Kyutai TTS

Kyutai TTS is a text-to-speech model for streaming use. It reads text out loud, but it doesn't wait for you to finish typing. The model begins speaking as soon as it has the first few words, which is what separates it from offline TTS engines that need the whole sentence before they make a sound.

Kyutai is a French open-science AI lab funded by the Iliad Group, the CMA CM Group and Schmidt Sciences. The lab built Kyutai TTS 1.6B, one of several Kyutai audio models, for its own work on Moshi, a speech dialogue system, then released the weights to the public. That build is aimed at low-latency speech synthesis, a job most engines handle only after they see the full text. A viewer researching voice tools for its own work on Moshi, a speech dialogue system, then released the weights to the public. A viewer researching voice tools can use it to build a talking assistant, generate narration, or add speech to an app without paying per character.

The main catch is how the model reaches its speed. It was trained on 2.5 million hours of public audio and runs best on a GPU, so this isn't a plug-and-play cloud service with a clean monthly price. You download the weights, set up an environment, and run it yourself. Two versions exist: the 1.6B streaming model covered here, and the newer Pocket TTS, a smaller model made for CPU machines.

Getting Started

1. Create a free Hugging Face account and accept the model license on the Kyutai TTS page.
2. Install the moshi Python package with pip, or install the moshi-server crate if you plan to serve it in production.
3. Load the checkpoint and the Mimi audio codec in Python, then point the model at a text prompt.
4. Generate audio in chunks for streaming playback, or run it on a file for offline output.
5. Pick a ready-made voice from the tts-voices repository, or supply your own voice embedding.