Analysis of mutual duration and noise effects in speaker recognition: benefits of condition-matched cohort selection in score normalization.