Highly Efficient Dimension Reduction for Text-Independent Speaker Verification Based on Relieff Algorithm and Support Vector Machines

Automatic speaker verification (ASV) systems are among the biometric systems used in security and telephone-based remote control applications. Recent years have witnessed an increasing trend in research on such systems. These systems usually use high dimension feature vectors and therefore involve high complexity. However, there is a general belief that many of the features used in such systems are irrelevant and redundant. So far, many methods for feature dimension reduction in these systems have been proposed, most of which are wrapper-based and thus computationally expensive since system performance is used for feature subset evaluation. This involves system training and performance evaluation for each feature subset, which is a time consuming task. In this paper, we propose a feature selection approach based on Relieff algorithm for ASV systems using support vector machine (SVM) classifiers. This method is wrapper-based but makes use of Relieff weights in order to have a lower using of system performance. Thus this method has lower complexity compared to other wrapper-based methods, can lead to 69% feature dimension reduction and has a 1.25% of Equal Error Rate (EER) for the best case that appeared in RBF kernel of SVM. The proposed method has been compared with Genetic Algorithm (GA) and Ant Colony Optimization (ACO) methods for feature selection task. Results show that the EER, number of selected features and time complexity of the proposed method is lower than these methods for different kernels of SVM.