Cross-Similarity Measurement of Music Sections: A Framework for Large-scale Cover Song Identification

Abstract

For large-scale cover song identification, most previous works take a single feature vector as the representation of a song. Although this approach ensures structure invariance, it may cause overcorrection since it totally neglects the structure feature of the song. To address this problem, we put forward a novel framework for large-scale cover song identification based on music structure segmentation, aiming at matching the irrelevant sections and ignoring the irrelevant ones. In our implementation, we apply the average and weighted average methods to integrating similarities of section pairs. We evaluate the proposed framework based on three representative previous methods, including 2D Fourier magnitude coefficients, chord profiles, and cognition-inspired descriptors. The experimental results show that the all the three methods in our framework significantly outperform those in their original works.