A framework for music similarity and cover song identification

Abstract

This paper presents a framework for music information retrieval tasks which relate to music similarity. The framework is based on a pipeline consisting of audio feature extraction, feature aggregation and distance measurements, which generalizes previous work and includes hundreds of similarity models not previously considered in the literature. This general pipeline is subjected to a comprehensive benchmark of analogously defined music similarity models over the task of cover song identification. Experimental results provide scientific evidence for certain preferred combined choices of features, aggregations and distances, while pointing towards novel combinations of such elements with the potential to improve the performance of music similarity models on specific MIR tasks.