Online Adaptive Score Normalization for Noise Robustness Speaker Verification on Cellular phone

Most commonly used score normalization methods can improve the performance of speaker verification systems, but need extra speech data or cohort models, more memory and computation MIPS. In this paper we present a low-cost adaptive online score normalization (LAOSN) method to improve the performance of speaker verification without any extra data. The computation and memory cost of LAOSN is very small. The procedure begins with initialization of the normalization parameters with existing scores of enrolment utterances from a given enrolment speaker model, and the normalization parameters will be online updated with the scores of subsequent test utterances. By this means, an accurate estimation of the unknown score distribution is archived to normalize current test score. Experiments on the Polycost corpus suggest that the LAOSN can achieve much better performance comparing to the well-known Z-norm method without any extra memory and computation cost.