I don't know, how about using a dictionary and doing research?

Otherwise wouldn't every commonly misspelled word show up incorrectly in spellcheck?

There is quite a bit of work in the field of speech recognition, too, and it's entirely possible to create a functional model without recording every user's voice and storing that information on a server. Even if they require speech samples, or samples from different languages, they should at least pay someone for that information.

I was disappointed when I learned that the voice recognition feature in Siri wasn't local to the phone.