2001: A Speaker Odyssey - The Speaker Recognition Workshop June 18-22, 2001 Crete, Greece

Integration and Imputation Methods for Unreliable Feature Compensation in GMM Based Speaker Verification

Andrzej Drygajlo, Mounir El-Maliki

Signal Processing Laboratory, Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland

This paper addresses the problem of text-independent speaker verification in the presence of unreliable (masked by noise) fea-tures. It presents and assesses several integration and imputa-tion approaches used for unreliable feature compensation in the framework of Gaussian mixture models (GMMs) of speakers. These approaches include marginalisation, bounded integra-tion, mean imputations, integrated speech-background model and Wiener filtering dependent on the most probable Gaussian component.

Full Paper   Presentation

Bibliographic reference.  Drygajlo, Andrzej / El-Maliki, Mounir (2001): "Integration and imputation methods for unreliable feature compensation in GMM based speaker verification", In ODYSSEY-2001, 107-112.