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.