Automatically Correcting Bias in Speaker Recognition Systems

In this paper we present a general machine learning framework for score bias reduction and analysis in speaker recognition systems. The general principle is to learn a meta-system using recognition systems' errors, given the training and testing conditions in which they occurred. In the context of speaker recognition, the proposed method is able to reduce the bias introduced in scores due to a variety of factors such as channel mismatch, additive noise, gender mismatch, different speaking styles, etc. Moreover, this framework enables a deep understanding of the origins of score bias in any system, which will support an optimized system redesign. Preliminary results obtained with several state-of-the-art systems showed considerable improvement in original performance, in addition to identifying sources of system bias.