Somehow need to overlap an LLM with vocal stream processing to identify semantically meaningful transition points to interrupt naturally instead of just looking for any pause or sentence boundary.