MusGO Framework: Assessing Openness in Music-Generative AI
This website builds on the paper MusGO: A Community-Driven Framework for Assessing Openness in Music-Generative AI, authored by Roser Batlle-Roca, Laura Ibáñez-Martínez, Xavier Serra,
Emilia Gómez, and Martín Rocamora, and published in the Proceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR 2025).
It serves not only as a companion to the publication, but also as a living resource, which is continuously updated and shaped by contributions from the community.
A music-domain-knowledge-inspired text-to-music latent diffusion model capable of fine-grained control over harmonic and rhythmic elements via an integrated MuNet architecture.
Mustango is trained on the MusicBench dataset, which is a collection of ~52k 10-seconds music samples. These samples were generated through targeted harmonic, rhythmic, and dynamic data augmentations of the original MusicCaps dataset and enriched with automated music theory text descriptions.
Training procedure is well documented in the paper, including hardware requirements and model configurations (see Appendix B for further details). Additional implementation details are available in the GitHub repository.
Evaluation methodology is well documented in the paper, including evaluation metrics and model performance. Evaluation dataset is composed of TestA, TestB and FMACaps. TestA and TestB are the testing splits of MusicBench, which is based on MusicCaps (see Section 2.3). Further information about how FMACaps dataset was created is available in Appendix F.
There is a data card available for the MusicBench dataset on HuggingFace, which provides information about the dataset's composition, usage, and licensing.
A demo is available on Replicate. There should also be a demo on HuggingFace Spaces, but it is currently not working (https://huggingface.co/spaces/declare-lab/mustango).
