Maybe using an n-gram model to predict how probable the next word is, could pass high probability phrases at a faster rate, while slowing down for "harder" words
(That sounds like a cool idea, hit me up if you're game for hacking on something like that)
What's fascinating is that in the entropy/information-theoretic sense, instead of using wpm, one could aim for "constant" information rate of reading (as in bits/token). That's really cool as a concept. I don't know if it would work, obviously.
I'd definitely be up to for hacking this up, but currently a bit too busy. Will ping you at some point though!