A Corpus-based Study of Rhythm Patterns

We present a corpus-based study of musical rhythm, based on a collection of 4.8 million bar-length drum patterns extracted from 48,176 pieces of symbolic music. Approaches to the analysis of rhythm in music information retrieval to date have focussed on low-level features for retrieval or on the detection of tempo, beats and drums in audio recordings. Musicological approaches are usually concerned with the description or implementation of manmade music theories. In this paper, we present a quantitative bottom-up approach to the study of rhythm that relies upon well-understood statistical methods from natural language processing. We adapt these methods to our corpus of music, based on the realisation that—unlike words—barlength drum patterns can be systematically decomposed into sub-patterns both in time and by instrument. We show that, in some respects, our rhythm corpus behaves like natural language corpora, particularly in the sparsity of vocabulary. The same methods that detect word collocations allow us to quantify and rank idiomatic combinations of drum patterns. In other respects, our corpus has properties absent from language corpora, in particular, the high amount of repetition and strong mutual information rates between drum instruments. Our findings may be of direct interest to musicians and musicologists, and can inform the design of ground truth corpora and computational models of musical rhythm.