Playing in Time: Integrating Temporal Information in the Computational Measurement of Musical Similarity

Abstract

In this article we discuss the importance of temporal information in assessing musical similarity and we present a content-based approach that emphasizes the sequential repetition of perceptually relevant, expressive musical features. To examine these approaches, we implemented three previously proposed MFCC-based music similarity assessment systems and compared their results with those obtained from the proposed new system. We employed human listeners to provide a “ground-truth” assessment of musical similarity for a body of 60 songs spanning 14 musical styles. Analysis of this body of music by the four methods showed higher correlation of human listener similarity assessments with the machine results for the ordering of musical features has the potential to increase performance of music similarity models.