Frequency-warped spectral distance measures for speaker verification in noise

Spectral distance measures on a warped frequency scale are investigated to improve noise robustness in speaker recognition. Frequency warping is applied to obtain a less noise-affected spectrum; higher energy parts, of the spectrum like those around formants, which are less affected by noise are expanded, and lower energy parts, which are corrupted by noise, are shortened. Linear predictive coding analysis on a warped frequency scale can be utilized to obtain frequency-warped spectral distance measures in which an autocorrelation function derived from a frequency-warped fast Fourier transform spectrum is used for the all-poles modeling. Such spectral distance measures are shown to provide significant performance improvement over conventional distance measures on the linear frequency scale in speaker verification in noise. Weighted cepstral distance measures are also discussed, focusing on their effectiveness for speaker verification