Consultation des notices

This master's thesis aims at exploring the challenge of automatically retrieving musical structures within an audio file. Our main contribution is to formulate this well-studied problem in the frame- work of computational information geometry, an emerging field at the frontier between statistics, differential geometry, and data mining. In this framework, we unify the fundamental tasks of event segmentation and similarity computing in a single sequential scheme for structure discovery. Furthermore, we propose an original metric on temporal segments, which combines several criteria of geometrical comparability : divergence between centroids, as well as inclusion and intersection ratios of corresponding information balls.