You can already achieve that by combining models - use a dedicated speech synthesis model for the narration, then layer that over background effects from AudioGen.
Given that, I don't think AudioGen particularly needs to add full narration. That seems like a very different problem to me, likely requiring a completely different architecture.
