Music Structure Analysis Based On Timbre Unit Distribution

Music structure is not only an important form of the music works to express artists'ideas, but also an effective way for the listener to understand the meaning of the music. The song structure generally comprises of Introduction (Intro), Verse, Chorus, Bridge and Ending (Outro). Therefore, the aim for music structure analysis is to segment a piece of music into small sections and to label each of them. Music structure analysis is an important and fundamental field in music understanding research, which plays an important role in many fields.There are two principle approaches for the music structure analysis, namely "state" and "sequence" approach. This paper proposes a timbre unit modeling method based on musical features, using unsupervised clustering method to analyze music structure according to the distribution of local timbre units, with the amended Fisher rule.The thesis contains the following works and contributions:1. Timbre unit modeling method based on musical features is used to analyze music structure, which comprises of three parts:preprocessing, timbre unit analysis, and structure analysis. State based approach reduces fragmentations in audio segmentation using clustering technique.2. DCT-based Chroma feature is used in music structure analysis, which indeed gains a significant boost towards timbre invariance.3. A histogram clustering method based on timbre unit distribution is used to get the optimal clustering result using the amended Fisher rule.4. A candidate boundary analysis method based on Bayesian Information Criterion method is proposed to locate the optimal boundary in order to optimize the histogram clustering result.5. Sequence based approach used similarity matrix to analyze the music structure. We notice that whether the selected threshold is suitable or not will directly affect the completeness of the similar lines, which will influence the segmentation results of music structure. While our method proposed in this thesis will find all of the similarity lines. The experiment results show that the unsupervised clustering algorithm with DCT-based Chroma feature is an effective way to analyze the music structure.