On optimum normalization method used for speaker verification

Score normalization has become necessary for speaker verification systems, but general principles leading to optimum performance are lacking. In the paper, theoretical analyses to optimum normalization are given. Under the analyses, four existing methods based on likelihood ratio, cohort, a posteriori probability and pooled cohort are investigated. Performance of these methods in verification with known imposters, robustness for different imposters and separability of the optimal threshold from the imposter model are discussed after experiments based on a database of 100 speakers.