Research On Text-independence Of Open-set Speaker Recognition

With the development of information technology, identity recognition plays a more and more important role in the field of information security insurance. Due to its stability, uniqueness and convenience, biological recognition becomes a crucial research direction of identity recognition. The technology of speaker recognition is a branch of biometrics technology, the important point of technology is distinguish speakers, according to difference of personality characteristics among speakers. Today, Text-dependence of closed set speaker recognition has performed quite perfect, but the recognition rate of open-set is lower. Open-set recognition involved with speaker identification and threshold verification, so speaker identification and threshold verification is very important. Aiming at Text-independence of open-set speaker recognition technology, this thesis analyzes the principle and framework of speaker recognition system in detail, it also researches on speech endpoint detection,feature extraction,open-set identification,threshold verification. The main work and contributions of the dissertation are in the several aspects as follows:(1) Concerning with preprocessing and speech endpoint detection, Firstly, we analyze Significance of preprocessing and speech endpoint detection. Secondly, we discuss temporal energy endpoint detection algorithm,temporal zero-crossing rate endpoint detection algorithm,temporal energy frequency value endpoint detection algorithm,basic spectral entropy endpoint detection algorithm and improved spectral entropy endpoint detection algorithm from theory. Finally, we simulate experiment on preprocessing and these endpoint detection algorithms, and compare the advantages and shortcomings of endpoint detection algorithms. Based on smaller of speaker pronunciation in laboratory conditions, this thesis applies improved spectral entropy endpoint detection algorithm.(2) On the side of speech feature extraction, we analyze the principle of combined feature extraction and principal components analysis, this thesis researches feature extraction method which combines PCA theory with combined feature extraction. The experiment results show that the method has obvious performance in recognition rate. At the same time, it decreases the amount of operation.(3) Concerning with open-set identification, Aiming at the limitations of the traditional vector quantization algorithm, this thesis researches a open-set identification method which combines fuzzy c means algorithm with PCA theory. The experiment results show that the method has better recognition rate than FCM,VQ+PCA,VQ identification methods.(4) On the side of open-set recognition, we discuss the traditional threshold algorithm,based on the range of the distortions algorithm,dynamic thresholds algorithm and RS threshold algorithm. Aiming at the Shortcomings of these threshold algorithms, this thesis researches two open-set verification methods which combines fuzzy c means algorithm with dynamic thresholds algorithm and combines fuzzy c means algorithm with RS threshold algorithm. The experiment results show that the two combined methods reduce the equal error rate partly, comparing with these threshold algorithms.