This Speech Synthesis Model Our Engineers Built Recreates A Human Voice Perfectly

On May 15, we announced that three of our Machine Learning Engineers, Hashiam Kadhim, Joseph Palermo and Rayhane Mama, have created what we think is the most realistic sounding AI voice we’ve heard to date. And it’s the voice of someone you’ve probably heard of before: Joe Rogan.

Using a text-to-speech synthesis model they built called RealTalk, the trio have crafted a series of audio clips from a ‘fake Joe Rogan’ that are practically indistinguishable from the prolific podcaster’s voice.

Check it out for yourself:

To prove how close the clips really sound to the real Rogan’s voice, the team also created a Turing-like Test featuring clips of both the real Joe Rogan and fake Joe Rogan speaking on a website they’ve coined

Obviously, there are some serious societal implications that come with the advancement of this technology. Because of this, at this time we have decided to refrain from making any parts of the model or dataset public. We’ve written more on both the implications and also on this decision on our Medium. Check out the post here.

Since we’ve published RealTalk, the work has also become viral on Reddit and been covered in The Verge, VICE, and other media outlets around the world.

This project was developed as part of Dessa’s Meta Labs initiative, an internal program we’ve developed that encourages employees to work on independent projects that advance their knowledge of machine learning — and then some. Ambitious in scope, these projects are a labour of love by our engineers (read countless hours outside of work), and often end up stretching the boundaries of what’s possible for the technology. Major props!

If you enjoyed reading about RealTalk, make sure to check out space2vec, another Meta Labs project here.

Please note that this project does not suggest that we endorse the views and opinions of Joe Rogan. Joe was selected as a demonstrative model for the purposes of displaying the capability of this technology.



This Speech Synthesis Model Our Engineers Built Recreates A Human Voice Perfectly

May 22, 2019