That sounds pretty cool, though I'd be concerned it will suffer from the classic problem of current AI (...and humans, but that's by the by) of confident incorrectness. Like an automatic transmission can miss meanings and types of context that a human will spot, programmatically generating speech can probably mess up punctuation and flow - even the way a human reader sometimes will get part way through a sentence and realise they need to start again for it to come out right.

That said, I can't see it being a big problem for most works, just unfortunate here and there. For once it seems an AI application short on downsides! (Except for the usual economic ones for many people previously trained in the field.)