Extracting audio from visual information

Researchers at MIT, Microsoft, and Adobe have developed an algorithm that can reconstruct an audio signal by analyzing minute vibrations of objects depicted in video. In one set of experiments, they were able to recover intelligible speech from the vibrations of a potato-chip bag photographed from 15 feet away through soundproof glass.

In other experiments, they extracted useful audio signals from videos of aluminum foil, the surface of a glass of water, and even the leaves of a potted plant. The researchers will present their findings in a paper at this year’s Siggraph, the premier computer graphics conference.

“When sound hits an object, it causes the object to vibrate,” says Abe Davis, a graduate student in electrical engineering and computer science at MIT and first author on the new paper. “The motion of this vibration creates a very subtle visual signal that’s usually invisible to the naked eye. People didn’t realize that this information was there.”

Joining Davis on the Siggraph paper are Frédo Durand and Bill Freeman, both MIT professors of computer science and engineering; Neal Wadhwa, a graduate student in Freeman’s group; Michael Rubinstein of Microsoft Research, who did his PhD with Freeman; and Gautham Mysore of Adobe Research.

Reconstructing audio from video requires that the frequency of the video samples — the number of frames of video captured per second — be higher than the frequency of the audio signal. In some of their experiments, the researchers used a high-speed camera that captured 2,000 to 6,000 frames per second. That’s much faster than the 60 frames per second possible with some smartphones, but well below the frame rates of the best commercial high-speed cameras, which can top 100,000 frames per second.

Commodity hardware

In other experiments, however, they used an ordinary digital camera. Because of a quirk in the design of most cameras’ sensors, the researchers were able to infer information about high-frequency vibrations even from video recorded at a standard 60 frames per second. While this audio reconstruction wasn’t as faithful as that with the
high-speed camera, it may still be good enough to identify the gender of a speaker in a room; the number of speakers; and even, given accurate enough information about the acoustic properties of speakers’ voices, their identities.

Read more . . .