Pocket TTS

Pocket TTS is a text-to-speech model and Python package from Kyutai that runs on CPU, without a GPU or a hosted API. The model has about 100M parameters, uses two CPU cores, streams audio with roughly 200 ms to the first chunk, and generates speech about 6 times faster than real time on a MacBook Air M4. It handles arbitrarily long text input.

You can use it as a Python library, through a CLI that generates WAV files, or through a local HTTP server with a web interface that keeps the model in memory between requests. It ships a catalog of pre-made voices and clones a voice from any WAV sample. Voice embeddings can be exported to safetensors files so a cloned voice loads quickly later.

Pretrained models cover English, French, German, Portuguese, Italian, Spanish, and Dutch, with larger 24-layer variants for non-English languages that trade speed for quality. Training code is released, and community-trained models (Czech, Hindi, Korean, Persian, Indonesian, Estonian, Welsh, Polish, Greek) load through a configuration file from a URL or Hugging Face.

GPU execution is not officially supported. On hardware with strong single-thread CPU performance no GPU speedup was observed, while a cloud x86 VM with a Tesla T4 measured about a 2.6x speedup over its CPU. Community ports run the model in the browser through WebAssembly or ONNX Runtime Web, and in Rust, C++, C#, MLX, and on Android through LiteRT.