Self-Similarity Analysis Applied on Tempo Induction from Music Recordings

Abstract

This paper presents a self-similarity analysis approach to tempo induction, assuming that tempo remains approximately constant throughout the music recording. The proposed method is based on the observation that rhythmic characteristics of the music signal manifest themselves as inherent periodicities that can be extracted by processing the diagonals of the self-similarity matrix. Such periodicities can then be processed in pairs to yield a pair of tempo candidates. The method was submitted to the MIREX 2006 Tempo Extraction contest, where it was investigated whether the returned tempi are related to perceived tempi extracted from ground truth data. In this paper results are also reported for a music corpus assembled by the authors.