Music Automatic Classification Based On Support Vector Machine

In essence, the music classification is a branch of audio classification, and has become a hot topic in pattern recognition, the direction of its research and development in general can be summarized in three aspects: first, the methods of music feature extraction and the form of feature vectors can be make improvements; second, the choice of classifier can do to improve; third, to solve the problem of multiple classifiers can do to improve on.In this dissertation, experiments have used support vector machine to classify the various music, which including pop music, classical music, piano music, folk songs, vocal, opera in six different styles, and experimental work are summarized as follows:First, by learning the theory of digital audio technology to master the short audio processing technology, combined with actual conditions, each music sample will be gone through the short-term pretreatment processes of pre-emphasis, framing, window, identification of silent frame. And then, to extract music samples'characteristics which are in time domain, frequency domain and extract the MFCC matrix as well as the statistical characteristics of this matrix.Second, after deep understand the principle of SVM classification, I made a comparison between standard SVM and SVM variants, and decided to choose support vector machine as the classifier, then, I tested the performances of classifier which was in different cases including MFCC in different dimension, different features of MFCC statistics, different combinations of MFCC and Perceptual characteristics, In addition, I made a comparison between SVM and other classifiers under the same conditions defined its advantages.Third, I have studied the application of multi-classification, analyzed advantages and disadvantages of multi-class classification methods and have proposed the method of making use of prior knowledge combined with the tree structure to construct the multi-classification system. I used standard SVM on training samples to get a lot of classifiers which were used in two types, moreover, analyzing these classifiers to get their classification accuracy. According the classification accuracy, I optimally clustered samples in different areas. Then to create the classify system in the tree of non-overlapping structure and tested results of the system's classification accuracy.