Kyutai has released PocketTTS, a text-to-speech system with 100 million parameters that runs on a dual-core processor

The Paris-based laboratory has released, under the MIT licence, a multilingual speech synthesis model – including French – which is six times faster than real time on a MacBook Air.

Kyutai, the open research laboratory founded in Paris, is launching PocketTTS, a voice synthesis model of only 100 million parameters designed to work without graphics. Two processor cores are enough: on a MacBook Air M4, it produces speech about six times faster than real time, and the first audio fragments arrive in about 200 milliseconds, making it suitable for interactive voice assistants.

Seven languages and voice cloning

The model supports English, French, German, Spanish, Italian, Portuguese and Dutch; deeper, 24-layer variants improve quality in some languages other than English. Any WAV recording can serve as a reference voice, the team recommending to clean the audio in advance.

On the research side, Kyutai trained the sample head of the model with a method called drifting, presented as an alternative to his previous approach. Code and weights are distributed under MIT license.