Study On Feature Extraction In Speaker Recognition

Speaker recognition as one of the biometrics techniques is to recognize speaker's identity from its voice which contains physiological and behavioral characteristics specific to each individual. Speaker recognition has caught many attentions for its particularly advantage on convenience, economy and veracity and become an important and popular authentication technique in human life and work. Therefore, a more robust method for speaker recognition with high accuracy of recognition rate is the aim for researchers at home and abroad. This paper focuses on the speaker recognition system based on Mel-frequency Cepstrum Coefficients (MFCC) and Gaussian Mixture Model (GMM), which are combined with the methods of F ratio, Principal Component Analysis and Independent Component Analysis. In this paper, the following research work is pursued:(1) Compared the system performance using the features of Linear Predictive Cepstral Coefficients (LPCC), Adaptive Components Weighting (ACW) and MFCC, and proved that by using MFCC a higher accuracy of recognition rate is obtained.(2) The MFCC feature with normalized short time energy and dynamic information is discussed based on the MFCC feature and the influence to the identification performance is analyzed. F ratio as an evaluation method for feature is used to choose feature components. F ratio is proved to be efficient from the results of experiments.(3) Choose the most contributive feature components for the recognition performance by using F ratio method, and a weighted F ratio method is proposed based on the feature of MFCC.(4) Principal Component Analysis and F ratio method are introduced into the MFCC feature extraction, which the correlation information between components is cancelled and the most differential components are chosen.(5) Independent Component Analysis (ICA) and F ratio method are introduced into the MFCC feature extraction, during which the redundancy information between components is taken out and the most contributive feature components for the identification performance are chosen.The approaches mentioned above are contrasted through simulation results, and elaborate analysis and the conclusion are given. Finally the expectation for feature work is suggested.