Back to Gap Map Model components / Speech & audio

Kokoro

hexgrad
open weights / Overall score: 3.8

An 82-million-parameter open-weight text-to-speech model with dozens of preset voices across English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese and Mandarin. Small enough to run on a CPU, it is widely deployed in local apps and commercial APIs.

Maintained by the individual developer hexgrad; the kokoro Python package is the inference library.

Openness

3 high confidence
3.0
weights
open (.pth checkpoint and voice packs, ungated)
data
described (a few hundred hours of permissively licensed audio, listed in part)
code
partial (inference library only)
license
Apache-2.0 (OSI)

Kokoro's weights and inference library are Apache-2.0, an OSI-approved license. The card lists some of the openly licensed audio it trained on and sizes the rest at a few hundred hours, but the training set and training code are not published.

Adoption

5 high confidence
5.0

Hugging Face downloads of the single weights repository. The kokoro package is marked as not a separate channel because it loads this repository.

Capability

3 medium confidence
3.0

Kokoro is mid-table in blind listening tests, well behind the commercial leaders, but it is by far the smallest model near that position.

Verified 2026-09-27