You asked why not. That's the answer -- why should platforms create an overcomplicated ML-based solution to try to screen read (assuming such a thing is even possible, which I don't believe it is), when existing solutions using standard frameworks work fine and _don't_ require devs to spend extra time?