The RU Submission to the Evalita’09 “application track” Speaker Recognition Evaluation

In this paper the RU submissions for the application track of the Evalita’09 speaker recognition evaluation are presented. The primary submission is a fusion of two independently developed acoustic systems that include channel compensation and full joint factor analysis (JFA), respectively. Both systems are UBM-GMM dotscoring systems, which are characterized by very efficient computation of scores. Of the two systems the JFA approach performs best on its own. All components for this system are trained on the Evalita’09 training data, using the UBM data for UBM, channel and speaker factor traning, Zand Tnorm cohorts, and using the development test trials for calibration. The plain dotscoring system with channel compensation uses an externally trained UBM, and utilizes the Evalita’09 UBM training data for channel compensation training. On the evaluation trials, the fused system obtains equal error rates ranging from 13.45% (TC1, TS1) down to 1.64% (TC6, TS2).