%Aigaion2 BibTeX export from Idiap Publications
%Saturday 21 December 2024 06:22:57 PM

@TECHREPORT{Sarfjoo_Idiap-RR-15-2019,
         author = {Sarfjoo, Seyyed Saeed and Madikeri, Srikanth and Hajibabaei, Mahdi and Motlicek, Petr and Marcel, S{\'{e}}bastien},
       keywords = {adaptation, batch normalization, speaker recognition},
       projects = {Idiap, ODESSA, EC H2020-ROXANNE},
          month = {11},
          title = {Idiap submission to the NIST SRE 2019 Speaker Recognition Evaluation},
           type = {Idiap-RR},
         number = {Idiap-RR-15-2019},
           year = {2019},
    institution = {Idiap},
        address = {Rue Marconi 19, 1920 Martigny},
       abstract = {Idiap has made a submission to the conversational telephony speech (CTS) challenge of the NIST SRE 2019. The submission consists of six speaker verification (SV) systems: four extended TDNN (E-TDNN) and two TDNN x-vector systems. Employment of various training sets, loss functions, adaptation sets and extracted speech features is among the main differences of the submitted systems. Domain adaptation is represented by a supervised method (developed using a limited data) with transfer learning of the batch norm layers. \% was applied. 
Mean shift and covariance estimation of batch norm allows to map  the target domain to the source domain, alleviating the problem of over-fitting on the adaptation data.  
The back-end of all the systems is represented by the conventional
Linear Discriminant Analysis (LDA) projection followed by Probabilistic LDA (PLDA) scoring for inference. The PLDA was also adapted unsupervisedly using the unlabelled part of the NIST SRE 2018 set. In addition, training the LDA and PLDA using in-domain data was investigated. The entire system was built around the Kaldi toolkit.},
            pdf = {https://publications.idiap.ch/attachments/reports/2019/Sarfjoo_Idiap-RR-15-2019.pdf}
}