A better mousetrap?

So Markov chains provide a good model for next-symbol probability, and touchscreens are amazingly shonky at working out what you really just typed.  Is anyone using a Markov chain to interpret the raw data from the users fat, mashy finger and determine what they really meant to press?  A second-order chain ought to be enough to pick out half a dozen likely candidate letters, most of which won't be in the area of the press.