The Research Of Feature Extraction Based On Fisher Criterion In Speaker Recognition

Speaker recognition is a technology which recognizing the identity of thespeaker by speaker’s voice. It has a broad application prospect in financial, legal,medical, military etc., and has become a hot topic in biometric authenticationtechnology. With the rapid development of computer and information technology,speaker recognition is gradually from the laboratory to practical application.As one of the key technologies in speaker recognition, feature extraction is theresearch focus of many domestic and overseas scholars. The practical environment isa challenge for speaker recognition, therefore extract the feature which can bettercharacterize the speaker’s personality, be more robustness and get better recognitionhas become a problem to be solved. Based on the research of some classic featureextraction algorithms and the feature selection method, a mixed feature extractionalgorithm based on Fisher criterion and a Gammatone feature extraction algorithmbased on Fisher criterion are proposed, and simulated in the MATLAB platform. Themain research works and achievements are described as follows:Firstly, some related knowledge of speaker recognition system was introduced,such as the basic principle, system structure and the performance evaluation standard.Then studied LPCC(Linear Prediction Cepstral Coefficient) and MFCC(MelFrequency Cepstral Coefficient), the most commonly features’ extraction algorithm,and given a comparative analysis of their advantages and disadvantages. Further more,the feature selection method based on Fisher criterion and its application in speakerrecognition was described. At the same time, GMM(Gaussian Mixture Model) whichused in this thesis as the recognition method was also introduced.Secondly, this thesis analyzed the extraction process of MFCC andIMFCC(Inverted Mel Frequency Cepstral Coefficient), especially the structure of Melfrequency filter banks and the inversed Mel frequency filter banks. In view of theadvantages and disadvantages of MFCC and MFCC, a mixed feature extractionalgorithm based on MFCC and IMFCC by using fisher criterion was exploited. Thesimulation experiments were conducted on TIMIT speech database and NOIZEUSspeech database, and the results showed that the improved method has a higherrecognition rate compared with MFCC and LPCC.Finally, because of MFCC has poor recognition in the noise environment, so weintroduced the Gammatone filter bank, which imitated the human ear hearing system, to replace the Mel frequency filter banks. But the Gammatone feature has too muchdimensions and too complicated calculation, in addition, each dimension’sdiscrimination is different. Therefore, based on the feature selection and dimensionreduction of Fisher criterion, this thesis proposed a Gammatone feature extractionalgorithm based on Fisher criterion. The simulation experiments were conducted onNOIZEUS speech database. Compared with Gammatone feature, the improved methodneeds less dimensions and has a better recognition rate.