this idea has been around for a long, long time (Score 2, Interesting)

Claude Shannon, the father of information theory, used the idea referenced here in his famous 1950 experiment to calculate the entropy of the English language. See "Shannon Game" at, for example, http://www.math.ucsd.edu/~crypto/java/ENTROPY/ There's also an entire field, often referred to as "Natural Language Processing," which uses empirical observations of large amounts of language data (text or speech) to construct statistical models which do speech recognition, language translation, text summarization, spelling correction (and, yes, people at Microsoft Research have worked on this), etc. Finally, Hemos writes "Stochastic modeling can also be used as a basis for speech recognition, with the recognizer using the model to choose a continuation when the speech signal is ambiguous or indistinct." FYI, speech signal is _always_ ambiguous, from the perspective of a machine trying to transcribe it to text. I very much doubt there's been any successful speech recognition work in the last 15 years on a non-statistical system.