Oxford scientists have created a system capable of lip reading with a tiny margin of error.
There is no worthwhile person who does not know how to read lips; It is a difficult and error-prone technique, but teachers are able to learn everything from a distance.
The difficulty of lip reading
Of course, these classes in the espionage course may not be needed any time soon. All thanks to LipNet, a new Artificial Intelligence system that reads lips with high precision.
There are already automated systems capable of lip reading, but they work by analyzing each word separately; They also often associate phonemes with certain lip movements.
But to read the lips, it is necessary to capture in detail all the complex movements of the mouth, and even then we can fail.
For example, the same gesture on the lips can correspond to different phonemes; the context and the person speaking are details that must be taken into account.
The near-perfect lip-reading system still has to improve
So LipNet focuses on analyzing complete sentences, instead of the separate words.
The Oxford researchers made the system swallows 29,000 videos of people talking to a camera. The videos were only three seconds long and included subtitles.
The results are evident, with 93.4% accuracy; In contrast, a normal system achieves an accuracy of 79.6%. In both cases far outnumber humans, which in the tests achieved an accuracy of 52.3%.
93.4% is very good, but not perfect. Furthermore, the great weak point is that this system you need your lips to be looking at the camera and have the ideal lighting.
So the next challenge for researchers will be teach AI with videos recorded in real life situations; The next tests will consist of videos at various angles.
If it works, LipNet could be of great help to deaf people, for example.