Clustering-Based Score Normalization for Speaker Verification

Score normalization can improve speaker verification (SV) performance by adjusting the distribution of test scores to follow a normal distribution. In this paper, all of the imposter scores for the target speakers are first obtained from the normalization cohort; then, these scores are clustered by an unsupervised clustering algorithm, and Gaussian mixture models (GMMs) are used to fit the score distribution. The mean and the standard deviation of the Gaussian component with the maximum mean value is used in the SV score normalization method. Experiments are carried out on the NIST SRE 2016 test set and the VOiCES test set. Compared with conventional score normalization methods, the proposed method can effectively improve SV performance.