AI isn't going to be able to do that. As good as it may get, it won't be able to read the mind of the author. It's going to be even more generic than a human reader.
But the LLMs don't, do they? If anything, they're written-first, or even on some deep level, binary-first. I don't think what they write even has "a way they meant the words to be emphasized in speech."
Why would we assume what comes from them will even aspire to "being like humans"?
The number of reasons AIs might not add the same things to an audio version of a book (the context we're talking about in this thread) is essentially infinite. It seems vastly more likely that they won't add what the author adds than that they will.
Humanity may not have a moat, but each individual human does, especially when it comes to art, where I'd include writing.
I can't say anything an an AIs aspirations, but that fact that we're imbuing them with all of our collective data, means they will be skewed to perceive the world similarly us, at least initially.
Autobiographies? Fine, but most of the time they are usually read by their authors.
I would like train a model on Allyson Johnson’s voice (narrated the Honor Harrington books) and then use that to re-narrate the 1-2 books in one of the spinoffs (I think it was in the Saganami Island series?) where they used a different narrator (who was horrible).
I also might be interested in using it to clean up the Wheel of Time series where, while it’s the same 2 narrators, they change the pronunciation of various names/words book-to-book. “Moghedien” being the one that stands out most. They pronounce it at least 3 different ways:
* Mo-gid-e-on
* Mo-ga-dean
* Mog-a-din