On line client-wise cohort set selection for speaker verification using iterative normalization of confusion matrices

T-normalization is a widely used method for normalizing the scores in a speaker verification system in order to reduce undesirable variation arising from acoustically mismatched conditions. In this paper we propose a particular form of T-normalization using iterative normalization of confusion matrix generated from impostor trials for each client speaker. The normalized confusion matrix along with a simple distance metric is then used to select a cohort set based on similarity modeling for each client speaker. The normalization statistics thus computed from this cohort set is used for both impostor and claimant scoring. Experiments on the NIST 2004 SRE data demonstrate reasonable improvements in terms of the equal error rate(EER) computed from the detection error trade(DET) curves, when compared to the baseline GMM-UBM schemes. Encouraging improvements in terms of DCF over the general T-normalization schemes are also illustrated for 8C-1C and 1C-1C conversation conditions.