Adversarial separation and adaptation network for far-field speaker verification

Abstract

Typically, speaker verification systems are highly optimized on the speech collected by close-talking microphones. However, these systems will perform poorly when the users use far-field microphones during verification. In this paper, we propose an adversarial separation and adaptation network (ADSAN) to extract speaker discriminative and domain-invariant features through adversarial learning. The idea is based on the notion that speaker embedding comprises domain-specific components and domain-shared components, and that the two components can be disentangled by the interplay of the separation network and the adaptation network in the ADSAN. We also propose to incorporate a mutual information neural estimator into the domain adaptation network to retain speaker discriminative information. Experiments on the VOiCES Challenge 2019 demonstrate that the proposed approaches can produce more domain-invariant and speaker discriminative representations, which could help to reduce the domain shift caused by different types of microphones and reverberant environments.