To date, there are many computer programs that are able to read with varying success the lips of a person what he is currently talking about. As a rule, the accuracy of recognition of such programs does not exceed 52%, which is quite good, but still far from perfect. Even a very professional surdopedist can attain accuracy only in 50-60%. And scientists from Oxford University managed to develop a LipNet algorithm that reads on the lips with an accuracy of 93.4%, which is by far the unsurpassed result.
The secret of such a high efficiency of the new program is that it does not try to analyze the words separately from each other, but accepts the sentences entirely, and then uses the technique of deep machine learning and proceeds to decipher the verbal sequence. In fact, we have an advanced neural network capable of guessing almost every word spoken by a person with high probability, without having access to auditory information. The scope of this algorithm is simply huge, but first of all it can become a lifeline for many people with a weakened or completely missing hearing.
Scientists had to shovel a lot of existing research to date in the field, explore dozens of other algorithms to identify their weaknesses and understand which direction to move further in order to improve the technology. The results of their work are impressive. So far, of course, LipNet is able to recognize the lips of an exclusively English-speaking speech. But in the future no one interferes with teaching the system a new language for it. You can see how the algorithm works in a video published on the official channel of one of the researchers.
The article is based on materials .
Comments
Post a Comment