Future Timeline

Deep fake technology has exploded in popularity since 2017. In a strange parody of Moore’s Law, the number of deep fake videos published online doubles every six months, according to one estimate.

Daniel Higgins knew something was up when he suddenly couldn’t log into his Instagram account and then discovered his password had been changed.

But he couldn’t believe his eyes when he saw a new post on his account: a video of man that looked and sounded just like him, telling all of his Instagram followers to buy Bitcoin.

“I just invested $300 into Bitcoin and got $10,000 back. Gotta try it,” the fake Higgins said in the video.

Suddenly, he received text messages from followers wondering what was going on.

That was weeks ago. Higgins has been trying ever since to regain access to his account. He says his attempts to convince Instagram to help have gone no where.

There’s no real person to contact, so Higgins has gone through the online procedure of alerting Instagram of the problem. He was even told by the system he would be sent a new password.

This work explores the task of synthesizing speech in non-existent human-sounding voices. We call this task "speaker generation", and present TacoSpawn, a system that performs competitively at this task. TacoSpawn is a recurrent attention-based text-to-speech model that learns a distribution over a speaker embedding space, which enables sampling of novel and diverse speakers. Our method is easy to implement, and does not require transfer learning from speaker ID systems. We present objective and subjective metrics for evaluating performance on this task, and demonstrate that our proposed objective metrics correlate with human perception of speaker similarity.

This page contains a set of audio samples in support of the paper: we suggest that readers listen to the samples in conjunction with reading the paper.

Note that the TacoSpawn systems described in the paper extend a baseline Tacotron -- that is, they do not use any additional acoustic or prosody embeddings to improve the naturalness of synthesized speech.