Our machines have gotten better and better at understanding us and the world around them. Whether it's computer vision or voice recognition, machine-learning has vastly improved these fields. And now, it’s opening another door in our quest to communicate naturally, by learning to read our lips.
Researchers at the University of Oxford Computer Science Department have developed LipNet, a service which can read lips and understand what people are saying. Though far from being the first of its kind, LipNet has become the most accurate such system to date, achieving 93.4 percent accuracy. That becomes even more impressive when considering that experienced human lip readers only achieve around 52% accuracy.
The system managed to achieve these results by analyzing entire sentences, not just individual words that users are saying. By doing so, its machine-learning algorithm could then reconstruct the words and meaning of the users, based on language models. Interestingly enough, a similar move in the field of voice recognition was also the reason that that area of research saw impressive performance growth recently. It seems context really is everything when listening to humans.
Given that our machines can now recognize the world better than humans, understand speech better than humans, and read our lips better than humans, and almost speak just like we do, we may as well welcome our new robotic overlords.