Kyutai's Pocket TTS runs voice generation on pure CPU

Pocket TTS is an open-source text-to-speech engine from Kyutai designed specifically to run on standard CPUs without touching a GPU. At just 100 million parameters, it cuts out cloud API bills and heavy hardware setups to deliver fast, local voice generation.

What it can do

Speed and responsiveness are front and center here. The model delivers its first audio chunk in roughly 200ms, making it snappy enough for live, interactive voice agents.

On a MacBook Air M4, it runs across just two CPU cores while pushing speech generation to roughly 6x real-time. Kyutai also released stripped-down English variants at 18.2M, 6.9M, and 1.9M parameters, which can reach up to 25x real-time on a single CPU thread.

It handles continuous audio streaming and infinitely long text inputs without choking, meaning you can throw entire audiobooks at it. It also includes voice cloning: feed it a short audio snippet, and it mimics the speaker's vocal profile.

Sound quality is surprisingly natural, with decent cadence and rhythm. It can even pick up on subtle cues like sarcasm directly from the text. That said, dynamic emotional range is noticeably flatter than what you would get from multi-billion-parameter cloud TTS systems. While the repository mentions experimental multilingual support, English remains the only polished experience right now.

How to get it running

Getting Pocket TTS up and running takes very little effort.

If you have uv installed, running uvx pocket-tts generate pulls the model and outputs a test wav file immediately. Spinning up a local web UI is just as easy with uvx pocket-tts serve.

For voice cloning, parsing raw audio files repeatedly adds overhead. The CLI lets you export voice embeddings into lightweight safetensors files, allowing instant reloading without extra compute.

If you want to bypass the Python ecosystem entirely, Kyutai provides standalone prebuilt Rust binaries for Linux, macOS, and Windows. They require neither Python nor PyTorch to run on your CPU. The community has also created WebAssembly ports that run directly inside a web browser.

The project is fully open: inference code, pretrained checkpoints, and complete training scripts are available on GitHub and Hugging Face.

Why it matters

Pocket TTS solves a major headache for developers building local-first software. Running high-quality speech synthesis used to require dedicated VRAM or recurring cloud API costs; this drops the requirement down to a basic laptop CPU.

Compared to bloated multi-gigabyte alternatives, Kyutai found a sweet spot between footprint and intelligibility. We think the lack of robust multilingual support and fine-grained emotional control keeps it from replacing premium studio tools, but for offline reading and snappy local voice bots, it sets a remarkable benchmark for tiny models.