Make that privacy and accents. There are too many regions where choice recognition just won't work without essentially building new voice models. And that's before we get to people using their second/third/... language.