We can assume that a Hidden Markov Model exists behind every possible grammatically correct sentence. If we can develop a rudimentary HMM for a language like English, we can use it to predict the likelihood of a sentence being said. We can take an imperfect speech recognition software and use the sounds the program hears to generate a list of possible sentences that sound similar. We use the HMM to determine how likely each of these sentences is to be said, relative to each other. This gives us a good idea of what the human actually said, rather than picking one possibility at random, since real sentences are far more likely than nonsense that simply rhymes with the sentence.
~Kalin Jonas