Back to Gap Map Model components / Speech & audio

Pocket TTS

Kyutai
open weights / Overall score: 2.6

Kyutai's small text-to-speech model, about 100M parameters, built to run faster than real time on two laptop CPU cores, with streaming output and voice cloning from a short sample. It speaks English and several other European languages, installs as a Python package with a command line and a server, and runs in the browser through community ports. The repository also carries the code to train new voices and languages.

Openness

3 medium confidence
3.0
weights
open (Hub checkpoints; the voice-cloning model sits behind an automatically approved form and the model without voice cloning is ungated)
data
closed (the data behind the released model is not stated; the training guide's example recipe uses the public HiFiTTS-2 corpus)
code
open (training code with configs for training from scratch and for fine-tuning)
license
CC-BY-4.0 (model weights on the Hub; the code is MIT)

The weights are licensed CC-BY-4.0, which asks only for attribution, and Kyutai has published the code to train the model from scratch. What it does not say is which data the released model was trained on.

Adoption

2 high confidence
2.0

Hugging Face downloads of the two checkpoints over the trailing month. The pocket-tts pip package fetches weights from the Hub as well, so its own download count is not added.

Capability

3 low confidence
3.0

Pocket TTS has no listener rating, and the only shared-test-set results Kyutai publishes are for models trained with its own recipe, which it says match the released model. On that basis it sits level with F5-TTS, which reports on the same LibriSpeech test set.

Verified 2026-11-12, except some scores, still awaiting re-confirmation