Supervised imbalanced multi-domain adaptation for text-independent speaker verification

Speaker verification is an important recognition task in speech signal processing. Domain adaptation for speaker verification is challenging and it is one of the practical problems put forward in the INTERSPEECH2020 Short-duration Speaker Verification (SdSV) Challenge 2020. Although there are several previous researches focused on the domain mismatch problem of the speaker verification task, many methods are not easy to show effectiveness in the real conditions. This is due to the suboptimal loss design, as well as the real-world datasets could contain multiple domains and imbalanced data in each domain. We have explored various domain adaptation methods and proposed one that is both effective and robust in this task by optimizing loss design and explicitly considering the data-imbalance problem. The proposed method is also designed to fit the scenarios where the datasets contain multiple domains. Significant single-model performance improvements have been observed by evaluating on the SdSV20 challenge testbench with our proposed method.