I came up with a mutation-based solution to this. Supply any string and it finds a valid hash in as few edits as possible.

Example: http://bochs.info/img/mutation-20140606-024906.png

One could definitely optimize this to be less destructive and produce more pronounceable results. It's basically two pieces: an engine for suggesting mutations, and a simple algorithm to score and pick mutations. Changes to either half (vowel distribution, ngrams, etc) could result in better strings.

(fyi, this kind of attack is a big reason to use cryptographic hashes: http://en.wikipedia.org/wiki/Cryptographic_hash_function)