Article
Downbeat Tracking for Western Classical Music Recordings: A Case Study for Beethoven Piano Sonatas (en)
* Presenting author
Abstract:
Tracking beats and downbeats are fundamental skills that enable humans to comprehend music and engage with it. While both beats and downbeats exhibit periodicity over time, estimating downbeats demands a deeper understanding of musical aspects, such as onsets, beats, melodies, phrases, and thus requires a larger musical context. To assess the efficacy of models in learning downbeats, it is crucial to utilize datasets of different musical styles that encompass varying degrees of complexity, tempo changes, and expressivity. However, due to the scarcity of high-quality annotated datasets of expressive classical music, the performance and behavior of state-of-the-art downbeat tracking models is largely unexplored in this context. In this study, we conduct a comprehensive performance analysis of existing downbeat tracking models using a carefully curated dataset of Beethoven Piano Sonatas with downbeat annotations, comprising pieces and performances of various levels of expressivity. In particular, we use context-sensitive and metric level-sensitive evaluation measures to better understand the models’ benefits and limitations. Furthermore, we explore the impact of training data, categorize sources of errors, and suggest potential directions for future research in this area.