I-Vector/PLDA Variants for Text-Dependent Speaker Recognition

The i-vector/PLDA approach currently dominates the field of text-independent speaker recognition and the question of how to translate this methodology to the text-dependent domain has recently become an active area of research. The essential difference between the two fields is that it is possible to do speaker recognition with enrollment and test utterances of very short duration in the text-dependent case but not in the text-independent case. The i-vector representation of short utterances turns out to be very sensitive to their phonetic content and this introduces a major source of nuisance variability when i-vectors are used in text-dependent speaker recognition. We show how, despite this complication, i-vector extractors can be successfully trained on short utterances (rather than on whole conversation sides as is usually done) and how this source of nuisance variability can be dealt with successfully in a PLDA classifier by making the PLDA model parameters phrase-dependent. Our results show that this phrase dependent version of PLDA is capable of outperforming the speaker-phrase version of PLDA prePreprint submitted to Computer, Speech and Language November 6, 2013 sented in [8] on the RSR2015 dataset. We also give a detailed account of uncertainty propagation in PLDA and we show that it combines very successfully with phrase-dependent PLDA.