afaik, it converts the raw audio to text STT, then feeds the text to LLM and then converts that text into raw audio; combining both texts should give a plausibe transcript