See the API docs.
Latin-ish inputs should have very stable output. All inputs are passed through an NFKD transform, and anything still in the unicode Letter and Number categories are passed through intact. Anything in the Mark or Lm/Sk categories (modifiers) are skipped, and runs of characters from any other categories are collapsed to a single hyphen.