ISCA Archive Odyssey 2008
ISCA Archive Odyssey 2008

MLLR techniques for speaker recognition

Marc Ferras, Cheung Chi Leung, Claude Barras, Jean-Luc Gauvain

Maximum-Likelihood Linear Regression (MLLR) and Constrained MLLR (CMLLR) have been recently used for feature extraction in speaker recognition. These systems use (C)MLLR transforms as features that are modeled with Support Vector Machines (SVM). This paper evaluates and compares several of these approaches for the NIST Speaker Recognition task. Single CMLLR and up to 4-phonetic-class MLLR transforms are explored using Gaussian Mixture Models (GMM) and large-vocabulary speech recognition Hidden Markov Models (HMM), using both speaker recognition and speech recognition cepstral front-ends and normalizations. Results for the individual systems as well as in combination with two standard cepstral systems are provided. Relative gains of 3% and 12% were obtained when combining the best performing CMLLR-based and MLLR-based systems with two standard cepstral systems, respectively.