Unsupervised learning and refinement of rhythmic patterns for beat and downbeat tracking

In this paper, we propose a method of extracting rhythmic patterns from audio recordings to be used for training a probabilistic model for beat and downbeat extraction. The method comprises two stages: clustering and refinement. It is able to take advantage of any available annotations that are related to the metrical structure (e.g., beats, tempo, downbeats, dance style). Our evaluation on the Ballroom dataset showed that our unsupervised method achieves results comparable to those of a supervised model. On another dataset, the proposed method performs as well as one of two reference systems in the beat tracking task, and achieves better results in downbeat tracking.