Feature extraction/selection is an important stage in every speaker recognition system. Dimension reduction plays a mayor roll due to not only the curse of dimensionality or computation time, but also because of the discriminative relevancy of each feature. The use of automatic methods able to reduce the dimension of the feature space without losing performance is one important problem nowadays. In this sense, a method based on mutual information is studied in order to keep as much discriminative information as possible and the less amount of redundant information. The system performance as a function of the number of retained features is studied.