Line Spectrum Coefficients And Gaussian Mixture Model-based Speaker Recognition Technology

The recent advances of information technology and increasing requirement for security have resulted in a rapid development of intelligent personal identification based on biometrics. Speaker recognition is a branch of biometrics. It has caught the attention of the world by its convenience, economy and accuracy, and it is an inevitable trend of security authentication system. The dissertation performs systemic study on the theories and technologies of small text-independent speaker recognition. In order to maintain unvoiced and voiced sound's different power spectrum structure, segmental wavelet de-noising was used. It could maintain integrity of the speech well. Traditional short time energy and zero-crossing-rate end-pointing detection based on the invariable threshold which has intrinsically limited to vary from the changing of the environment, so dynamic dual-threshold are taken as arguments to renew the thresholds. A better test results are obtained with the adaptive threshold than traditional ways. Linear spectrum pairs has many good properties, such as good filter stability, representational efficiency,desirable interpolation and quantization properties, better characterise both formant locations and bandwidths, and it also has an ordering related to the spectral properties of the underlying data, which leads to advantages when used for analysing speech and other signals. It's spectrum properties were studied, and compared with some of the existing features. Results show that LSP feature parameter has a better performance in presenting speaker features. K-Means algorithm is used to initial EM algorithm, when EM algorithm is utilized to realize Gaussian Mixture Model based clustering, to improve it's performance. In this paper, an experimental speaker recognition system based on LSP feature parameter and GMM model used for the research of text-independent speaker recognition has been developed. Results show that, the system has good performance and practicality. When signal noise ratio is 20dB to 40dB, the recognition rate can reach over 90%.