Allpass modelling of Fourier phase for speaker verification

This paper proposes features based on parametric representa-tion of Fourier phase of speech for speaker verification. Directcomputation of Fourier phase suffers from phase wrapping andhence we attempt parametric modelling of phase spectrum us-ing an allpass (AP) filter. The coefficients of the AP filter areestimated by minimizing an entropy based objective functionmotivated from speech production process. The AP cepstralcoefficients (APCC) derived from the group delay response ofestimated AP filter are used as features for speaker verification.An i-vector based speaker verification system is employed toevaluate the performance of the proposed APCC features onNIST 2003 speaker recognition evaluation database. The equalerror rates (EER) obtained from speaker verification systemsbuilt using APCC features and baseline mel-frequency cepstralcoefficients (MFCC) are reported. A relative improvement of12% was obtained over MFCC features by combining evidencesfrom both MFCC and APCC based system