A comparison of symbolic similarity measures for finding occurrences of melodic segments

Abstract

To find occurrences of melodic segments, such as themes, phrases and motifs, in musical works, a well-performing similarity measure is needed to support human analysis of large music corpora. We evaluate the performance of a range of melodic similarity measures to find occurrences of phrases in folk song melodies. We compare the similarity measures correlation distance, city-block distance, Euclidean distance and alignment, proposed for melody comparison in computational ethnomusicology; furthermore Implication-Realization structure alignment and B-spline alignment, forming successful approaches in symbolic melodic similarity; moreover, wavelet transform and the geometric approach Structure Induction, having performed well in musical pattern discovery. We evaluate the success of the different similarity measures through observing retrieval success in relation to human annotations. Our results show that local alignment and SIAM perform on an almost equal level to human annotators.