How to Reduce Dimension while Improving Performance

Abstract

This paper addresses the feature subset selection for an automatic Arabic speaker recognition system. An effective algorithm based on genetic algorithm is proposed for discovering the best feature combinations using feature reduction and recognition error rate as performance measure. Experimentation is carried out using QSDAS corpora. The results of experiments indicate that, with the optimized feature subset, the performance of the system is improved. Moreover, the speed of recognition is significantly increased, number of features is reduced over 60% which consequently decrease the complexity of our ASR system