All your voices are belong to us: Stealing voices to fool humans and machines

Abstract

In this paper, we study voice impersonation attacks to defeat humans and machines. Equipped with the current advancement in automated speech synthesis, our attacker can build a very close model of a victim’s voice after learning only a very limited number of samples in the victim’s voice (e.g., mined through the Internet, or recorded via physical proximity). Specifically, the attacker uses voice morphing techniques to transform its voice – speaking any arbitrary message – into the victim’s voice. We examine the aftermaths of such a voice impersonation capability against two important applications and contexts: (1) impersonating the victim in a voice-based user authentication system, and (2) mimicking the victim in arbitrary speech contexts (e.g., posting fake samples on the Internet or leaving fake voice messages). We develop our voice impersonation attacks using an off-the-shelf voice morphing tool, and evaluate their feasibility against state-of-the-art automated speaker verification algorithms (application 1) as well as human verification (application 2). Our results show that the automated systems are largely ineffective to our attacks. The average rates for rejecting fake voices were under 10–20% for most victims. Even human verification is vulnerable to our attacks. Based on two online studies with about 100 users, we found that only about an average 50 % of the times people rejected the morphed voice samples of two celebrities as well as briefly familiar users.