Cross-Domain Speaker Recognition using Cycle-Consistent Adversarial Networks

Abstract

Speaker recognition systems often suffer from severe performance degradation due to the difference between training and evaluation data, which is called domain mismatch problem. In this paper, we apply adversarial strategies in deep learning techniques and propose a method using cycle-consistent adversarial networks for i-vector domain adaptation. This method performs an i-vector domain transformation from the source domain to the target domain to reduce the domain mismatch. It uses a cycle structure that reduces the negative influence of losing speaker information in i-vector during the transformation and makes it possible to use unpaired dataset for training. The experimental results show that the proposed adaptation method improves recognition performance of a conventional i-vector and PLDA based speaker recognition system by reducing the domain mismatch between the training and the evaluation sets.