Just my personal experience: If anyone remembers how hard it was to make things render consistently across IE and Netscape, the current situation with different screen readers is 1000x harder.

I mainly test with NVDA on the latest Firefox ESR on Windows 10. Something always breaks if I upgrade any part of this chain in my testing VM. Few months ago I spent many days to make abbreviations in a table header to be read correctly. I tried the abbr tag, abbr attribute, aria-label and many other possibilities and found a super hacky solution that involves hiding the expanded form with some css while exposing it to the screen reader. Then sometime after that I updated the Firefox. At the same day a colleague questioned why we needed all these hacks, and I wanted to show her how broken it was without my hack, but voila it had started working with the non-hacky, semantic way.

I have many stories like this.

I also do test with other configurations but just before releasing. VoiceOver (mac) is 80% there. JAWS Inspect is giving some colleagues headaches. None work 100% without breaking others, so the most popular stack decides how we implement things.

When I saw this post, I thought perhaps doing everything with spans and adding necessary attributes like role would be easier. Well nobody is passing something like that through a code review :)