Groove Kernels as Rhythmic-Acoustic Motif Descriptors

Proceedings of the International Society for Music Information Retrieval
The “groove” of a song correlates with enjoyment and bodily movement. Recent work has shown that humans often agree whether a song does or does not have groove and how much groove a song has. It is therefore useful to develop algorithms that characterize the quality of groove across songs. We evaluate three unsupervised tempo-invariant models for measuring pairwise musical groove similarity: A temporal model, a timbre-temporal model, and a pitch-timbre-temporal model. The temporal model uses a rhythm similarity metric proposed by Holzapfel and Stylianou, while the timbre-inclusive models are built on shift invariant probabilistic latent component analysis. We evaluate the models using a dataset of over 8000 real-world musical recordings spanning approximately 10 genres, several decades, multiple meters, a large range of tempos, and Western and non-Western localities. A blind perceptual study is conducted: given a random music query, humans rate the groove similarity of the top three retrievals chosen by each of the models, as well as three random retrievals.