# Model weight licensing status

**Project weight license: [Creative Commons Attribution-NonCommercial 4.0 International](licenses/CC-BY-NC-4.0.txt) (CC BY-NC 4.0).** This applies to all 29 project checkpoints identified in [manifest.json](manifest.json), only to rights the licensors are authorized to grant. Credit ACappellaSet and preserve upstream attribution; indicate modifications when sharing. It does not grant additional third-party rights or change upstream license terms. The full-remix and single-clone singer-augmentation checkpoints were trained with voice-converted singer augmentation generated using Moises; their project license does not grant additional rights controlled by service providers or other third parties.

## Upstream notices

Demucs source code uses MIT; preserve the [Meta notice](licenses/Demucs-MIT.txt).
Pretrained initialization identities and tensor correspondence are recorded in the [initialization provenance](https://github.com/pting126/ACappellaSet/blob/main/third_party/initialization.json).

Historical fixed and specialist training uses Music-Source-Separation-Training;
preserve its [MIT notice](licenses/MSS-MIT.txt).
Singer- and dual-conditioned models include fine-tuned WeSpeaker ReDimNet2
B6-LM encoder weights; preserve the upstream [Apache license](licenses/WeSpeaker-Apache-2.0.txt)
and [WeSpeaker attribution](https://huggingface.co/Wespeaker/wespeaker-voxceleb-redimnet2-B6-LM).
Project fine-tuning changes pretrained tensors. All released checkpoints are inference-only exports, retaining every separator and matching encoder tensor while omitting optimizer state, training arguments, and internal filesystem paths.

Original documentation and metadata retain their [MIT license](LICENSE); that MIT license does not apply to model weights.
Dataset audio, corpus terms, and third-party rights remain separate.