HIERARCHICAL MULTI - CLASS SELF SIMILARITIES Tristan Jehan Media Laboratory Massachusetts

The music similarity question is yet to be defined. From an objective signal point of view, at least three classes of similarities may be considered: pitch, rhythm, and timbre. Within each class, several scales of analysis may be evaluated: segment, beat, pattern, section. We propose a recursive approach to this hierarchical and perceptual similarity problem. Our musically meaningful set of measurements is computed from the ground up via dynamic programming. This may be regarded as an analysis foundation for content-based retrieval, song similarities, or manipulation systems.