In the problems of practical application, GMM - UBM is adopted as the background model when the training data is insufficient in speaker recognition system. EigenVoice is used as adaptation ways, then it structured a new combination kernel function combined with homogeneous polynomial kernel with good generalization ability and radial basis kernel function with good earning ability by linear weighted method to optimize model parameter. The optimal parameters of kernel function are determined through the multiple grid search method. DAG method is adopted to realize multivariate classification of SVM kernel function. Then the linear kernel, homogeneous polynomial kernel, radial basis kernel function and combination kernel function are evaluated in the experiments. The experimental results show that the identify performance of the combination kernel is more ideal than that of other kernel functions in the different classification strategy, different adaptive time, different signal-to-noise ratio and different number of speakers.