Yes, dealing with the various disabled users is tough because in order for them to make use of it, they need the computer to see it (and then be spoken in the case of someone that is blind - or sight limited or whatever the current PC terms are), and if the computer can see it - then any bot can see it.
I guess I should consider myself fortunate that I don't ever have to program for that - every application I have written has been for an environment where sight is assumed.

The advantage over using an image is that it can't be scanned by a bot and read - if you have an image on a page, the bot can look for the image and then pull the data from the image (easiest way is using neural net training - well, I guess not "easiest" but most effective for varied image types).

But if there is no image there, then that particular bot can't find anything. The text is also not on the page, so it can't find anything either. The bot then has to parse the appropriate content on the page (which is easy if it is the only thing on the page, harder as you add more content and dynamically change how you reference the classes) and rebuild it as an image, and then do the analysis on it.
There are ways of making it much harder for the bot to rebuild it.

Yeah, it is about 10K to represent the same as what a 1K PNG could have done - certainly not ideal for showing images - but this wouldn't be something that you would do on every page either. That is about a 2 to 3 second download for a 33kbps modem user.