Cultural Style Classification of Music Signals

Abstract

Music classification based on cultural style is useful for automatic music analysis and has potential applications in music retrieval systems. This project is to develop classification techniques for classifying Chinese music and Cantonese popular music. The project is divided into 2 parts, the first part is about investigating the musical styles and features of the different music samples and the second part is to extract the useful data from each music sample for the computer learning. Music features can mainly be divided into two groups: one is timbre features, the other one is rhythm features. Rhythm feature like peak and valley count; timbre features like zero crossing and spectral flatness. Both of them are being investigated. By plotting the graph for each feature of different music, we can find out those features which are useful for classifying pieces into the corresponding musical style. The 2nd part is to extract those useful data from the music to train the HMM models. The HMM model used has two groups of tied output probabilities. The output probabilities are drawn from a finite set of 256 values, each corresponds to a codeword symbol in a VQ codebook. Two HMM models are trained: One is for Chinese music and the other one is for the Cantonese popular music. Experimental results show that the proposed method can achieve an accuracy of 85%, which prove that music audio signals can be classified according to their culture styles using machine learning methods.