Polyphonic Music Retrieval by Local Edition of Quotiented Sequences

Melody is an important property for the perceptual description of Western musical pieces. A lot of applications rely on the comparison of two melodies. Several existing techniques assume a monophonic context or extracts a monophonic melody from polyphonic pieces. In this paper, we evaluate the contribution of a method which considers a polyphonic context. We propose an algorithm that computes a similarity measure between polyphonic musical sequences. All the notes are taken into account, even if they sound at the same time. Melodies are represented by quotiented sequences. A quotiented sequence is a sequence graph defined with an additional equivalent relation on its vertices and such that the quotient graph is also a sequence graph. The core of the method relies on an adaptation of edit-distance metrics, regularly applied in bioinformatic context. These methods are currently being used to evaluate the similarity between a monophonic query and a database of polyphonic musical pieces. First experiments show that the adaptation to polyphony does not degrade the quality of the algorithm with monophonic musical pieces. Furthermore, the results of experiments with polyphonic pieces are promising, even if they show some limitations.