Not sure if what you’re describing couldn’t be also done with audio snippets and good splicing
A good example is Mortal Kombat where each character has several lines for every OTHER character (https://www.youtube.com/watch?v=L85QApISlvA) they are about to fight, that's a LOT of voice work as opposed to using the character lore and their history and or relationship with the opponent to come up with something fresh that's (witty|snarky|sad) to say etc.
Once we get flexgen running Llama (or some other combination of optimizations) there will be no audio files and these LLM’s can be run client side on consumer hardware
probably as a shared resource at the OS level