INTERSPEECH 2011

We propose a strategy for discriminative training of the ivector extractor in speaker recognition. The original ivector extractor training was based on the maximumlikelihood generative modeling, where the EM algorithm was used. In our approach, the ivector extractor parameters are numerically optimized to minimize the discriminative crossentropy error function. Two versions of the ivector extraction are studied  the original approach as defined for Joint Factor Analysis, and the simplified version, where orthogonalization of the ivector extractor matrix is performed.
Bibliographic reference. Glembek, Ondřej / Burget, Lukáš / Brümmer, Niko / Plchot, Oldřich / Matějka, Pavel (2011): "Discriminatively trained ivector extractor for speaker verification", In INTERSPEECH2011, 137140.