ProGress: Structured Music Generation via Graph Diffusion and Hierarchical Music Analysis

Stephen Ni-Hahn, Chao Péter Yang, Mingchen Ma, Cynthia Rudin, Simon Mak, Yue Jiang
September, 2025
Code

Abstract

ProGress (Prolongation-enhanced DiGress) is a novel symbolic music generation framework that integrates Schenkerian analysis with discrete graph diffusion models. Unlike existing black-box generative systems, ProGress emphasizes musical interpretability and structural cohesion. It adapts the DiGress discrete diffusion model for symbolic phrase generation, introduces a Schenkerian-inspired phrase fusion methodology, and allows controllable generation of harmonically and melodically coherent compositions. Through blinded human subject experiments, ProGress significantly outperforms prior state-of-the-art models and even surpasses Bach chorales in enjoyability metrics, despite using orders of magnitude fewer parameters (~3M vs. 500M+).

Type
Conference paper
Publication
Thirty-Ninth Conference on Neural Information Processing Systems (NeurIPS 2025), Creative AI Track
Symbolic Music Generation Diffusion Models Graph Neural Networks Interpretability Schenkerian Analysis
Chao Péter Yang
Chao Péter Yang
ML Research Assistant

My research focuses on symbolic music generation and graph machine learning, with broader interests in generative modeling, agentic AI systems, and large models. I aim to bridge theory and practice in data science to create both scientific and real-world impact.