I came across an interesting paper today, "Keyboard Acoustic Emanations Revisited" [PDF] Li Zhuang, Feng Zhou, and J. D. Tygar, researchers at UC Berkeley, have discovered a way to tell what someone is typing using only the recorded audio of the typing session.  It works based on the fact that every key on your keyboard makes a slightly different but unique sound.  Armed with this knowledge and a standard cryptographic frequency analysis attack and a little bit of machine learning they are able to turn the sounds of the keyboard into text, with absolutely no training of the program beforehand and no previous samples.  Whats more, this attack has a %96 success rate and even works against random text such as a password.
A cryptographic frequency analysis attack works like this, we know that the most commonly used letters in the english language are E,T,A and S.  If we have a code where a symbol consistently represents one letter (or in this case a sound consistently represents one letter) then we can assume the most common symbol (sound) represents E, the next most common is T, and so forth.  This is the basic theory that the keyboard emanations works on.
Apparently this paper came out back in 2005, but it went largely unnoticed by the media, including blogs.  Definitely the most interesting thing about this attack is its ease of implementation and success rate.  Keyboard acoustic sniffing attacks had been written about before, but they required sample data for training and had a much lower success rate.  This attack requires no training data and even works with a microphone outside the room if using a parabolic microphone.  Some might say it is similar to Van Eck Phreaking.
Interestingly while researching this I came across a patent for the "method and apparatus for masking acoustic keyboard emanations." So you may already be able to stop this attack, and you may be committing patent infringement if you do it DIY.