Automatic music summarization methods using two stage clustering and segment similarity

Abstract

In this thesis, we present effective methods for music summarization which automatically extract a representative portion of the music by signal processing technology. One proposed method, which has the framework of two-stage clustering, uses 2-dimensional similarity matrix, tempo tracking, and clustering techniques to extract several segments which have different moods or dissimilar semantic structure in the music. The other method proposed uses segment similarity based on identity of codeword indexes of each music frame. The method aims to find repetitive structure of music and to utilize a process of human perception which recognizes a hook of music. The segments extracted are combined to generate a complete music summary in both two methods proposed. Experiments show the proposed methods capture the main theme of the music more effectively than a conventional method. The experimental results also show that one of the proposed methods could be used for real-time application since the processing time in generating a music summary is much faster than other ones.