Pruning abnormal data for better making a decision in speaker verification

In development of a speaker verification system, a priori threshold estimation is often needed based on a training set for decision making. Such a threshold critically determines performance of a speaker verification system. From a statistical viewpoint, a speaker's voice could be modeled by a certain distribution. Thus, data for training are only some samples of this distribution in a subspace, and the statistical information acquired from the training set is usually biased to that of the whole space. In this paper, we propose a method for better estimation of underlying statistics by abnormal data elimination. Without use of more data, our method provides an alternative way to improve performance of those statisticsbased a priori threshold estimation methods in terms of generalization capability. On the basis of a benchmark database, KING, and a baseline system with a priori threshold estimation, we demonstrate the effectiveness of our method.