Bayesian Adaptation of PLDA Based Speaker Recognition to Domains with Scarce Development Data
Recently, speaker verification based on i-vectors and PLDA has become state-of-the art. This approach relays on models whose parameters need to be estimated from a development database with a large number of speech segments and speakers. That is one of the reasons why it has been very successful on NIST evaluations where we have sufficient data available. However, when we need to do speaker verification in a domain where the development data is scarce, training accurate models is complicated. In this paper, we propose a method to do Bayesian adaptation of the PLDA parameters from a domain with sufficient development data to a domain with scarce development data. The method is based on the variational Bayes recipe. We perform experiments adapting models trained with the NIST databases to the EVALITA09 database. Results show interesting improvements.