Reproduce “Noise-to-Notes: Diffusion-based Generation and Refinement for Automatic Drum Transcription” on OpenResearch

About this paper

Michael Yeung, Keisuke Toyama, Toya Teramoto, Shusuke Takahashi, Tamaki Kojima

Researchers at Sony Group Corporation developed Noise-to-Notes (N2N), a diffusion-based generative framework that redefines Automatic Drum Transcription (ADT) as a conditional generation task. This approach achieved new state-of-the-art F1 scores on ADT benchmarks, reaching 89.68 on E-GMD and 94.90 on IDMT, by jointly predicting drum onsets and velocities and leveraging Music Foundation Model features for enhanced robustness to out-of-domain data.

Key insights

• Reframing ADT as a generative diffusion task can yield state-of-the-art performance, surpassing traditional discriminative models and enabling new capabilities like inpainting.
• A dynamically weighted loss function, such as the Annealed Pseudo-Huber loss, is crucial for effectively optimizing disparate targets like binary onsets and continuous velocities within a unified generative framework.
• Integrating high-level semantic features from Music Foundation Models significantly improves the generalization capability of ADT systems, making them more robust to variations in drum sounds across different datasets.

Summary by alphaXiv.