- It reads footnote links ("one", "fourteen", etc)
- It reads years like numbers ("One thousand, seven hundred and fifty six" instead of "seventeen fifty six").
- It reads introductory metadata, warnings, etc.
- It reads tables.
- It read image captions.
Ideally it wouldn't do any of these things. Basically, it reads the screen from top to bottom, which isn't what a human would do if you asked somebody to read a Wikipedia article to you. If anyone knows a good way to synthesize Wikipedia to speech, that would be so nice!
EDIT: I wonder if Apple would reject such an app from the App Store because it would compete with ebooks.