The Researchers Who Wrote Voice AI's Foundations Are Now Selling It

01 /The Concept

The foundational research papers behind most of today's voice AI — EnCodec, SoundStream, Moshi — were written by four researchers who are now selling API access to what they built.

Gradium launched in December 2025 as a spinout from Kyutai, the Paris-based AI research lab, co-founded by Neil Zeghidour, formerly at Google Brain, DeepMind, and Facebook AI Research. The three other co-founders share the same lineage. Together, the team built the audio compression and real-time speech synthesis architectures that ElevenLabs, Deepgram, and AssemblyAI are each implementing from published papers. In July 2026, Nvidia backed an extension to Gradium's seed round, bringing total disclosed funding to $100 million.

The product is a unified voice infrastructure stack for developers building production voice agents: streaming speech-to-text, expressive text-to-speech, high-fidelity voice cloning, WebSocket multiplexing, on-device models, and Gradium Translate, which converts speech to speech across five languages in real time. The technical differentiator is Semantic Voice Activity Detection — turn detection based on whether the speaker has finished a thought, not whether acoustic silence has occurred. Pricing runs from a free tier through five paid tiers, XS at $13 through L at $1,615 per month, with voice cloning and streaming included on all paid plans and five deployment models available, from shared cloud to full on-premises. Renault is among the production customers.

02 /Why It Matters

The latency ceiling is the core commercial problem in voice AI. Conversational agents that pause 800 milliseconds before responding feel unnatural in a way that text chatbots do not — the gap breaks the social contract of conversation. For consumer applications, that means abandonment. For enterprise voice use cases — receptionists, booking systems, outbound sales automation — it means measurable conversion loss.

When Semantic VAD was introduced — detecting conversational turn-taking based on semantic completion rather than acoustic silence — most commercial voice AI vendors were still doing silence detection because that was the method described in every prior published implementation. Gradium's founders knew the semantic approach was coming not because they read the paper but because they wrote it. That means the product shipped Semantic VAD as a feature while competitors were still evaluating whether to implement the older approach. This iteration speed — understanding the architectural roadmap because the team drew it — compounds over time in ways that a vendor catching up from published research cannot replicate.

The five-deployment-model architecture (cloud, inference partners, dedicated instances, private cloud, on-premises) directly addresses the enterprise procurement obstacle. Regulated industries cannot route customer audio through shared cloud infrastructure. An on-premises deployment with the same API surface as the cloud product resolves the security review that has blocked enterprise voice AI adoption in banking, healthcare, and government — markets where voice agent economics are largest and competitor penetration remains lowest.

03 /Opportunities

The voice infrastructure market will compress toward commodity pricing on raw ASR and TTS within 18 months as model distillation reduces compute requirements. Margin will shift to turn detection quality, emotional expressiveness range, and multi-language robustness — capabilities that require ongoing research rather than efficient serving. ElevenLabs competes on voice cloning and TTS quality; Deepgram on transcription accuracy at speed; AssemblyAI on structured audio intelligence. Gradium's research origin gives it the most defensible position in the capabilities that require the deepest model understanding.

The specific unaddressed gap is real-time multilingual voice agents for markets with high accent diversity. Customer service in India, Southeast Asia, and Brazil requires not just language support but dialect robustness and code-switching — conversations that flow between two languages mid-sentence. IVR replacement in these markets is the highest-value unsolved use case: call volumes are enormous, existing automation rates are low, and the language complexity has kept current voice AI from deploying effectively. A builder targeting IVR replacement in one of these markets, using Gradium's multilingual stack as the infrastructure layer, is addressing a problem where the TAM is large, the incumbent (legacy IVR hardware) is both dominant and universally resented, and no current voice AI vendor has built specifically for the use case.