Abstract: Channel interference factor for the identification results is prevalent among the existing speaker recognition algorithm. In order to improve the accuracy of the system, in this paper, feature warping is used to compensate the channel factor of Mel-Frequency Cepstral Coefficient (MFCC) features. Then, factor analysis technique is applied to deal with the channel factors of the speaker's Gaussian Mixture Model (GMM). In the endpoint detection phase of speech of this recognition system, the GMM for speech modeling is built to accurately determine the beginning and end points of the speech segment, and then the features after feature warping are used to establish speaker GMM. Using factor analysis technique to fit the differences between the speaker characteristics space and the channel space, the algorithm removes channel factor from GMM, and then extracts GMM super-vectors as input of the Support Vector Machine (SVM) to obtain recognition results. Experimental results show that the combination of channel compensation technique and SVM can obtain better recognition rate, and ensure the robustness of the system.

Key words: computer application, speaker recognition, support vector machine(SVM), Gaussian mixture model(GMM), feature warp, latent factor analysis(LFA)