In this paper we introduce two methods to improve text-independent speaker verification. In feature extraction process the feature vectors of voiced speech and unvoiced speech independently. In test process, the test speech is adapted to a new model instead of calculating the log-likelihood, then the Mahalanobis Distances among the UBM model, the speaker model and the test speech model are calculated. These three models distances formed a triangle. The angle of the models can be obtained as the test scores. Further more the scores of the log-likelihood and the angle of the models can be fused to improve performance. When we employ the proposed algorithms, the EER of the speaker verification can be reduced by 20% compared with the baseline system.