Deep learning for speech synthesis: The good news, the bad news, and the fake news

Description

Mention speech synthesis and most people will think of robotic voices characteristic of historic systems such as that used by the late Stephen Hawking. However, modern developments in deep learning allow us to build end-to-end speech synthesis systems that not only require no linguistic domain expertise but can generate speech indistinguishable to the human ear from a real speaker.

While these technologies have myriad benevolent applications, they also usher in a new era of fake news. What if a malicious actor could generate an audio recording of a political adversary saying anything they want and leak this to the press for political gain?

Scott Stevenson discusses the technical developments in deep learning that make this possible and the potential impacts on public discourse of modern political disinformation. He then explores how the public can be inoculated against such disinformation campaigns and how to use additional modern machine learning techniques such as adversarial networks to build countermeasures against such disinformation campaigns.