Google’s AI capable of better lip reading than expert humans

Researchers have developed lip reading software capable of deciphering spoken words more accurately than expert humans. Scientists from the University of Oxford and Google’s DeepMind artificial intelligence (AI) division trained a program to lipread by watching 5,000 hours of BBC programme footage, including Question Time, Newsnight and BBC News. The teams used machine learning, a form of AI capable of analysing and learning from data free from human intervention, to match sounds of words to the shapes made by human mouths when speaking them. The program, called Watch, Listen, Attend and Spell (WLAS), then transcribed the mouth motion of the speakers into characters, from 118,000 unique sentences and 17,500 individual words. The programmes were broadcast between January 2010 and December 2015, then a second set of footage shown between March…


Link to Full Article: Google’s AI capable of better lip reading than expert humans

Pin It on Pinterest

Share This