Couldn't you just do a bunch of different faxes, perhaps in different fonts or different font sizes, which would lead to different randomly distributed errors? Then you can do OCR for all of them, and just take the median of the result, and get exponentially less error.