Study Of Channel Compensation Based On Speaker Recognition

Speaker recognition technology has a widely application future in field of biometrics recognition, owing to its special advantages in facility, economy and accuracy and so on . Although in the lab environment, speaker recognition system has achieved great success, its performance deteriorate obviously in the practical application . In order to enhance the practicality of the system, it needs to find the solution of the channel robustness.How to raise the channel robust of the speaker recognition system is researched based on channel compensation. In the side of channel compensation, considering the channel could affect the residues, the paper combines the normalization of residues and CMS, comprehensively considers the channel linear and nonlinear influence on the voice signals, and obtains a new joint characteristics compensation method based on channel compensation. On the basis, we use the GMM model and speaker adaptive model algorithm, and judge based on the score. First we use the Log-Likelihood score judgment for those who have cluster comparatively great differences and decide the category. Then we do the Z-norm judgment for the similar features of category assignment, and through test to obtain the decision threshold, and further improve the recognition effect.In addition, the paper successfully realizes the soft development of the telephone speech recognition system. The system, to be working well, has been verified through experimental study. After using the joint compensation in the same channel, the recognition ratio could reach over 98%, and over 67% when channel being different, improved by 17% to the baseline system.