Do you get good results when adding scraped youtube audio? My model performance on LibriSpeech dev drops a bit when adding youtube audio to the training dataset ( my guess is likely due to poor alignment from auto generated captions ).

would it be possible to train on any of the more recent Text to speech engines out there? some of them are very realistic.

this would give you absolutely perfect sync down to the word, I assume... I don't know about the cost if you paid ratecard though, perhaps you can do some partnership with them since yours is a symetrical product