm-hmm”) to maintain coherent exchange.

Building one model that handles all of this simultaneously requires solving several problems at once:

• Identity consistency: Speech output must match a target voice; visual output must match a target face across frames.
• Expression coherence: Facial expressions need to synchronize with speech content and emotional register in real time.
• Low latency: The loop from input to output needs to be short enough that the interaction feels conversational, not transactional.

Separating these into multiple specialized models is the standard approach. Nuance Labs is betting that a unified model — trained jointly on all three modalities — will generalize better and run faster because it can share representations across the signal types rather than running separate inference passes.

Why NVIDIA

NVIDIA’s participation is the signal that warrants attention in this round. At the seed stage, Nuance Labs drew Lightspeed, Accel, and South Park Commons — experienced AI infrastructure investors. NVIDIA does not typically invest in pre-product research labs at the Series A stage.

The exception to that pattern involves companies working on modalities that are computationally intensive in ways that extend NVIDIA’s hardware demand. Full-duplex audiovisual models — if they work — require real-time inference on video streams with sub-100ms latency budgets, at scale. That workload profile differs from batch language model inference in ways that favor NVIDIA’s Vera Rubin architecture, with its emphasis on sustained throughput under tight latency constraints.

NVIDIA’s strategic interest is not difficult to reconstruct: it backs companies that will need significant GPU capacity and whose technical direction validates a use case that existing infrastructure handles poorly.

The Team

Nuance Labs was founded by Fangchang Ma, who leads the company as CEO. The founding team came out of Apple, which has invested heavily in on-device multimodal inference over the past several years. The specific Apple background — on-device latency optimization, hardware-software co-design — maps directly onto the latency constraints of real-time full-duplex inference.

The company is hiring across pretraining infrastructure, reinforcement learning research, model optimization, inference, data infrastructure, and human evaluation. The breadth of the hiring list suggests they have not yet begun large-scale pretraining; the $50 million would fund the compute and team for a first full training run.

What Comes Next

Nuance Labs plans to open a public research preview later in 2026, with access through a waitlist. No product is live. The first meaningful test of the core technical claim — that one model can handle all three modalities in real time — will come from the preview.

The funding positions them alongside a small set of labs working on multimodal real-time AI: Hume AI, which raised substantial capital for empathic voice interfaces, and Tavus, which focuses on video generation for personalized communications. Nuance Labs is aiming at a harder version of the same problem, specifically the full-duplex real-time constraint that none of the current video generation approaches have solved.

Whether a unified audiovisual model is achievable at interactive latency remains the open question. The round is a bet that it is, and that getting there first is worth funding at this stage.