Fake speech: Generating and detecting convincing voice impersonations by humans and machines

该博士项目旨在针对音频领域中deepfake技术的犯罪使用，即假语音（即模仿语音）。虽然近年来Deepfake应用程序越来越为公众所熟知，但对Deepfake的普遍看法仍然与视觉模仿有关，即将一个人（通常是名人）的脸数字移植到另一个人的脸上，并且通常取得相当大的成功，因为Deepfake图像和视频变得越来越令人信服。然而，尽管虚假言论的领域不太为公众所知，但它仍然给社会带来了真实的问题和后果。因此，这个项目将通过四种方式的排列来解决语音问题：i）人类检测由人类生成的假语音ii）人类检测由机器生成的假语音iii）机器检测由人类生成的假语音iv）机器检测由机器生成的假语音通过提高四种场景中的每一种中的假语音检测率，预计该项目的成果可以成为打击假讲话的整体折衷努力的一部分。

This doctoral project aims to target the criminal use of deepfake technology in the audio domain, namely fake speech(i.e. impersonated speech). While deepfake applications are becoming more well-known to the public in recent years, thegeneral perception of deepfake still relates to visual impersonation, i.e. digitally transplanting an individual's (typically afamous person) face on top of another person's face, and often with considerable success as deepfaked images andvideos are becoming more convincing. However, although the domain of fake speech is less well-known to the generalpublic, it nevertheless presents real problems and consequences to society. Hence, this project will tackle the problem offake speech through a four-way permutation: i) Humans detecting fake speech generated by humans ii) Humansdetecting fake speech generated by machines iii) Machines detecting fake speech generated by humans iv) Machinesdetecting fake speech generated be machines By improving the rate of fake speech detection in each of the fourscenarios, it is anticipated that the results arising from this project can form a part of an overall eclectic effort to combatfake speech.