It would be better if we had a much larger dictionary, since we could map more bytes to a single word -- with a list of 65,536 words we would need only 8 words per 16 byte hash, which I think is a lot more manageable.
Note that I never intended this to have all the uniqueness of the original digest. It's merely supposed to be 'good enough' for UI purposes.
/usr/share/dict/words is likely to be pretty big, over 200k words, but it's still human readable. You could prune it by removing proper words (i.e. anything capitalized).
https://en.wikipedia.org/wiki/English_phonology#Phonotactics
I think this is getting a little beyond my linguistics skills...
Edit: I suppose which syllables are easily corruptable depends on the medium: written hashes would require a different set of graphemes than spoken hashes would phonemes, for example, while memorised hashes would likely be different again. I wonder if there's been any research or anything into this.