An efficient DTW-Based approach for melodic similarity in flamenco singing

Abstract

We study melodic similarity in flamenco singing by using the Dynamic Time Warping (DTW) distance. Given two melodic contours, the score of the alignment of the two melodies is taken as a similarity measure. Concretely, we consider a particularly representative flamenco repertoire, the tonás, a cappella flamenco singings with free rhythm and high degree of complex ornamentation. We show that the DTW-distance discriminates correctly variations between the styles. In order to speedup the quadratic time and space complexity of the standard DTW, our strategy is to perform an efficient segmentation on the pitch contour before applying dynamic programming. We show that our method achieves better results (both in efficiency and accuracy) than other existing DTW-based similarity measures.