Research on music classification has been processing years, but the performance of each method is not very well. This paper proposes a new method based on entropy and support vector machine for music classification. It uses bank of filters to decompose the music clip into different channels. Then the filters turns it into spectrum through discrete Fourier transform and compute the information entropy and uses support vector machine training and testing on a dataset containing four categories of music. The experiment compares three different kinds of filters, among which the Bark filter achieves an accuracy of 80%. The result shows that the proposed feature vector is better than MFCC.