New AI Algorithm Can Alter Videos of Anyone to Say Anything

Artificial intelligence researchers just developed a new video editing tool that can create realistic videos from any audio clip used as input.

Tests of the technology were used on former President Obama, taking audio from one of his speeches and using the algorithm to develop videos of the former president in various settings.

While you might expect the tool to be used to create fake news and invalidate video media, it’s designed to do just the opposite. Researchers from the University of Washington designed the AI tool to better spot video fakes and autonomously analyze the authenticity of video media. Since the tool knows nearly perfectly how to fake a video, it makes it very good at detecting other forgeries.

You can see a demo of the tool below.

The team at the University of Washington say that as long as the audio source for a video is of quality, the algorithm can create realistic nearly 100% accurate mouth movements of a subject in video form.

In more standard applications of this technology, it could also be used to enhance video conferencing. Inputting this algorithm onto a photo or even as a supplemental tool to prevent lagging, it could create consistent visual output that would line up with the audio feed.

A quick guide from the research team on how the algorithm works can be seen below.



The system works in two steps. First, the AI analyzes large amounts of video to format a specific algorithm for how mouth shapes change along with audio cues. Then, the system is fed a small amount of video and images from a specific person in order to alter its base algorithm as necessary.

If you watched the video above, you can see just how good the AI is at altering mouth movements. The only drawback to the technology is that it requires authentic, or seemingly authentic source audio. It would take a highly skilled impressionist or further AI advances to “put words into someone’s mouth” in a fake video. Even still, skilled editing of audio clips could yield authentic-sounding base audio files.

Ultimately though, the main goal of this project is to enhance fraud detection in video files. So it seems that this AI will benefit us for now – until it gets into the wrong hands.

Source: University of Washington


Trevor is a civil engineer (B.S.) by trade and an accomplished writer with a passion for inspiring everyone with new and exciting technologies. He is also a published children’s book author and the producer for the YouTube channel Concerning Reality.

It's only fair to share...Share on Facebook0Share on Google+0Tweet about this on TwitterShare on LinkedIn0