Using post-classifiers to enhance fusion of low-and high-level speaker recognition

This paper proposes a method for automatic correction of bias in speaker recognition systems, especially fusion-based systems. The method is based on a post-classifier which learns the relative performance obtained by the constituent systems in key trials, given the training and testing conditions in which they occurred. These conditions generally reflect train/test mismatch in factors such as channel, noise, speaker stress, etc. Results obtained with several state-of-the-art systems showed up to 20% decrease in EER compared to ordinary fusion in the NIST'05 Speaker Recognition Evaluation.