It looks pretty much the some, except that you assume the input is already in your library's canonical encoding (probably utf-8 nowadays).

I realize this sounds like a total cop-out, but when the use-case is destructively best-effort tokenizing an input string using library functions, it doesn't really matter whether your internal encoding is utf-32 or utf-8. I mean, under the hood, normalize still has to map arbitrary-length sequences to arbitrary-length sequences even when working with utf-32 (see: unicodedata.normalize("NFKC", "a\u0301 ffi") == "\xe1 ffi").

So on the happy path, you don't see much of a difference.

The main observable difference is that if you take input without decoding it explicitly, then the always-decode approach has already crashed long before reaching this function, while the assume-the-encoding approach probably spouts gibberish at this point. And sure, there are plenty of plausible scenarios where you'd rather get the crash than subtly broken behaviour. But ... I don't see this reasonably being one of them, considering that you're apparently okay with discarding all \W+.