Symbolic Segmentation: A Corpus-Based Analysis of Melodic Phrases

Gestalt-based segmentation models constitute the current state of the art in automatic segmentation of melodies. These models commonly assume that segment boundary perception is mainly triggered by local discontinuities, i.e. by abrupt changes in pitch and/or duration between neighbouring notes. This paper presents a statistical study of a large corpus of boundary-annotated vocal melodies to test this assumption. The study focuses on analysing the statistical behaviour of pitch and duration in the neighbourhood of annotated phrase boundaries. Our analysis shows duration discontinuities to be statistically regular and homogeneous, and contrarily pitch discontinuities to be irregular and heterogeneous. We conclude that pitch discontinuities, when modelled as a local and idiom-independent phenomenon, ] can only serve as a weak predictor of segment boundary perception in vocal melodies