Towards better making a decision in speaker verification

Speaker verification is a process that accepts or rejects the identity claim of a speaker. How to make a decision is a critical problem; a threshold for decision-making critically determines performance of a speaker verification system. Traditional threshold estimation methods take only information conveyed by training data into consideration and, to a great extent, do not relate it to production data. It turns out that a speaker verification system with such threshold estimation suffers from poor performance in reality due to mismatches. In this paper, we propose several methods towards better decision-making in a practical speaker verification system. Our methods include the use of additional reliable statistical information for threshold estimation, elimination of abnormal data for better estimation of underlying statistics, and on-line incremental threshold update. To evaluate the performance of our methods, we have done simulations based on a baseline system, Gaussian Mixture Model, in both text-dependent and text-independent modes. Comparative results show that in contrast to the recent threshold estimation methods our methods yield considerably better performance, especially on miscellaneous mismatch conditions, in terms of generalization. Thus our methods provide a promising way for real speaker verification applications.