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.
Source code incudes model achitecture, training pipeline, inference, and data processing.
Training data
✘ Closed
Training data is not available nor properly described. In the technical report (Section 3.1.1), authors indicate that the model was trained on about 100,000 hours of music (~1.8 million musical pieces). However, no information on specific data sources, acquisition methods, and data licensing is given.
Codebase is properly documented, including details on model's installation and usage. In addition, authors provide a specific page with training instructions (https://github.com/ace-step/ACE-Step/blob/main/TRAIN_INSTRUCTION.md).
Training procedure is describe with detail in the technical report. Appendix A provides additional details on training parameters and hardware requirements.
Evaluation procedure is described in the technical report, including human-evaluation and objective metrics evaluation. Despite evaluation data not being openly available, Appendix B includes the detailed description of the 20 prompts (with lyrics) used to generate the samples for the model evaluation.
A supplementary page is available, including a brief description of the model, links to relevant sites (e.g., demo, paper and code), some examples showcasing the model's abilities, limitations and future work.
