Gradium-V1 Omni-Voice Model Hits 120ms Latency in $100M Seed

Paris-based AI startup Gradium has secured a $100 million seed funding round led by Nvidia via NVentures. The capital injection provides prioritized access to H200 and B200 Tensor Core GPU clusters for training audio foundation models. Gradium will use the funding to open a secondary headquarters in San Francisco and hire 50 engineers to compete directly with US labs in the enterprise voice sector.

Eurazeo, Iris Capital, and several European tech angel investors also participated in the round. The investment highlights the aggressive capital requirements necessary to train and serve real-time multimodal inference models.

The Gradium-V1 Architecture

Gradium launched its Gradium-V1 foundation model family alongside the funding. Traditional Text-to-Speech systems rely on a cascaded approach with separate linguistic and acoustic models. Gradium-V1 uses a unified “Omni-Voice” architecture that processes text and audio simultaneously.

The model achieves an end-to-end latency of under 120ms. For developers building real-time voice agents, this sub-200ms threshold is necessary to prevent unnatural pauses in conversational turn-taking.

Gradium trained the model on a proprietary dataset containing 1.2 million hours of multilingual speech. This dataset allows the model to interpret and replicate micro-emotions like hesitation, laughter, and breathing patterns natively within the audio output.

The initial release supports 32 languages. Gradium specifically prioritized high-fidelity performance and native-level prosody in French, German, and Spanish to differentiate its offerings from predominantly English-centric models.

Enterprise Strategy and Availability

Gradium is adopting a dual-homed strategy. The company will maintain its core research hub in France while establishing commercial and executive operations in Silicon Valley. This structure allows the startup to tap into European engineering talent while accessing the US enterprise market.

The company is targeting automated customer service and real-time media translation. Gradium has opened a private beta of its API for select enterprise partners in the telecommunications and gaming industries. This positions them against other well-funded voice infrastructure providers focusing on high-volume commercial deployments.

Gradium plans to launch a public developer platform in Q4 2026. If you build production voice applications, the upcoming API will require evaluating whether unified acoustic models provide enough latency reduction to justify migrating from existing text-to-speech pipelines.