In order to realizethe text-independent speaker recognition and improve its accuracy rate, avariable method in combination with Gaussian mixture model （GMM）and support vector machine （SVM） is used in this paper. By vectoring GMM mean parameters, we use SVM to recognize the real speaker of the testing speech. We put forward two methods of building hyper plane between two kinds of training speeches and another way to build the hyper plane between Universal Background Model （UBM） and the training speech. Upon vectoring the GMM mean parameters, both of the two methods show better performance in speaker verification thathe traditional GMM-UBM withless accuracy. The vectoring GMM mean parameters can amplify the characters of speakers and thus make verification between two speakers in a more obvious manner.