Similarity of structures in popular music

Abstract

The study of the similarity matrix of a song has been a particularly efficient technique to characterise song structures. This method transforms a song into a matrix representing the proximity between its different sections and is usually used to automatically detect structural properties such as its verse, its chorus, its tempo, etc. In this paper, these matrix representations are used not to study the inherent structure of a song, but to compare them with each other. This allows to create a metric on songs related to their pattern matrices, on which statistical tools can be applied. This metric is used to create groups of songs with similar structures and leads to interesting observations on patterns commonly used by certain artists, for certain years, and in certain genres. Moreover, this approach also unveils structures used across different features, such as songs from different decades and genres. Finally, this metric on songs is evaluated on classification tasks and shows that its interest lies in its ability to highlight specific behaviours rather than general trends.