June 2004

Read My Lips

By Technology Review

Computers that observe facial expressions could personalize online communications, make video games more interactive, and read lips to help recognize speech. But facial expressions change so quickly and unpredictably that computers have trouble keeping up. At Tsinghua University in Beijing, China, computer scientist Guangyou Xu has developed software that takes video of a person's face and traces an accurate outline of the lips from frame to frame. Xu first trained a computer to draw the 2-D outline of a goldfish swimming back and forth in a bowl; from this, he developed an algorithm that represents the complex motion of lips using just a few mathematical parameters. That enables the computer to learn the lips' typical range of movement, so it can distinguish them from other features, like teeth, more accurately than current techniques. Xu is in discussions with companies to commercialize the software, which could be used to make video games and online chat rooms more vivid and, eventually, to analyze surveillance video.