You should probably also go Bayesian on word frequencies for the suggestion ordering. Maybe also on error types (missing letter, repeated letter, nearby (on standard keyboard) letter.
Your comment and the reply to it are good if we're talking spell checker software for stuff people type.
What I have to deal with as well are recognition errors. From trivial replacements (I/l/1,C/G,i/j,P/R etc.) to 'engine said that there's some character here according to the character segmentation results, but it has no damn clue what character it could be' that's a different look at the same problem.
Also, for typed content I see quite a lot of shortcuts. SwiftKey for Android, nice as it is, seems to be too braindead to correct words that _start_ with the wrong character (if I type vharacter, with just one single slip at the start, it fails miserably and looks baaad imho). Assumptions like 'just check adjacent characters on the keyboard (special points for assuming a keyboard layout)' or 'ah, the user is certainly starting the jnput correctly and just might misspress a letter in the middle of a word' are nice models to make the developer's life easier, but they don't work for a good number of cases as well..
I used soundex and metaphone to correct spelling mistakes for a project. My assumption: that most spelling mistakes are people trying to sound out the spelling and not simple typos. For the limited case I was solving (mostly Biblical names) it worked quite well.
