Gradium

Gradium is releasing a beta of its text-to-speech model aimed at developers building voice agents, with a headline claim of under 50 millisecond time-to-first-audio while most competing systems sit above 100ms. The launch matters because in real-time voice agents the audio must start within 200 to 300 ms for a conversation to feel natural, and TTS latency typically eats most of that budget. This model is positioned as the voice layer that agent builders can drop underneath an LLM without introducing the awkward pauses that give current assistants away. The company itself is a real-time voice AI platform for developers and companies deploying voice agents, built by researchers and co-founded by Neil Zeghidour, Laurent Mazaré, Olivier Teboul, and Alexandre Défossez . The same group previously co-founded Kyutai and worked at Meta, Google DeepMind, and Google Brain , and shipped Moshi and Hibiki before turning the underlying audio language model research into a commercial API covering TTS, speech-to-text, and voice cloning. The beta arrives in a market where latency benchmarks have become contested territory, and Gradium has been publishing its own methodology alongside third-party leaderboards like Coval. Gradium says its new TTS beta starts speaking in under 50ms without losing quality, while public benchmark pages still measure its existing model above 200ms , so the launch is effectively an invitation for founders and infra teams to test the new model directly against whatever voice stack they currently run in production.