Research On Content Based Music Feature Extraction And Classification

With the development of data mining and machine learning,more and more data can be processed by computer with artificial intellegence.Plenty of achievements have been reached in the field such as text processing,pattern recognition,and speech processing.However,in the field of music signal processing,further research are still needed.Based on the theories of time domain signal processing,frequency domain signal processing and tinme-frequency domain signal processing,a new music feature extraction method is proposed in this paper.And further classification and clustering research of this paper is on the feature of this method.The main difference between this method and traditional audio signal processing method is,through signal processing theories,traditional audio signal processing method calculate the specific index of signal such as centroid of frequency spectrum,bandwisth,average energy.The method in this paper process the signal with wavelet analysis theory from a time-frequency domain perspective,extract the main feature of music by the method of matrix singular value decomposition,and describe the tone of music by Mel-frequency cepstral coefficient.Additionally,a new algorithm is proposed in this paper that can track the beats in music from a time domain perspective which is used for describing the beats feature of music.The reason is traditional signal processing method can not well used in music signal,obviously music signal is non-stationary signal and traditional signal processing theory can only apply to stationary signal,and can barely show the general view of the signal.To prove the validity of the feature extration method,distance measurement method is used in this paper on the data bank which is consist of 6 different style of music with different musical instrument.The 6 different music style are acoustic music with guitar,hip-hop music,piano music,relaxing music,rock music,and pure vocal music.Based on the feature extracted by the method in this paper,both logistic regression and support vector machine classification algorithm show a high precision around 95%.The result shows this method can markedly reduce the data size of music,and make music computable.Furthermore,this method can be widely used in content based recommendation,retrieval,classification and cluster system,also it has good prospect in music data managerment,analysis and music service field.