y'know, it's farcical that the unicode math/cursive text people use on Twitter and the like is not a solved problem for screen readers from trillion dollar companies. It was farcical ten years ago too.

that's 52 letters at most for those examples, each of them announce all the unicode description including the letter. The Machine is perfectly able to be made to know that it's that letter, it's in the name, it can be picked up with probably one search pattern per character set. You could have a feedback form where people just give you the set of unicode text they saw someone use and hardcode that too. it's not like there's a huge chunk of them people use.

"math people might use them" math people use the block letters, and even those probably should just be parsed as their letter if it's in a word. I get that screen readers right now use some black box ML magic but you can... use the lookup table before you send the data.

Update: so, I checked and the google voices don't do it in Reader Mode, but they do with talkback. So does espeak-ng(? they've got the same names as the espeak voices) voices through firefox's tts via readaloud. They just pronounce them as words, at least when it's reading a Firefox window. Which fonts break which readers? is it a browser thing?

these still break on block/circle text, but it just doesn't render it as a word at all instead of spamming them, which... better and worse I suppose

Found a working windows machine: narrator fails the test
Remembered my ipad was an ipad: voiceover does too

I don't have a mac to test and getting orca/jaws running on my hardware is an issue since I uh. have neither an insert key or a capslock key on any of my machines and iirc they need that for navigation. Narrator does too, but it took me a lot of fighting with it to make it work without the key.

man.