Is masking a relevant aspect lacking in MFCC? A speaker verification perspective

We hypothesize that spectral masking may account for most of the gains in robustness against noise using ensemble interval histogram (EIH) and zero crossing with peak amplitude (ZCPA) compared to Mel-frequency cepstral coefficients (MFCCs). To test this hypothesis, we focus on this issue by comparing two MFCC implementations for which the only difference is spectral masking. The comparison involved biometric speaker verification tasks using two publicly available databases. The results confirm the superiority of MFCC with masking, thus corroborating our hypotheses that masking is a key aspect for improved robustness in feature extraction.

► Masking should be a pivotal concern for robustness issues in speaker recognition.
► Fast masking effect can be obtained through a slightly modified MFCC extractor.
► Rectangular masking windows outperform usual triangular ones under strong noise.
► Under additive noise a good masking bandwidth is roughly twice the critical band.