What we found when creating audiobooks: You can add a decoding seed parameter and stick to that (like when Minecraft generated worlds) you can try online system prompts, you can chunk aggressively .. and at the end of the day it’s still gonna sound different but it takes the edge off a bit.