Comparing repetition-based melody segmentation models

Abstract

This paper reports on a comparative study of computational melody segmentation models based on repetition detection. For the comparison we implemented five repetition-based segmentation models, and subsequently evaluated their capacity to automatically find melodic phrase boundaries in a corpus of 200 folk melodies. We systematically investigate the effects that the choice of melodic representation, similarity measure, and parameter settings have on each model’s performances. We discuss at length issues such as parameter sensitivity, generalization capability, and efficiency. The best performing model employs a similarity matrix to identify repetitions, and selects which repetitions are used to segment the input melody using an optimisation-based search algorithm.