Robust speech recognition using missing features