Distant Speaker Verification Using a Combined Family of MVDR Estimates

Abstract

Distant speaker verification involves explicit spectral estimation of speech acquired over microphone arrays. The choice of the appropriate set of microphones is important here. In this paper we describe an implicit approach to minimum variance distortionless response (MVDR) spectral estimation of distant talking speech and its application in distant speaker verification. A mathematical formulation for computing an implicit spectral estimate for speech acquired over a uniform linear array (ULA) is first presented. This formulation is based on a simple mathematical relation between a fixed order MVDR spectral estimate, the harmonics in speech, and the noise power. This relationship is used for spectral modeling of distant talking speech by jointly combining a family of MVDR estimates and the number of elements in the ULA. The performance of the proposed implicit MVDR spectral estimation method is evaluated in terms of cepstral distance measure indicating improvements over the Fourier spectral estimates obtained from the individual elements of the ULA. Experiments on distant speaker verification using speech data from the NIST 2004 corpus indicate reasonable improvements when compared to conventional MFCC from the individual elements from the ULA.