Yes, I was thinking of speech synthesizers when I wrote my comment :) Closing the gap is devillish and requires increasingly complex grammatical analysis and more and more rules. One of the big barriers for early speech synthesizers was enough RAM/ROM to cram all the text-to-phonetic transformation rules.

I believe that modern approaches mostly use deep learning. And definitely still aren't perfect. I don't use speech synthesizers regularly personally as a tool, but synthesized speech is common enough that I do hear obvious stress errors in synthesized speech every day. Probably because English orthography is most opaque about stress, which basically doesn't exist as far as the writing system is concerned.