Think of creating some variation of a
Markov program to generate words based on a collection of proper speech/words, say a dictionary or large literary work. It will improve your chances of producing workable results.
BTW, that was a quick google search result I provided, a search on 'Markov Chanining' should produce some decent references.