Melodic segmentation: evaluating the performance of algorithms and musical experts

We review several segmentation algorithms, qualitatively highlighting their strengths and weaknesses. We also provide a detailed quantitative evaluation of Temperley’s Grouper and Cambouropoulos’ Local Boundary Detection Model. In order to benchmark the algorithms’ performance against musicians, we compiled a corpus of different melodic excerpts and collected individual segmentations from 19 musicians. We then evaluated each algorithm’s flexibility and practicality by running several experiments using large, hand-segmented portions of the Essen folk song collection and the melodies from our corpus.