thinking outside the box, i'm a dev, and a feedback loop might be helpful here. If i were to build this, i would save the audio generated --- transcribe it, and pipe it back into a llm and ask for refinements.. maybe won't do much but it might help.