One of the highlights of the recent Siggraph conference was a demonstration of a provocative new technique for facial animation called Trainable Videorealistic Speech Animation. Developed by Tony Ezzat and a team at MIT's Center for Biological and Computational Learning, the technology allows the user to alter a simple video camera recording of a person saying something to make it appear that he or she is saying something completely different. With the new method, a synthesized speech sequence generated on a computer can be composited with a video sequence so that it is virtually impossible to detect the alteration. In the images below, the top row shows stills from a sequence of unaltered video images of a person speaking. The middle row is an animation of a synthetic mouth generated in the lab. The bottom row shows the synthetic mouth animation superimposed on the original sequence. —Phil LoPiccolo
 |
| Images courtesy Tony F. Ezzat, MIT. |