But I disagree with you when you write "it simply doesn't have the information to improve beyond sounding like a human reading words fluently" - it has the same information when reading it as a human does, meaning that the best implementation would have to not only adapt tone to explicit instructions like "... she shouted", but also read between the lines / make subjective choices to suit the different characters.
AI is already capable of doing sentiment analysis on text, and text to speech models are getting better at being able to simulate moods/emotions rather than just speaking flatly, and I don't think we're many years away, if that, from those two sides being paired together in a way that produces the sort of quality output we're talking about for the first time without human involvement. Add to that the fact that AI can train on the many good examples of humans reading things, they may get to the point of emulating not just the core accent but also how each accent should adopt to what meanings in the text and arrive at a great solution without even needing to go through the steps of analysing what the text means to use that to know how to modify the voice being generated.