# Third-party notices

This repository preserves Kokoro's licensing. Packaging, tensor conversion, and ONNX export do
not replace or relicense upstream work.

## Kokoro models and voices

- Kokoro v1.0 model, configuration, and all 54 published voices:
  [`hexgrad/Kokoro-82M`](https://huggingface.co/hexgrad/Kokoro-82M), revision
  `f3ff3571791e39611d31c381e3a41a3af07b4987`.
- Kokoro v1.1-zh model, configuration, 100 Mandarin voices, and 3 English voices:
  [`hexgrad/Kokoro-82M-v1.1-zh`](https://huggingface.co/hexgrad/Kokoro-82M-v1.1-zh),
  revision `01e7505bd6a7a2ac4975463114c3a7650a9f7218`.
- Polish and German fine-tuned checkpoints and speaker embeddings:
  [`software-mansion/react-native-executorch-kokoro`](https://huggingface.co/software-mansion/react-native-executorch-kokoro),
  revision `3744b57964eab7df6e8c48f0b84badb29e14df07`.
- ONNX exporter runtime code: [`SuvioMedia/kokoro`](https://github.com/SuvioMedia/kokoro),
  revision `8ee8ec3179c38990dd0564fb703c4deb68be58e9`.

These materials and the resulting ONNX model files and raw speaker-style matrices are distributed
under Apache License 2.0. The license text is available in [`LICENSE`](LICENSE) and
[`third-party/kokoro/LICENSE`](third-party/kokoro/LICENSE).

## Phonemis

The neural phonemizer models for `de`, `en-gb`, `en-us`, `es`, `fr`, `hi`, `it`, `pl`, and `pt`,
plus the English full lexicons and tagger models, come from
[`IgorSwat/Phonemis`](https://github.com/IgorSwat/Phonemis), revision
`71eb1ce33bd586d38cbac037843b8539d7829c3b`.

For release `v2.1.0`, the nine neural weight containers are deterministically wrapped as ONNX
graphs ahead of distribution. Exact source and output hashes, sizes, and converter provenance are
recorded in [`sources/phonemis-onnx-v2.1.0.json`](sources/phonemis-onnx-v2.1.0.json).

Phonemis is distributed under the MIT License. Its unchanged license text is available in
[`third-party/phonemis/LICENSE`](third-party/phonemis/LICENSE). This separate MIT license applies
to the Phonemis files and does not alter Kokoro's Apache-2.0 license.

## Provenance

Exact source paths, immutable revisions, original LFS hashes, converted artifact hashes, and sizes
for all 157 upstream voices are recorded in
[`sources/upstream-assets.json`](sources/upstream-assets.json). Additional checkpoint hashes are
recorded in [`SOURCES.lock.json`](SOURCES.lock.json).

No eSpeak binary, GPL component, Misaki package, Open JTalk dictionary, or UniDic dictionary is
distributed in this repository.