> such as people trying to find meaning of misspelled words.

That worked just fine for quite a while. There's apparently enough misspelling in the training data, we don't need precise spelling for it. You can literally write drunken gibberish and it will work.