Research Of Self-adapting Algorithm For Speaker Recognition

The substance of this magisterial thesis is the research and improvement of speaker recognition which is based on the VQ (Vector Quantization) and HMM (Hidden Markov Model).With cheerful prospect, speaker recognition is a biometrics that recognizes people via their voice, and the combination of VQ and HMM is one of the best and most prevailing method in the field of speaker recognition.The author widely studies the knowledge of every part of speaker recognition which involves voice signal preprocessing, character vector extracting and recognition method based on different model.During this work, the main body is the recognition method, namely, recognition algorithm. Three aspects of it are as follows:(l)Study of speaker recognition based on VQ: It is applicable to SD(Speak Dependent) recognition but unable to deal with the variety in voice character which result from different speaker.(2)Study of speaker recognition based on HMM: As a kind of statistical model, it is applicable to SI(Speak Independent) recognition because it includes the variety in voice character which result from different speaker.(3)Study of speaker recognition based on FVQ(Fuzzy VQ)/HMM: It is the special form of HMM. Compared with original HMM, it has less parameters to reduce training data for learning, higher constringency speed of learning to be applicable to real-time self-adapting learning and higher recognition speed to be applicable to real-time continuous voice recognition with large vocabulary. Compared with original division VQ, it has better effect of division and less quantization error of codebook by FCM(Fuzzy C-Means) clustering analysis.