[an error occurred while processing this directive] [an error occurred while processing this directive]
Kaizhi Qian

Regularized Estimation of GMMs for SVM-based Speaker Recognition

Kaizhi Qian, 10/28/2014, 4:30-5:30pm, BI 2369

Speaker adaptation based on the Universal Background Model (UBM) has become a standard approach for speaker recognition. A GMM supervector is constructed by normalizing and stacking the means of the adapted mixture components, which provides a compact representation of the speaker-dependent model in speaker recognition tasks. The estimation of the unknown GMM parameters is usually obtained by the method of maximum a posteriori estimation (MAP), which can be regularized to increase the model interpretability with insufficient training data. In this thesis, the speaker-adapted models are estimated using the MAP with L1-regularization, referred to as the elastic net, based on the assumption that the distinctions between any two speakers are sparse. Experiments on the NIST2008 speaker recognition evaluation task show error rate reduction with the elastic net.

[an error occurred while processing this directive]