A perceptual masking approach for noise robust speech recognition