I was asking because back around 2008 when I was more involved with Accessibility on the Web, we built a website that had a legal requirement to be as accessible as possible. So therefore we were trying to generate accessible double-paged PDFs, voicing over DAISY books and all. And doing so was so much work. We spent thousands of man hours just on document conversion, even when the underlying source format for the documents was RTF which is at least theoretically easy to parse in regards to layouting.
Every time we tried to make things compatible with JAWS, we realized that JAWS was just a pile of dirty Trident hacks that wasn't integrated as nicely as someone would expect such a software to be.
It was before the rise of AI/CNNs so therefore converting a vectorized PDF back into a semantic one was totally impossible. These days tesseract seems to make huge progress, but is still unusable for the task in practice due to its high failure rate in recognized words that you cannot fix with tricks like a Levenshtein distances or dictionary statistics.
Eversince I've been more on the Linux side of things, though. Here the ecosystem is so bad that I cannot even start to describe it. Most TTS engines are literally from the last millenia, and projects like Orca aren't made for anything serious when trying to embed it into your software to give users more access and control.
Maybe you have also some hints here? Are there better alternatives that I'm not aware of?