A new method of text-independent speaker recognition using discriminative feature selection is proposed in this paper. The method is compared with the classical UBM-GMM and hybrid codebook approaches. The characteristics of the proposed method are as follows: feature parameters extraction, vector quantization with the growing neural gas (GNG) algorithm and discriminative feature selection (DFS) according to the uniqueness of personal features. The speaker recognition algorithms are evaluated on three databases: A2000, B2000 and King. The PLP cepstral coefficients have been used together with four channel normalization techniques. The test results showed that channel normalization techniques greatly improve the speaker recognition performance and they showed that use of discriminative features selection procedure yields significant performance compared with classical UBM-GMM and codebook hybrid approaches.