With the rate speech synthesis is developing, it's probably going to be hard to distinguish between human speakers and synthetic speakers in a few years from now.
I'm really looking forward to those actually being available to use.
I've just started to use AWS Polly in some demos, but the speech synthesis quality feels 2-3 years old at least :/
I think what's missing is to understand and adapt the speech rhythm and pauses based on the language. At first sight I would say that this requires human intuition, but then again, lots of literature is repetative, so I think ML could actually work quite well well, with actual understanding of the text, given enough data.