Recently, Google DeepMind artificial intelligence is learning with all new and new functions with amazing speed. He has already begun to have fun playing video games, and in general, for self-study, he practically does not need a person. However, there is no limit to perfection, and not so long ago, the AI acquired another new skill: lip reading.
True, in this case it was not possible to manage without human participation. Scientists had to install an advanced mimic recognition algorithm into the neural network, after which DeepMind had to “watch” 5,000 hours of the BBC show, such as BBC Breakfast, Newsnight and Question Time, which contained about 120,000 offers. As a result, the AI has learned to read lips, even in cases where a person swallows parts of words. After this, the most interesting phase of the experiment began. Speech recognition and artificial intelligence were given the opportunity to view 200 unrelated fragments. As a result, the professional correctly recognized only 12.5% of words, while the AI was not mistaken in 46.8% of cases.
This effect was achieved due to the fact that in the process of learning, the facial recognition algorithm on which the machine was based was based on a fairly advanced lip reading system called LipNet, developed at Oxford University. At the moment, experts from Google together with the developers of LipNet are preparing an Internet site with free access for developers involved in the creation of automatic lip reading systems so that even more advanced recognition mechanisms are created. According to one of the authors of the experiment, Dr. Zheng Zhou,
“All this is a huge step towards the creation of fully automatic lip reading systems. This became possible only due to the huge set of basic data on which this system was trained. ”
The article is based on materials .
Comments
Post a Comment