BUT Text-Dependent Speaker Verification System for SdSV Challenge 2020

In this paper, we present the winning BUT submission for thetext-dependent task of the SdSV challenge 2020. Given thelarge amount of training data available in this challenge, we exploresuccessful techniques from text-independent systems inthe text-dependent scenario. In particular, we trained x-vectorextractors on both in-domain and out-of-domain datasets andcombine them with i-vectors trained on concatenated MFCCsand bottleneck features, which have proven effective for thetext-dependent scenario. Moreover, we proposed the use ofphrase-dependent PLDA backend for scoring and its combinationwith a simple phrase recognizer, which brings up to 63%relative improvement on our development set with respect to usingstandard PLDA. Finally, we combine our different i-vectorand x-vector based systems using a simple linear logistic regressionscore level fusion, which provides 28% relative improvementon the evaluation set with respect to our best singlesystem.