SongGen

Beihang University, Shanghai AI Laboratory, The Chinese University of Hong Kong, Harbin Institute of Technology, CPII (InnoHK) · 2025

A single-stage auto-regressive transformer designed for controllable song generation.

Application types: Text-to-songLyrics-to-song
Architecture:Autoregressive transformer

Essential categories

CategoryStatusNotes
Source code✔︎ OpenThe source code is available on GitHub, but there are some parts missing regarding data processing (indicated as to-do in GitHub).
Training data~ PartialThis model is trained on 8,000 hours of openly licensed music data, including MSD, FMA and MTG-Jamendo. However, the curated version of the dataset is not available (indicated as to-do in GitHub).
Model weights✔︎ OpenWeights are available on Hugging Face.
Code documentation✔︎ OpenCodebase is well documented, with a README file that includes installation and inference instructions, and links to the training procedure.
Training procedure✔︎ OpenThe training procedure is described in the article (see Appendix A). Also, there are specific training guidelines in the GitHub repository.
Evaluation procedure✔︎ OpenEvaluation procedure is described in the article (see Appendix B). The dataset used for evaluation is available on Hugging Face.
Research paper✔︎ OpenAccepted at ICML 2025.
Licensing✔︎ OpenApache License 2.0

Desirable categories

CategoryStatusNotes
Model card∅ Not included
Datasheet∅ Not included
Package∅ Not included
User-oriented application∅ Not included
Supplementary material page⭐ IncludedA demo page with sound examples is available. The page also includes a link to the GitHub repository and research paper.

Raw YAML file with complete evaluation.