Abstract: In order to improve the performance of speaker recognition, a method for optimal speaker acoustic features selection is proposed, using greedy algorithm to select some dimensional features from a large dimensional feature set in turn. The recognition rate is increased by 1.6% on pure voice than traditional method and by 6% on noisy voice, the recognition speed is increased about 5 times in experiments based on TIMIT. The experiment results show that the optimal selective features which eliminate disturbance of other redundant features can effectively improve both the recognition accuracy and the computational speed on the speaker recognition system.

Key words: speaker recognition, feature selection, Gaussian mixture model (GMM), greedy algorithm, Mel frequency cepstrum coefficient（MFCC）