Modelling Musical Similarity for Drum Patterns
Fred Bruford, Mathieu Barthet, SKoT McDonald, Mark B. Sandler · 2019
Computational models of similarity for drum kit patterns are an important enabling factor in many intelligent music production systems. In this paper, we carry out a perceptual study to evaluate the performance of a number of state-of-the-art models for estimating similarity of drum patterns. 24 listeners rated similarity between 80 pairs of drum patterns covering a range of styles. We find that many of the models perform well, especially those using density-based features, and a more simplistic rhythm-pattern distance. However, many of the most perceptually important factors reported by listeners (such as swing, genre and style, instrument distribution) are not adequately accounted for. We also introduce a velocity transform method to better incorporate variable onset intensity into rhythm similarity analysis. Inter-rater agreement analysis shows that models are also limited somewhat by individual perceptual differences. These findings will inform future research into improved approaches to drum pattern similarity modelling that integrate existing features with new features modelling a wider range of characteristics.