Researchers at India's International Institute of Information Technology (IIIT) have developed artificial intelligence (AI) that tweaks video footage in real time to sync audio to people's lips.
The team trained a generative adversarial network algorithm on short clips, and it marked out people's lip shapes as they spoke.
For a given sound and video footage of someone talking, a generator AI adjusted lip imagery to match spoken words, while two discriminator AIs differentiated real from fake footage.
One discriminator focused on the realism of mouth shapes, and penalized the generator for mismatched sound and lip movements; the second discriminator checked visual quality, flagging glitches or unnatural visual artifacts around the mouth.
Because it was trained on human faces, the algorithm is more accurate on video footage of people than on computer-generated characters; the IIIT team hopes to use the AI for multilingual video dubbing.
From New Scientist
View Full Article
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA