The technique uses the edges of objects in the video. The idea is they vibrate from the sound waves. The vibrations are too small to have the edges displaced in the pixel resolution of the video, but the edge pixels change in color tone as they vibrate. It is that variation that is tracked to recreate the sounds.
It is a fascinating technique that is actually quite creepy considering the slowly evolving surveillance state some desire. We don't need Big Brother surreptitiously listening as well as watching.
