Score Normalization of X-Vector Speaker Verification System for Short-Duration Speaker Verification Challenge

摘要

In this paper we present our contribution to the task 2 of the short-duration speaker verification (SdSV) challenge. The main task for this challenge is to find new technologies for text-dependent and text-independent speaker verification in short duration scenario. Some of the approaches used by the authors during participation in the challenge are presented. Described speaker verification systems include baseline x-vector system with PLDA backend and score normalization, x-vector system with neural PLDA backend and fusion of both systems. The main goal of this paper is to analyze influence of different score normalization methods on x-vector based speaker verification systems performance. We found that system with PLDA backend and ZT-normalization method (single system) gives superior performance in Farsi trials, but gives lower performance improvement in English trials. Overall, in terms of minDCF single system performs 46.3% better than baseline x-vector system. We found that enroll data augmentation is useless for Neural PLDA backend, as performance of the system does not improve after adding augmented enroll data. Single system with ZT-score normalization and additional enroll audio augmentation performs 14.8% better than Neural PLDA backend system.