Multi-strategy Segmentation of Melodies

Abstract

Melodic segmentation is a fundamental yet unsolved problem in automatic music processing. At present most melody segmentation models rely on a ‘single strategy’ (i.e. they model a single perceptual segmentation cue). However, cognitive studies suggest that multiple cues need to be considered. In this paper we thus propose and evaluate a ‘multi-strategy’ system to automatically segment symbolically encoded melodies. Our system combines the contribution of different single strategy boundary detection models. First, it assesses the perceptual relevance of a given boundary detection model for a given input melody; then it uses the boundaries predicted by relevant detection models to search for the most plausible segmentation of the melody. We use our system to automatically segment a corpus of instrumental and vocal folk melodies. We compare the predictions to human annotated segments, and to state of the art segmentation methods. Our results show that our system outperforms the state-of-the-art in the instrumental set.