Replay detection using CQT-based modified group delay feature and ResNeWt network in ASVspoof 2019

Abstract

Automatic Speaker Verification (ASV) technology is vulnerable to various kinds of spoofing attacks, including speech synthesis, voice conversion, and replay. Among them, the replay attack is easy to implement, posing a more severe threat to ASV. The constant-Q cepstrum coefficient (CQCC) feature is effective for detecting the replay attacks, but it only utilizes the magnitude of constant-Q transform (CQT) and discards the phase information. Meanwhile, the commonly used Gaussian mixture model (GMM) cannot model the reverberation present in far-field recordings. In this paper, we incorporate the CQT and modified group delay function (MGD) in order to utilize the phase of CQT. Also, we present a simple 2D-convolution multi-branch network architecture for replay detection, which can model the distortion both in the time and frequency domains. The experiment shows that the proposed CQT-based MGD feature outperforms traditional MGD feature, and performance can be further improved by combining both magnitude-based and phase-based feature. Our best fusion system achieves 0.0096 min-tDCF and 0.39% EER on ASVspoof 2019 Physical Access evaluation set. Comparing with the CQCC-GMM baseline system provided by the organizer, the min-tDCF is relatively reduced by 96.09% and EER is relatively reduced by 96.46%. Our system is submitted to the ASVspoof 2019 Physical Access sub-challenge and won 1st place.