Improving the Effectiveness of Speaker Verification Domain Adaptation with Inadequate In-Domain Data

This paper addresses speaker verification domain adaptation with inadequate in-domain data. Specifically, we explore the cases where in-domain data sets do not include speaker labels, contain speakers with few samples, or contain speakers with low channel diversity. Existing domain adaptation methods are reviewed, and their shortcomings are discussed. We derive an unsupervised version of fully Bayesian adaptation which reduces the reliance on rich in-domain data. When applied to domain adaptation with inadequate in-domain data, the proposed approach yields competitive results when the samples per speaker are reduced, and outperforms existing supervised methods when the channel diversity is low, even without requiring speaker labels. These results are validated on the NIST SRE16, which uses a highly inadequate in-domain data set.