Advanced Fill in the Blank (accessibility)

I'm testing theaccessibility of the Advanced Fill in the Blank learning object, and have a few suggestions to improve accessibility. As is it's somewhat accessible, but would be largely unusable with a screen reader.

When a correct choice is made, feedback should be provide, first announcing "correct" then reading the whole sentence. aria-live, or perhaps role="alert", might be used with the correct checkmark image with alt text "correct" (or other readble text), and script could select the sentence so it gets read as a whole. As is, there is no indication for screen reader users that the section was correct, without being able to see the green select menu or the check mark.

Similarly, when an incorrect selection is made, there's no indication by the screen reader that an error has occurred. Part of the sentence is highlighted, and a small dialog opens with details on why it's an error. None of this info is read by the screen reader,  making it difficult to know what just happened after making an incorrect selection. If one navigates to the end of the object, then one can tab navigate into the error feedback, though that's unlikely to be useful for a screen reader users, or at least require significantly more effort. In this case, the "incorrect" alt text (or other readable text) with the "X" could be read using role="alert",  the highlighted text could recieve focus so it gets read, and the error feedback dialog could be setup with aria-live="polite" so immediately after reading the focused text, the content of the feedback dialog would read out. Alternately, the error feedback dialog could be added to the tab sequence, so it could be reached with the tab key without having to navigate to the bottom of the object.

Tested with current VoiceOver & Chrome, Jaws + Chrome, NVDA + FF. I'm open to providing some feedback on an updated version of this object, if you need expert feedback.