8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

SVM Kernel Adaptation in Speaker Classification and Verification

Purdy Ho, Pedro Moreno

Hewlett Packard Labs, USA

Techniques for speaker classification and verification based on discriminant based kernel methods such as Support Vector Machines are becoming more and more popular. However, when compared with state of the art statistical based techniques such as Gaussian Mixture Models, their performance suffer for two main reasons: first their inability to scale up and handle a large number of classes, and second their inability to adapt model parameters. In this paper we address the second issue.Previously [8] we have introduced a kernel based classifier that combines the best of generative methods and discriminative classifiers. Each utterance is fitted with a generative model such as a Gaussian Mixture Model (GMM) and a kernel distance is defined among GMMs. In this paper we extend this kernel with the ability to adapt to the speaker utterance by adapting the utterance GMM using Maximum Likelihood Linear Regression (MLLR) techniques. Our experimental results on two different speaker databases show that kernel adaptation is a promising technique highly effective on long utterances when compared with non-adapted kernels.

Full Paper

Bibliographic reference.  Ho, Purdy / Moreno, Pedro (2004): "SVM kernel adaptation in speaker classification and verification", In INTERSPEECH-2004, 1413-1416.