Research Of Music Classification Algorithm Based On Linear Discriminative Analysis And Support Vector Machine

Along with the development of technology in Internet and broadcast, there are great opportunities for people to have access to the large quantities of multimedia contents. But since the fast growing of the data volume, how to manage the contents automatically has emerged as an urgent problem. Especially to the all kinds of music signals around us, fast and efficient methods are required to classify and manage them(according to different styles or singers). This thesis is in hope of finding a better algorithm to solve this problem.Based on the existing music classification architecture, this thesis propose an improved music classification architecture, adding LDA module to the former one to perform dimensionality reduction on the original high-dimensional vector, also use SVMs classifier in the final classification stage, then simulate the classification results by Matlab software.Most of the contemporary algorithms for audio signal classification include two stages: feature extraction stage and classification stage. Lots of music features can be applied to implement this algorithm, including the short-time energy and short-time zero-crossing-rate etc. from the time domain, the bandwidth and brightness etc. from the frequency domain, also the MFCC(Mel-frequency cepstral coefficients) coefficient which is based on the perception. And the many high efficient algorithms in the Pattern Recognition and Pattern Classificatin such as Gaussian Mixture Model(GMM)[29],Neural Network(NN),Hidden Markov Model(HMM) etc. can be utilized to implement the classification.When facing such many features and classification algorithms, how to combine them to achieve a better classification accuracy rate? Is it possible to do some preprocessing on some of the features or do some optimization on the classifiers base upon the speciality of music classification to achieve a higher classification accuracy rate? To answer these questions, this thesis propose a new music classification method base on the many already existing ones.The now existing music classification methods all isolate the two stages of feature extraction and classification, the extracted features are directly passed to the classifiers for classification, but have not took into account the fact that the already extracted features may not be the best ones for classification(the feature points representing the feature vectors are not the most separable in the high- dimensional space), it's probable to achieve more separable music features by performing some linear or non-linear transforms. This thesis utilize a new music...