I wonder if it makes any assumptions about layout and keyboard shape (for example, it might not know that I was using Dvorak, or that I had an ergo keyboard where my fingers were in atypical places).
I'm sure it uses some kind of Markov chain statistical analysis technique that would have to be programmed to assume you were typing on a particular keyboard layout in a particular language. On the other hand, there's nothing stopping them from trying to decode the raw data with several different configurations and seeing which one sticks. The IBM Selectric bugs the Soviets planted [1] did a similar thing, where they only transmitted four bits per character, but the messages could be rehydrated knowing letter and word frequencies from the English language.
