On desensitizing the Mel-cepstrum to spurious spectral components for robust speech recognition