Extension and further analysis of higher order cepstral moment normalization (HOCMN) for robust features in speech recognition