/
Variation Transformer: New datasets, models, and comparative evaluation for symbolic music variation generation

Variation Transformer: New datasets, models, and comparative evaluation for symbolic music variation generation

Chenyu Gao, PhD Student, School of Arts and Creative Technologies

Summary

This project addresses the lack of open datasets in symbolic music variation generation by introducing an algorithm to automatically extract theme-and-variation pairs from arrangements. The resulting annotated datasets (POP909-TVar and VGMIDI-TVar), models, and evaluation materials are fully open-sourced. Clear documentation was provided to support reuse and reproducibility, and a demo page was created to engage non-specialist audiences.

Case Study

This project tackles a gap in symbolic music variation generation: the limited availability of open datasets and reproducible models. An algorithm was developed to automatically extract theme-and-variation pairs, which was then applied to two open-sourced music datasets: POP909 and VGMIDI. The annotated results produced two new open datasets: POP909-TVar (2,871 pairs) and VGMIDI-TVar (7,830 pairs).

Alongside the dataset development, two symbolic music variation generation models, one non-deep learning and one deep learning, were proposed. Listening studies and feature-based evaluations were conducted. The newly developed Variation Transformer outperformed all others in perceived “variation success.”

The full project was shared openly, including code and documentation via GitHub, trained models via Zenodo, and a demo for non-specialists via GitHub. Special attention was paid to producing clear README documentation to support reuse. The Variation Transformer has also been integrated into Cocreate, a web-based co-creative AI music tool, and is expected to be used in the 2025 AI Song Contest and in Master’s-level teaching

Other contributors:

  • Dr. Federico Reuben (Supervisor)

  • Dr. Tom Collins (Supervisor)

Links:

Funding acknowledged:

  • The Viking cluster (University of York)

  • China Scholarship Council and the University of York (PhD support)

Licensing Information

Except where otherwise noted copyright in this work belongs to the author(s), licensed under a Creative Commons Attribution-NonCommercial 4.0 International License