Rule Mining for Local Boundary Detection in Melodies

Abstract

The task of melodic segmentation is a long-standing MIR task that has not been solved, yet. In this paper, we shortly review existing approaches, most of which are either based on rule-sets derived from Gestalt principles, or on a statis- tical learning approach. We use a method related to both approaches. A rule mining algorithm is employed to find a rule set that classifies notes within their local context as phrase boundary. The advantage of a rule-based model is its interpretability. By inspecting the rules, some important clues are revealed about what constitutes a melodic phrase boundary, notably a prevalence of rhythmic features over pitch features. Both the discovered rule set and a Random Forest Classifier trained on the same data set outper- form previous methods on the task of melodic segmenta- tion of melodies from the Essen Folk Song Collection, the Meertens Tune Collections, and the set of Bach Chorales.