Research On Music Classification Algorithm Based On Deep Belief Network And Hidden Markov Model

With the development of information and multimedia technologies, digital music has been widely used in different media, including radio broadcasting, digital storage, the Internet, etc. Therefore, efficiently retrieving and managing music develop rapidly in recent years. Music classification is an important research topic and the performance optimization on music classification is even more important.This master thesis presents a music classification algorithm based on the combination of Deep Belief Network(DBN)and Hidden Markov Model(HMM). First, we extract multi-features consisting of music timbre, rhythm, and mood. Then the DBN-HMM model is used to map these features to a new feature space, and calculate the probability of each music category. Finally, the music is classified to the largest probability.The innovation points are as follows:(1) Replace Mel-frequency cepstral coefficients with Mel-Phon coefficients, which is more suitable for applying and training the multi-layer network model.(2) Combine the Mel-Phon coefficients with the music mood feature vector as the input music feature vector, which reflects the music information more precisely.(3) Obtain a modified network classification model with more efficiency and higher accuracy based on the combination of the Deep Belief Network(DBN) and the Hidden Markov Model(HMM)。In order to verify the performance of this algorithm, ten categories music in GTZAN database are used for test. The test results show that the classification accuracy is 92.4%. The comparison with the existing algorithms shows that both the proposed feature vector and the classifier can improve the performance of the music classification system.