Screaming Penguin - Source Does Matter

The crew at Keyboard-emanations.org has a very detailed research paper which outlines the process by which they are able to listen to the sounds "emanating" from a keyboard as you type and, assuming English is in use, determine what is being typed only from the sound with a 90% accuracy.

I wont pretend to have complete comprehension of the very detailed process but a very rough overview is that the researchers have mathematically identified groups of keys that are commonly typed with known pressure and angles and thus different sounds (and delays in between - lack of audio may also be used?) and using those groups with dictionary technology they can determine with a high degree of accuracy the probability of what key is being typed. Once they have listened in for a while they can then generate a sort of key or style for the particular typist in a language with a known dictionary (only English thus far) and then can use that key to even identify NON dictionary words, such as passwords, etc.

This is really remarkable research on "emanations". This tactic would of course require some degree of local physical access to get the audio data (and with said physical access there are of course many other ways to get keystroke data) but it is still very impressive (though not potentially useful or dangerous to most people in its present form, but with a powerful focused parabolic microphone, hmmmm).