Cross-Domain adaptation in Distance Space for Speaker Verification

Significant performance degradation often occurs when a well-trained speaker verification system is applied to an unseen domain. Data augmentation and domain adaptation are two common approaches to tackling this problem. However, data augmentation would not be helpful when language mismatch rather than environment noise causes the domain shift. Domain adaptation also suffers from a label mismatch problem, making feature distribution alignment unreliable. We propose incorporating a distance metric space into model adaptation to address these issues. The idea is to align not only the embeddings across domains but also the distributions of their pairwise distances, resulting in embeddings tolerant to domain shift. To validate the idea, we used the non-Chinese utterances in VoxCeleb2 and the Chinese utterances in CN-Celeb2 as the source and target domain training data, respectively. Results show that the alignments reduce the EER on the CN-Celeb1 test set by 15.2%.