Linguistic versus personal variation in speech recognition