Modeling Rhythmic Attributes in Music At Scale with Tree Ensembles and the Music Genome Project

Musical meter and attributes of the rhythmic feel (e.g., swing) are crucial when defining musical style, but they are often difficult to capture in music audio signals. In this work, we employ a set of tree ensemble models trained using a rhythm-inspired acoustic feature set. We model a total of nine rhythmic attributes covering meter and feel using over one million examples labeled by experts from Pandorar Internet Radio’s Music Genome Project r. While linear models are shown to be somewhat effective, the complexities of rhythm are better represented using more powerful, non-linear, tree ensemble methods.