At tomorrow's TeX Hour I'll present my solution, via the concept of rich DVI files. This goes back 25 years, to my TUGBoat article "Editing .dvi Files, or Visual TEX". Although in some ways dated, the basic concepts are still sound and even more relevant than before.
My approach is to pre-process the source tree so that every word (and similar source objects) is tagged in the resulting DVI file. This will then give a rich DVI, from which an external program can create an accessible PDF. Frank's approach is similar, except that he uses complex and ingenious LaTeX macros for everything, and tags only the paragraphs.
Frank's approach leads to a multi-year journey to modernize LaTeX, and avoids the need for a post-processor. My approach can use current LaTeX, by use of an input filter to add tagging, and an output filter to render the rich DVI to accessible PDF.
This is an important choice, which will affect TeX for over 30 years. If you're keen, you can read about these two approaches at
Arising from that is a project to help users find existing advice on LaTeX best practice,