Identifying Cover Songs Using Information-Theoretic Measures of Similarity

摘要

This paper investigates methods for quantifying similarity between audiosignals, specifically for the task of of cover song detection. We consider aninformation-theoretic approach, where we compute pairwise measures ofpredictability between time series. We compare discrete-valued approachesoperating on quantised audio features, to continuous-valued approaches. In thediscrete case, we propose a method for computing the normalised compressiondistance, where we account for correlation between time series. In thecontinuous case, we propose to compute information-based measures of similarityas statistics of the prediction error between time series. We evaluate ourmethods on two cover song identification tasks using a data set comprised of300 Jazz standards and using the Million Song Dataset. For both datasets, weobserve that continuous-valued approaches outperform discrete-valuedapproaches. We consider approaches to estimating the normalised compressiondistance (NCD) based on string compression and prediction, where we observethat our proposed normalised compression distance with alignment (NCDA)improves average performance over NCD, for sequential compression algorithms.Finally, we demonstrate that continuous-valued distances may be combined toimprove performance with respect to baseline approaches. Using a large-scalefilter-and-refine approach, we demonstrate state-of-the-art performance forcover song identification using the Million Song Dataset.