Modeling variability in speech recognition