Have you ever tried to read someone’s lips before? If you’ve played games like the Whisper Challenge (see video above if you haven’t), then you know how difficult it is simply because there are certain words that we say that shares the same mouth and lips shape as another, thus adding to the confusion and difficulty.
Now there are some people who are proficient at lip-reading, but unfortunately it seems that computers have managed to beat us once again. Over at Oxford University, it seems that a couple of the university’s departments have come up with different but effective ways that have taught computers how to read the lips of people.
The Department of Computer Science developed an AI system called LipNet and built it using a data set called GRID, which is based on well-lit videos of people facing the camera and reading three seconds worth of words. From there it trains a neural network which according to the tests, is capable of identifying 93.4% of the words correctly. Human lip-reading volunteers only managed to get it right 52.3% of the time for the same tasks.
Oxford’s other department, the Department of Engineering Science, developed another system while working with Google’s DeepMind. This one is based on a series of 100,000 videos from the BBC and according to their findings, it was successful 46.8% of the time, versus 12.4% which was based on a professional lip-reader’s attempts.
While these advancements are no doubt great for technology, we have to wonder what the implications and implementations are.
Filed in . Read more about AI (Artificial Intelligence).