How would you be able to pull out any information out of an audio recording of keyboard typing? Wouldn't the training set differ between keyboards too? Sounds interesting it it worked, do you have any sources?
You get a week's worth of typing; you cluster the sounds; then based on statistical frequency analysis you decode which cluster maps to the spacebar and which to the letter 'x'; then you can transcribe the whole typing history and decide which of all that is a password.