This Thesis aims to develop new classifier combination techniques applied in sub-bands in order to improve the robustness of text-independent speaker identification systems. The advantages observed in previous experiments using multiple classifiers in sub-bands for robust speaker recognition motivated the development of combination techniques for these algorithms. New strategies to combine the classifiers responses are proposed in this Thesis. The main purpose is to increase the recognition performance in situations when there is no knowledge about the type of noise that corrupts the testing speech signal. The different proposals consist in applying non-uniform weights, null space, multicondition training, dynamic features and autocorrelation based MFCC features. The employment of the new strategies significantly contribute to increase the recognition performance. It was obtained an increase, for instance, compared to the Sum combination technique shown in the literature, of about 47% in tests with white noise, and 32% with non-white noise in 15 seconds of speech in 10 dB of SNR (Signal-to-noise ratio), just using a new strategy which employ the null space to combine the sub-band classifiers. Even better results were obtained by using the other proposals.