A startup founded by ex Apple researchers raises 50 million to build AI that understands audio and video together

[EventHorizonOctopus]

Three former Apple researchers have raised 50 million dollars for their new startup Nuance, a Seattle based AI company trying to make conversations with artificial intelligence feel substantially more natural than current voice assistants and chat interfaces typically manage. Lightspeed Venture Partners led this Series A funding round, with additional participation from Accel, Nvidia's NVentures corporate venture arm, South Park Commons, and Define Ventures. The startup had previously raised a smaller 10 million dollar seed round before this larger institutional round closed.

Nuance is specifically targeting a well known weakness in most current voice assistants and AI avatars, namely that many existing systems effectively stitch together separate speech recognition, text based language model processing, and speech generation components as three distinct steps rather than treating the whole interaction as one unified process. That kind of multi stage pipeline can add meaningful latency to responses and often loses important visual cues, facial reactions, expressions, and other subtle signals that humans naturally rely on constantly during genuine face to face conversation. Nuance instead wants to build a single unified multimodal model that takes audio and video as direct combined inputs and generates audio and video outputs directly together, rather than processing each modality entirely separately through disconnected pipeline stages.

The company is specifically collecting specialized conversational training data purpose built for this task rather than relying primarily on material simply scraped from the broader open internet the way many other AI companies typically do. Nuance remains genuinely pre product at this stage but plans to release a research preview later this year, with potential applications the company has mentioned spanning interview training, language learning, presentation practice, sales training, customer service, and AI powered interviewing. The funding round also underscores growing investor interest more broadly in AI systems that move meaningfully beyond simple text chat interfaces toward genuinely real time multimodal interaction that more closely mirrors natural human conversation

Top Tech News Today, September 14, 2026: Apple, Anthropic, Microsoft, Revolut, SoftBank & More - Tech Startupstechstartups.com

[Merchant27]

Three former Apple researchers moving directly into building exactly this kind of natural multimodal conversational AI makes a lot of intuitive sense given how heavily Apple has historically emphasized seamless natural human computer interaction as one of its core design and product principles. Their specific prior institutional experience clearly maps directly onto what Nuance is now trying to build as an independent company. Wonder what specific technical bottleneck or limitation they personally hit while still working inside Apple that ultimately pushed them to leave and pursue this particular problem independently instead

[Finley_27]

Collecting purpose built specialized conversational data rather than relying entirely on scraped internet material is a smart and increasingly important strategic differentiator that more AI startups working in specific narrow niches like this should probably be pursuing themselves. Generic internet scale training data simply doesn't reliably capture the specific nuanced dynamics of natural face to face human conversation particularly well.

Curious exactly how they're sourcing and collecting that specialized conversational data at sufficient scale to properly train a genuinely capable model

[ECWCole36]

Nvidia's NVentures participating directly in this funding round is a notable detail worth flagging on its own, since Nvidia tends to invest fairly selectively and strategically in AI infrastructure adjacent startups rather than spreading its corporate venture capital broadly across the entire space. Their specific involvement here suggests genuine belief that multimodal conversational AI specifically represents a meaningfully significant emerging compute market worth strategically positioning around early. Their investment decisions in this space are generally worth watching closely as a leading signal

[Connor97]

Remaining pre product with a research preview still not scheduled to launch until later this year is a reminder that this specific 50 million dollar round represents a genuine and meaningful bet purely on the founding team's underlying technical vision and prior track record, rather than on any already demonstrated working product people can actually try or evaluate today. That's a substantial amount of capital and investor confidence to raise purely on vision and reputation alone. Says something notable about how much credibility this particular team apparently carries already within investor circles based on their previous Apple experience

[StevenArroyo]

The stated application areas here, spanning interview training, language learning, and sales practice specifically, all genuinely benefit enormously from picking up on subtle non verbal cues that pure text based AI interaction simply cannot capture or convey at all. A genuinely good language tutor for instance needs to reliably notice confusion or hesitation registering on your face, not just parse whatever words you happen to type or say out loud. If Nuance can actually deliver on that kind of capability reliably, the identified use cases described here all feel like genuinely legitimate and valuable target markets to pursue

[Sequence]

The stitched together pipeline problem described here is genuinely the exact reason so many current voice assistants still feel noticeably robotic and stilted despite years of steady incremental underlying model improvements. Each individual stage in a multi step pipeline adds both meaningful latency and inevitably loses some amount of information along the way as data passes between components. A truly unified multimodal approach handling everything together could represent a genuinely meaningful architectural improvement rather than just another marginal incremental gain