Segmentation of Pop Music Based on Histogram Clustering

Abstract

As a method of music segmentation in music structure analysis, segmentation of music based on histogram clustering is to detect and discover musical repetition patterns on the basis of simulating human auditory perception. The selection of clustering algorithm is an important factor affecting the segmentation accuracy. In this paper, a music segmentation method based on histogram clustering is implemented. The beat-based pitch class profile (PCP)feature is selected, and the pop music is segmented according to the music structure through similar feature vector clustering, histogram clustering and marginal adjustment. The best parameters of histogram clustering were obtained by parameter optimization experiment. Using the traditional K-means, K-means++ and Isodata clustering algorithm, 200 Chinese pop songs were segmented and the performance of K-means++ algorithm was the best with an average accuracy of 71.34%. The results show that although the K-means++ algorithm has an increase in segmental redundancy, the average accuracy is greatly improved and the time complexity is lower, so it is more suitable for music segmentation based on histogram clustering.