Crossing Phrase Boundaries in Music

Abstract

This paper presents a new model for segmenting symbolic music data into phrases. It is based on the idea that melodic phrases tend to consist of notes, which increase rather than decrease in length towards the phrase end. Previous research implies that the timing of note events might be a stronger predictor of both theoretical and perceived segmentation than pitch information. Our approach therefore relies only on temporal information about note onsets. Phrase boundaries are predicted at those points in a melody where the difference between subsequent note-to-note intervals reaches minimal values. On its own, the proposed model is parameter-free, does not require adjustments to fit a particular dataset, and is not biased towards metrical music. We have tested the model on a set of 6226 songs and compared it with existing rule-based segmentation algorithms that had been previously identified as good performers: LBDM and Grouper. Next, we investigated two additional predictors: meter and the presence of pauses. Finally, we integrated all approaches into a meta-classifier, which yielded a significantly better performance than each of the individual models.