The CORAL+ Algorithm for Unsupervised Domain Adaptation of PLDA

I4U Submission to NIST SRE 2018: Leveraging from a Decade of Shared Experiences

“…In general, sub-systems that utilized more recent xvector deep speaker embedding were more successful. On the CMN2 partition, the CORAL+ [16] unsupervised PLDA adaptation technique has shown to be effective. The VAST partition is more difficult compared to the CMN2.…”

“…By domain adaptation, we assume that the channel variability learned from one domain shares some common behaviors in another domain. Simple covariance transformation techniques [15,16] have shown to work well compared to a much complicated counterpart [31]. This is a topic for future research.…”

“…Given SRE18-CMN2-Unlabeled, an unlabelled set of indomain data (see Table 1), the central idea of domain adaptation is to estimate the in-domain between and within-speaker covariance matrices from the in-domain, yet unlabelled, dataset with some helps from out-of-domain covariance matrices. In I4U SRE'18 submission, two unsupervised domain adaptation techniques have been found to be useful, namely, (i) modellevel correlation alignment with CORAL+ [16], and (ii) Kaldi's PLDA adaptation 1 . We refer the interested reader to [15,16,17] and references therein for more details.…”

“…In I4U SRE'18 submission, two unsupervised domain adaptation techniques have been found to be useful, namely, (i) modellevel correlation alignment with CORAL+ [16], and (ii) Kaldi's PLDA adaptation 1 . We refer the interested reader to [15,16,17] and references therein for more details.…”

“…In the past decade, I4U participated in five SREs, namely SRE'08, 10,12,16, and 18 [6,7,8,9,10]. Aside from a joint submission, the I4U consortium was formed with a common vision to promote research collaboration and facilitate active exchange of information and experience towards the open evaluation of speaker recognition technology.…”