In speaker recognition systems, if the original data distribution is unknown, the choice of inappropriate kernel functions will result in poor support vector machine (SVM) learning performance. Thus a speaker recognition method based on a multi-grid search of parameters and a combination of kernel functions is proposed in this paper. First, the method constructs a hybrid kernel function by linearly weighted polynomial and RBF kernels. Then it proposes a multi-grid search method to adjust the weights, and thus the hybrid kernel function can adapt to the current data distribution. Finally, a SVM classifier is trained to obtain the classification results. Simulation experiments on TIMIT datasets and noisy datasets show that the recognition performance of SVM classifiers using a combination of kernel functions is better than that using linear kernels, polynomial kernels, and RBF kernels. Therefore, the proposed method can effectively improve the performance of speaker recognition systems.

speaker recognition  /  support vector machine  /  combination of kernel functions  /  multi-grid search