Masked Diffusion Enables Coherent Beat Tracking

Official inference code and pretrained models for the ISMIR 2026 paper “Masked Diffusion Enables Coherent Beat Tracking” by Francesco Foscarin, Filip Korzeniowski and Richard Vogl. This work was done at Moises AI. If you use this work, please cite the paper.

For research only. These models were developed and released for research purposes only, under a non-commercial license. If you need beat and downbeat annotations for commercial use, you can get them through Music.AI or the Moises mobile app. These services rely on a separate model, built from the ground up for commercial use and offering improved accuracy.

Contents

• Installation
• Quick start
• Usage
• Available models
• Released predictions and activations
• Results
• Reproducing the results
• Differences from Beat This!
• License
• Citation

##…