Imagine someone listening in to your private conversation by filming the bag of chips sitting on the other side of the room. Oddly specific, I know, but researchers at MIT did just that: They've created an algorithm that can reconstruct sound (and even intelligible speech) with the tiny vibrations it causes on video.
When sound hits an object, it makes distinct vibrations. “There’s this very subtle signal that’s telling you what the sound passing through is,” said Abe Davis, a graduate student in electrical engineering and computer science at MIT and first author on the paper. But the movement is tiny – sometimes as small as thousandths of a pixel on video. It’s only when all of these signals are averaged, Davis said, that you can extract sound that makes sense. By observing the entire object, you can filter out the noise.
This particular study grew out of an earlier experiment at MIT, led by Michael Rubinstein, now a postdoctoral researcher at Microsoft Research New England. In 2012, Rubinstein amplified tiny variations in video to detect things like the skin color change caused by the pumping of blood. Studying the vibrations caused by sound was a logical next step. But getting intelligible speech out of the analysis was surprising, Davis said.
The results are certainly impressive (and a little scary). In one example shown in a compilation video, a bag of chips is filmed from 15 feet away, through sound-proof glass. The reconstructed audio of someone reciting “Mary Had a Little Lamb” in the same room as the chips isn’t crystal clear. But the words being said are possible to decipher.
In most cases, a high-speed camera is necessary to accomplish the feat. Still, at 2,000 to 6,000 frames per second, the camera used by the researchers is nothing compared to the best available on the market, which can surpass 100,000 frames per second. And the researchers found that even cheaper cameras could be used.
“It’s surprisingly possible to take advantage of a bug called rolling shutter,” Davis said. “Usually, it creates these artifacts in the image that people don’t like.” When cameras use rolling shutter to capture an image, they don’t capture one single point in time. Instead, the camera scans across the frame in one direction, picking up each row at a slightly different moment.
By doing so, the camera happens to encode information at a much higher rate than its actual frame rate. For the researchers, that meant being able to analyze vibrations that should have happened too quickly for capture on film. “It kind of turns a two-dimensional low-speed camera into a one-dimensional high-speed camera,” Davis explained. “As a result, we can recover sounds happening at frequencies several times higher than the frame rate of the camera, which is remarkable when you consider that it’s just a complete accident of the way we make them.”
There are definitely limitations to the technology, Davis said, and it may not make for better sound reconstruction than other methods already in use. “Big brother won't be able to hear anything that anyone ever says all of a sudden,” Davis said. “But it is possible that you could use this to discover sound in situations where you couldn’t before. It’s just adding one more tool for those forensic applications.”
Davis and his colleagues care more about applications in scientific research. “This is a new dimension to how you can image objects,” he said. “It tells you something about how they respond physically to pressure, but instead of poking and prodding at them, all you need is to play sound at them.”
More from The Washington Post: