Thats a cool idea. Could the LLM find the right location for the audio stream by simply having the context of the buffer, and the location of the text and audio cursor when the intersction starts?