Abstract:

This paper proposes to embed an auto-associate neural network (AANN) in the Gaussian mixed
model-universal background model (GMM-UBM) for speaker verification. The scheme integrates the merits of
both GMM and AANN. GMM and AANN are trained as a whole in terms of maximum likelihood (ML), and
spatial information of the feature vectors is disclosed. AANN reshapes distribution of the data and improves
similarity of intra-class data. In the training, parameters of GMM and AANN are updated alternately. Experimental
results show that the proposed method together with TNorm can provide improvement of verification
rate by 26% over the baseline GMM-UBM.

Key words: speaker verification, Gaussian mixed model-universal background model (GMM-UBM), autoassociate, neural network (AANN), embed