Modified MFCCs for robust speaker recognition

Mel-scale frequency cepstrum coefficients (MFCCs) are commonly used featues in speaker recognition systems, but MFCC values are not very robust in the presence of noise. thus, the modified MFCCs (named as SMN-CMN-MFCC) based on the general noisy speech model is proposed in this paper, which uses spectrum mean normalization (SMN) to suppress the additive noise, and uses cepstral mean normalization (CMN) to remove the effect of convolutional noise. Theoretical analyses show that the combination of SMN and CMN can inhibit additive and convolutional noise at the same time. To verify the performance of the SMN-CMN-MFCC, we have conducted some speaker recognition tests. With the same convolutional noise component, the additive white noise experiments and the additive factory noise experiments show that SMN-CMN-MFCC provides 10.5% and 9.6% relative improvement than the conventional MFCC and ΔMFCC features, respectively.