By Gareth Cook, Globe Staff, 5/15/2002
AMBRIDGE - Scientists at the Massachusetts Institute of Technology have created the first realistic videos of people saying
things they never said - a scientific leap that raises unsettling questions about falsifying the moving image.
In one demonstration, the researchers taped a woman speaking into a camera, and then reprocessed the footage into a new video that showed her
speaking entirely new sentences, and even mouthing words to a song in Japanese, a language she does not speak. The results were enough to fool
viewers consistently, the researchers report.
The technique's inventors say it could be used in video games and movie special effects, perhaps reanimating Marilyn Monroe or other dead film
stars with new lines. It could also improve dubbed movies, a lucrative global industry.
But scientists warn the technology will also provide a powerful new tool for fraud and propaganda - and will eventually cast doubt on everything
from video surveillance to presidential addresses.
''This is really groundbreaking work,'' said Demetri
Terzopoulos, a leading specialist in facial animation who is a professor of computer science and mathematics at New York University. But ''we are
on a collision course with ethics. If you can make people say things they didn't say, then potentially all #### breaks loose.''
The researchers have already begun testing the technology on video of Ted Koppel, anchor of ABC's ''Nightline,'' with the aim of dubbing a show
in Spanish, according to Tony F. Ezzat, the graduate student who heads the MIT team. Yet as this and similar technology makes its way out of
academic laboratories, even the scientists involved see ways it could be misused: to discredit political dissidents on television, to embarrass
people with fabricated video posted on the Web, or to illegally use trusted figures to endorse products.
''There is a certain point at which you raise the level of distrust to where it is hard to communicate through the medium,'' said Kathleen Hall
Jamieson, dean of the Annenberg School for Communication at the University of Pennsylvania. ''There are people who still believe the moon landing
Currently, the MIT method is limited: It works only on video of a person facing a camera and not moving much, like a newscaster. The technique
only generates new video, not new audio.
But it should not be difficult to extend the discovery to work on a moving head at any angle, according to Tomaso Poggio, a neuroscientist at the
McGovern Institute for Brain Research, who is on the MIT team and runs the lab where the work is being done. And while state-of-the-art audio
simulations are not as convincing as the MIT software, that barrier is likely to fall soon, researchers say.
''It is only a matter of time before somebody can get enough good video of your face to have it do what they like,'' said Matthew Brand, a
research scientist at MERL, a Cambridge-based laboratory for Mitsubishi Electric.
For years, animators have used computer technology to put words in people's mouths, as they do with the talking baby in CBS's ''Baby Bob'' -
creating effects believable enough for entertainment, but still noticeably computer- generated. The MIT technology is the first that is
''video-realistic,'' the researchers say, meaning volunteers in a laboratory test could not distinguish between real and synthesized clips. And
while current computer- animation techniques require an artist to smooth out trouble spots by hand, the MIT method is almost entirely automated.
Previous work has focused on creating a virtual model of a person's mouth, then using a computer to render digital images of it as it moves. But
the new software relies on an ingenious application of artificial intelligence to teach a machine what a person looks like when talking.
Starting with between two and four minutes of video - the minimum needed for the effect to work - the computer captures images which represent
the full range of motion of the mouth and surrounding areas, Ezzat said.
The computer is able to express any face as a combination of these faces (46 in one example), the same way that any color can be represented by a
combination of red, green, and blue. The computer then goes through the video, learning how a person expresses every sound, and how it moves from
one to the next.
Given a new sound, the computer can then generate an accurate picture of the mouth area and virtually superimpose it on the person's face,
according to a paper describing the work. The researchers are scheduled to present the paper in July at Siggraph, the world's top computer
The effect is significantly more convincing than a previous effort, called Video Rewrite, which recorded a huge number of small snippets of video
and then recombined them. Still, the new method only seems lifelike for a sentence or two at a time, because over longer stretches, the speaker
seems to lack emotion.
MIT's Ezzat said that he would like to develop a more complex model that would teach the computer to simulate basic emotions.
A specialist can still detect the video forgeries, but as the technology improves, scientists predict that video authentication will become a
growing field - in the courts and elsewhere - just like the authentication of photographs. As video, too, becomes malleable, a society
increasingly reliant on live satellite feeds and fiber optics will have to find even more direct ways to communicate.
''We will probably have to revert to a method common in the Middle Ages, which is eyewitness testimony,'' said the University of Pennsylvania's
Jamieson. ''And there is probably something healthy in that.''
Compare original and synthetic videos from MIT on www.boston.com/globe.