Google also has a ton of computing power and disk space...
That said, homophones and homonyms only would make the 3-gram sets smaller, but it would only detect homophones and homonyms being used incorrectly. I could still use a word incorrectly, and if I'm banking on that software to detect my mistakes it wouldn't.
Indeed. My false positive rate [flagged errors that are really omissions from the dict.] on spell-check in-browser appears to have been about 95% over the past couple of months. I wouldn't be expecting perfection; surely some mistakes corrected is better than none.
How I envision it is also a learning tool - "can here you" would pop up a "can {here} you [here refers to location, hear to hearing sound; 'can hear you']" allowing the author to click the 'corrected' phrase. At the same time you're learning the distinctions.
That said, homophones and homonyms only would make the 3-gram sets smaller, but it would only detect homophones and homonyms being used incorrectly. I could still use a word incorrectly, and if I'm banking on that software to detect my mistakes it wouldn't.