Abstract: An audio feature extraction method which can improve classification rate by utilizing the class information is proposed in this paper. Contrary to the traditional feature extraction method based on independent component analysis, the mutual information among the dimensions of the feature vectors is computed on each set of data belonging to different class, rather than on the whole training set. The average then is adopted as the contrast function when generating the basic functions of the training space. Experiments show that higher classification rate can be achieved on the new feature vectors extracted by this method, and that the new basic functions produce smaller mutual information among the dimensions of the feature vectors belonging to the same class.