Beyond the Beats: Can Downbeat Trackers Predict Hypermetre?

Hypermetre is the metric structure between the timescales of bars and sections in a piece of music. Modelling hypermetre could make existing downbeat tracking and structure analysis algorithms more robust. However, this topic is underexplored compared to downbeat tracking and structure analysis, perhaps due to the lack of annotated data for training. In this work, we evaluate the success of using existing downbeat tracking algorithms to perform hypermetre tracking, proposing five methods to do so. We identify the McGill Billboard Dataset as a large, usable evaluation set for hypermetre tracking. We define four levels of hypermetric complexity, and find that a straightforward baseline is best for the simplest cases, but that adaptations of downbeat trackers can perform better for more complex cases.