You could use your parser as a language model to improve recognition along the lines of things that would actually parse correctly, I suppose. That's probably much of a micro-optimization, though.