Partial Fake Speech Attacks in the Real World Using Deepfake Audio

Abstract

Advances in deep learning have led to dramatic improvements in generative synthetic speech, eliminating robotic speech patterns to create speech that is indistinguishable from a human voice. Although these advances are extremely useful in various applications, they also facilitate powerful attacks against both humans and machines. Recently, a new type of speech attack called partial fake (PF) speech has emerged. This paper studies how well humans and machines, including speaker recognition systems and existing fake-speech detection tools, can distinguish between human voice and computer-generated speech. Our study shows that both humans and machines can be easily deceived by PF speech, and the current defences against PF speech are insufficient. These findings emphasise the urgency of increasing awareness for humans and creating new automated defences against PF speech for machines.