This doesn't particularly matter since there's a perfectly good fallback for people who do happen to have "exotic" passwords: just type the password in correctly. As long as your mapping is consistent, it's totally fine if the bit sequence you're hashing is linguistically nonsense, because you're never going to display that to the user.

What you don't normally want to do is normalize passwords before hashing and then only store the hash of the normalized string, because that's fragile to changes in your normalization algorithm, e.g. updating your Unicode data tables.