Many of the readers of Librivox are not great, but I'd rather have the worst of them over AI.
Even a future better AI that knows more about when to pause and how to pronounce things the right way from context, will still be lifeless because it can't actually understand the full context, only the surrounding words. The actual full context of a given word or phrase requires not only pattern-matching the surrounding words, but having a sense of self and inserting that self into the described situation, and applying that selfs response to that, which will be not only more lifelike and realistic and meaningful, but also be a bit different for every different human reader. Good readers aren't simply doing it more correctly, their own unavoidable interpretation of the text is in their reading, even if they are trying to be a robotic neutral newscaster.
Sometimes I'm actually annoyed by how a reader reads a given sentence differently than I would have. Like they failed to realize what the thrust of the thing was and focussed on some other part. But even those are better than an AI that will either always apply the same algorithm, or worse, maybe varies it but only randomly, not as a result of having a personality that colors everything.
I guess at some point they will be able to make AI's that can fake the quasi-random-consistency of a personality. If someone can describe the problem like I just did, then someone else can code up a response to it. And then it will be good enough to essentially fool most not-very-perceptive people, and that will suck because it means we won't have a choice to avoid that crap any more. Much content will simply only be produced this way, because it will be cheaper to make and good enough for most people, and so will sell. Once it's good enough to sell, that's it, that's what happens almost exclusively after that. Like bluetooth headphones.