The issue with this is you are always going to get transcriber input from your user and proper LLM generated text for your agent to train on that defeats the purpose of making it more conversational. It’ll help a tad maybe