Traditional cascaded systems consist of three components—a speech recognizer, a large language model, and a separate turn detector—which coordinate poorly with each other and require manual tuning