Kyutai Labs has done it again.

Last week, they quietly released a project called Pocket TTS, a lightweight text-to-speech solution that can run entirely on CPU. With only 100 million parameters, it requires no GPU, no API key, and no payment.


Conventional TTS models rely on Transformer-based token generation, which brings high latency and heavy computational load. Pocket TTS bypasses this mechanism and outputs audio streams directly.

Measured test results:

- On a MacBook Air M4, generation speed reaches 6 times real-time speech speed

- Latency for the first audio chunk is approximately 200ms

- Only uses 2 CPU cores

- Voice cloning requires just a 5-second audio sample

- Supports 6 languages out of the box

The model is small enough to fit on any edge device. It is released under the MIT license, with all training data sourced from public datasets. There are no hidden terms or "for research purposes only" restrictions.

Installation is straightforward:

```bash

pip install pocket-tts

```

You can then generate speech via the Python API or CLI. It supports Python 3.10 to 3.14, depends on PyTorch 2.5+, and does not require the CUDA-enabled version.

For developers, this means you can run TTS on Raspberry Pi, old laptops, even cloud functions without renting GPU instances.

For regular users, this means you can generate speech for free on your own device, clone your voice, or have text read out in any of the 6 supported languages — all without sending private data off your local machine.

Kyutai Labs is a non-profit AI research lab based in France, previously known for releasing the Moshi speech model. Pocket TTS follows their signature style: no chasing trends, no empty hype, just a working, ready-to-use model dropped straight to the public.

The repository link is in the replies to the original tweet — go check it out yourself.