Use of Claimed Speaker Models for Replay Detection

Abstract

Replay attacks are the simplest form of spoofing attacks on automatic speaker verification (ASV) systems and consequently the detection of these attacks is a critical research problem. Currently, most research on replay detection focuses on developing a stand-alone countermeasure that runs independently of a speaker verification system by training a single common spoofed model as well as a single common genuine model. This paper investigates the potential advantages of sharing speaker data between the speaker verification system and the replay detection system. Specifically, it explores the benefits of using the claimed speaker's model in place of the common genuine model. The proposed approach is validated on a modified evaluation set of the ASVspoof 2017 version 2.0 corpus and show that the use of adapted speaker models is far superior to the use of a single common genuine model.