Gradium

Gradium is a Paris-based startup building foundational audio language models designed to enable natural, real-time voice interfaces between people and machines. The company emerged from stealth in December 2025 after being founded in September 2025. It was established by a team of prominent researchers from Google DeepMind and Meta: Neil Zeghidour (CEO), Olivier Teboul (CTO), Alexandre Défossez (CSO), and Laurent Mazaré (CCO). This founding team is recognized for inventing and open-sourcing key technologies in generative audio, such as neural audio codecs and audio language models.

Gradium was spun out of Kyutai, a non-profit AI research lab in Paris co-founded by Zeghidour and backed by notable figures like Xavier Niel and Eric Schmidt. The company launched with an exceptionally large seed funding round of $70 million, co-led by FirstMark Capital and Eurazeo, with participation from DST Global Partners, Eric Schmidt, and Xavier Niel, among others. This significant investment highlights strong investor confidence in the team's expertise and its mission to solve core challenges in voice AI.

The company's core product is a family of audio-native models that unify generation, transcription, transformation, and dialogue into a single neural architecture. This technology aims to eliminate the traditional trade-offs between speed, cost, and naturalness in voice AI. Gradium's platform provides ultra-low latency, emotionally expressive, and realistic voice synthesis and transcription. Its services, accessible via an API, are designed for developers and enterprises for applications in gaming, customer support, healthcare, and digital advertising. At launch, the platform supports English, French, German, Spanish, and Portuguese.

Keywords: audio language models, voice AI, speech synthesis, text-to-speech, speech recognition, generative audio, low latency voice, conversational AI, neural audio codecs, real-time voice, speech-to-text, AI voice generation, natural language processing, deep learning, machine learning, AI API, voice interface, expressive speech, multilingual AI