8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


Dependence of GMM Adaptation on Feature Post-Processing for Speaker Recognition

Robbie Vogt, Jason Pelecanos, Sridha Sridharan

Queensland University of Technology, Australia

This paper presents a study on the relationship between feature post-processing and speaker modelling techniques for robust text-independent speaker recognition. A fully coupled target and background Gaussian mixture speaker model structure is used for hypothesis testing in this speaker model based recognition system. Two formulations of the Maximum a Posteriori (MAP) adaptation algorithm for Gaussian mixture models are considered. We contrast the standard single iteration adaptation algorithm to adaptation using multiple iterations. Three post-processing techniques for cepstral features are considered; feature warping, cepstral mean subtraction (CMS) and RelAtive SpecTrA (RASTA) processing. It is shown that the advantage gained through iterative MAP adaptation is dependent on the parameterisation technique used. Reasons for this dependency are discussed.

Full Paper

Bibliographic reference.  Vogt, Robbie / Pelecanos, Jason / Sridharan, Sridha (2003): "Dependence of GMM adaptation on feature post-processing for speaker recognition", In EUROSPEECH-2003, 3013-3016.