# Model weight licenses and attribution

This repository contains model weights and configuration files. The root
[LICENSE](LICENSE) applies to the project's `s1`, `s2`, `aligner` and `vocoder`
weights under GPL-3.0-only. All four were trained from random initialization.

Upstream pretrained components retain their own terms:

| Component | Upstream model | Weight terms and attribution |
| --- | --- | --- |
| `plbert` | [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M), phoneme ALBERT subset | [Apache-2.0](plbert/LICENSE); [component notice](plbert/NOTICE.md) |
| `hubert` | [TencentGameMate/chinese-hubert-base](https://huggingface.co/TencentGameMate/chinese-hubert-base) | Upstream model card declares [MIT](hubert/LICENSE); [component notice](hubert/NOTICE.md) |
| `speaker` | `wavlm_large_finetune.pth`, available through [Seed-TTS-Eval](https://github.com/BytedanceSpeech/seed-tts-eval#metrics) | [Download and upstream terms](speaker/NOTICE.md) |

Retain the applicable license text, attribution and modification notices when
redistributing these components.

Download the [speaker checkpoint](https://drive.google.com/file/d/1-aE1NfzpRCLxA4GUxX9ITI3F9LlbtEGP/view)
from the link provided by Seed-TTS-Eval and save it as
`speaker/wavlm_large_finetune.pth`. Its upstream terms apply.

`bundle.json` records SHA256 hashes for the configurations and weights.
Rights to recordings and voices are separate from model licenses.