Masked Diffusion Improves Beat Tracking Coherence

A recent study published on arXiv (2608.04624) introduces a masked diffusion technique aimed at tackling a core issue in beat tracking. Current neural networks frequently yield invalid results, such as consecutive downbeats and unpredictable tempo shifts, even when such occurrences are not present in the training datasets. The researchers suggest that this problem arises from insufficient modeling of various plausible beat grid outputs, resulting in conflicting interpretations. To address this, they propose a masked diffusion strategy that accurately represents multiple outputs and facilitates consistent predictions through iterative inference. The method includes three significant enhancements: independent masking of beats and downbeats during both training and inference, a balanced masking schedule for inference, and peak-picking throughout inference stages. This strategy minimizes erratic behavior and enhances beat tracking accuracy. The full paper can be accessed at https://arxiv.org/abs/2608.04624.