Optimization of the DET curve in speaker verification under noisy conditions

The increasing need for secure authentication systems has motivated recent interest in effective algorithms for Speaker Verification (SV). In particular, there is increasing need for noise robust algorithms for SV, which will allow SV systems to operate successfully in real conditions, which are typically noisy. Speaker verification addresses a pattern classification problem, in which there is a tradeoff between false acceptance and false rejections. Traditional approaches optimize the parameters of a classifier for a single operating point emobidied by the proportions of positive and negative examples in the training data, or by learning the parameters without considering the tradeoff. In a real situation where noise is present, the operating point is effectively unknown and may not match training conditions. We believe that for such situations the optimization of the parameters should not be limited to a single operating point, and that a more robust strategy is to optimize the parameters for all operating points by minimizing the area under the detection error tradeoff curve. In this paper we investigate the minimization of the area under the detection error curve in noisy conditions. Experiments performed on the database NIST2008 show our method improves the performance with respect to conventional methods.