Phonon : Gradium, from Kyutai, publishes an open source voice synthesis of 31 million parameters that rotates on the phone

The Parisian start-up launches two models of 31 and 40 million parameters, multilingual and licensed under Apache 2.0, designed to speak without connection on mobile, PC or browser.

Gaudium, the voice company founded by the co-founders of the Paris laboratory Kyutai, unveiled on 21 October during the NVIDIA GTC in Berlin, Phonon, a family of voice synthesis models designed to work entirely on the user's device. No return to a server: both text and audio remain local. Two sizes are proposed, 31 and 40 million parameters, about two and a half times less than Kokoro (82 million), the open reference model in this category. According to blind listening tests conducted by Gradium, Phonon would be preferred to Kokoro, Magpie and Pocket TTS.

Five languages in one model

The multilingual model covers English, French, German, Spanish and Portuguese from the start, with a catalogue of ready-to-use voices and a speed adjustment. It weighs less than 100 MB and integrates into Rust, Python, Swift or browser via WebAssembly and WebGPU. On a server equipped with a NVIDIA L40S card, the first voice fragment comes out in 5 ms, against 43 ms on an AMD EPYC processor.

The weights are published on Hugging Face and the inference code on GitHub, under Apache 2.0 license. Cloning and voice creation are still reserved for the Gradium Pay API. Launching partners include Hugging Face, Arm, Qualcomm, LM Studio, Ollama and FORVIA Appning, which clearly targets mobile applications, automotive embarkation, video game and robotics.