LipNet : when the AI is better than the man to read on the lips – ZDNet France

The learning curve of artificial intelligence becomes exponential. It was already mentioned here wins multiple AlphaGo the game of Go against a human opponent. Or even the steady progress of virtual assistants, thanks to Machine Learning algorithms. Today, scientists from the University of Oxford and Google Deepmind once again demonstrate the superiority of the machine in a very specific field : the lip-reading.

initially, it was to develop an artificial intelligence capable of automatically translating a text or simulate a voice almost human. Now, this AI is able to read lips (no sound source), reveals New Scientist magazine. The program called LipNet can convert the lip movements into text and is a lot better than an expert human of the question. The application gets a score of 47% recognition accurate compared to 24% for the human.

LipNet operates, of course, the Machine Learning, the system has digested not less than 5000 hours of video of six TV programmes that are different from the BBC, in order to learn to walk, the better it anticipates that it decrypts. “Based on what we know, LipNet is the first model of lip-reading, which works at the sentence level”, the researchers say.

With Deepmind, Google is at the forefront of the AI and he intends to exploit unexpected applications to improve its technology. On the occasion of the conference BlizzCon from the publisher Blizzard, the subsidiary AI of Google, has announced the creation of a test environment open in StarCraft II. This partnership with the game editor will be able to use this as iconic for gamers as a test platform for artificial intelligence and the research on machine learning.

This platform will be open to all, including DeepMind, starting in 2017. “We have worked in close collaboration with the team of StarCraft II to develop an API that supports something similar to bots in previous writings, with an interface that “scripted”, enabling control to programmatic individual units and access to the complete state of the game, ” said DeepMind.

The firm believes that video games are one of the best ways to “develop and test AI algorithms more intelligent and flexible” and this more quickly and efficiently.