Robustness to additive noise of locally-normalized cepstral coefficients in speaker verification

In this paper the performance of a new feature set, Locally Normalized Cepstral Coefficients (LNCC) is evaluated for a speaker verification task with short testing utterances in additive noise. The results presented here show that LNCC outperforms baseline MFCC features when SNR is lower than 15 dB. The average relative reduction in EER achieved by LNCC is 33%. The use of LNCC in combination with spectral subtraction provides a reduction in EER averaging 18% when compared to MFCC features also with spectral subtraction. In addition, sub-band LNCC is proposed to improve the estimation of noise energy and hence the effectiveness of spectral subtraction. When compared with MFCC features, the use of sub-band LNCC led to greater reductions in EER than LNCC with non-stationary noise.