Vulnerability of voice verification systems to spoofing attacks by tts voices based on automatically labeled telephone speech

Abstract

This paper explores the robustness of a text-dependent voice verification system against spoofing attacks that use synthesized speech based on automatically labeled telephone speech. Our experiments show that when manual labeling is not used in creating the synthesized voice, and the voice is based on telephone speech rather than studio recordings, False Acceptance error rate decreases significantly compared to high-quality synthesized speech.