I'm disappointed that this summary doesn't touch on the actual results; the system recognizes 19/20 characters in the best of cases, and 9/10 is an optimistic number otherwise (so many different trials make naming one number impossible). I am glad to discover this awesome, extremely impressive work; but including numeric results would have ensured that I saw this as a information security problem. Instead, I imagined that I could unplug my keyboard forevermore :)
Hi, I wrote this post - you're right that I should have emphasised the application in security/espionage. With regard to the results, it's very difficult to judge, because in natural language text, even with a low accuracy rate per character people would still be able to deduce the message (see http://www.mrc-cbu.cam.ac.uk/people/matt.davis/cmabridge/) With regard to password cracking, even a low accuracy rate could make a previously impractical brute force search possible,