Pattern Clustering in Monophonic Music by Learning a Non-Linear Embedding From Human Annotations

Musical pattern discovery algorithms find instances of repetition in symbolic music, allowing for some user-specifiable amount of variation between identified repetitions; however, they can yield an intractably large number of discovered patterns when allowing for even small amounts of variation. This is commonly addressed by defining some heuristic notion of pattern significance, and returning only the most significant patterns. This paper develops a method of pattern discovery that models human judgement of what constitutes a significant pattern by incorporating annotations of repeated patterns, avoiding the need to design heuristics. We take pattern discovery as a clustering task, where the input is a set of passages of monophonic music, represented as vectors of extracted features, and the output clusters correspond to discovered patterns. The human annotations are used to train a neural network to learn a low-dimensional embedding of the feature space that maps passages of music close together when they are occurrences of the same ground-truth pattern. The results of this approach match up with the annotations significantly better than the results of an approach using clustering without subspace learning. We provide examples of the types of patterns that this method tends to discover and discuss its feasibility and practicality as a tool for extracting useful information about repetitive structure in music