Automatic speaker verification (ASV) systems usually use high dimension feature vectors and therefore involve high complexity. However, many of the features used in such systems are believed to be irrelevant and redundant. So far, many wrapper-based methods for feature dimension reduction in these systems have been proposed. Meanwhile, the complexity of such methods is high since system performance is used for feature subset evaluation. In this paper, we propose a new feature selection approach for ASV systems based on ant colony optimization(ACO) and support vector machine (SVM) classifiers which uses feature relief weights in order to have a lower number of feature subset evaluation.
This method has led to 64% feature dimension reduction with a 1.745% Equal Error Rate (EER) for the best case appeared in polynomial kernel of SVM. The proposed method has also been compared with Genetic Algorithm (GA) regarding feature selection task. Results indicate that the EER and the number of selected features for the proposed method are lower for different kernels of SVM.