Just as with redacted documents (consistently blocked terms) or bad OCR jobs (wrong or missing characters), even if only a certain percentage comes out unmangled it is more readable than having no data at all.

A stable base corpus and some dynamic programming will allow you to clean up the remainder[0].

[0]: http://stackoverflow.com/a/11642687/2449774