Improving Robustness of Speaker Recognition to New Conditions Using Unlabeled Data

ALICIA LOZANO DIEZ; MITCHELL MCLAREN; AARON LAWSON; DIEGO CASTÁN; LUCIANA FERRER

Estocolmo

Congreso; Interspeech 2017; 2017

ISCA

Unsupervised  techniques  for  the  adaptation  of  speaker recognition are important due to the problem of condition mismatch that is prevalent when applying speaker recognition technology  to  new  conditions  and  the  general  scarcity  of  labeled ?in-domain?  data.   In  the  recent  NIST  2016  Speaker  Recognition  Evaluation  (SRE),  symmetric  score  normalization  (S-norm)  and  calibration  using  unlabeled  in-domain  data  were shown  to  be  beneficial.   Because  calibration  requires  speaker labels for training, speaker-clustering techniques were used to generate pseudo-speakers for learning calibration parameters in those cases where only unlabeled in-domain data was available. These  methods  performed  well  in  the  SRE16.   It  is  unclear, however, whether those techniques generalize well to other data sources.  In this work, we benchmark these approaches on several distinctly different databases,  after we describe our SRI-CON-UAM team system submission for the NIST 2016 SRE. Our analysis shows that while the benefit of S-norm is also observed across other datasets, applying speaker-clustered calibration provides considerably greater benefit to the system in the context of new acoustic conditions.