A novel feature extraction approach based on PCA and improved Fisher score applied in speaker verification

Support vector machines (SVMs) have been applied in speaker verification successfully. But they cannot easily deal with the dynamic time structure of audio data, since they are constrained to work with fixed-length vectors. In this paper, we propose a new feature extraction approach based on PCA and improved Fisher score for the sake of solving this problem existed in SVM. This new feature extraction approach can not only map the variable length sequence to a single point in fixed-dimension space, but also can improve the speed of speaker verification system. Improved Fisher score will help a lot in preserving the original information of the utterances and will finally lead to high classification accuracy. Experimental results show the superior performance of this proposed method in comparison with the main SVM system in reducing EER from 9.18% to 4.57%.