SVM-based Speaker Classification in the GMM Models Space
Nir Krause, Ran Gazit · 2006
This paper describes a new approach to speaker classification, based on using an SVM classifier over the GMM models space. Adaptation of a speaker-independent GMM universal background model with speaker specific data creates a speaker-dependent GMM model. The vector representation of this model is used by an SVM classifier to recognize the speaker. When used with multiple, channel-specific background models, this scheme has the potential to improve speaker recognition performance in channel mismatch conditions. Performance improvement is demonstrated over a multi-channel corpus, as well as over the NIST 2004 evaluation data