Semi-Supervised Far-Field Speaker Verification with Distance Metric Domain Adaptation

Achieving high performance in far-field speaker verification is challenging due to the scarcity of data and prevalence of low-quality labels. A common approach is to use a semi-supervised learning framework. A pre-trained near-field speaker model extracts embeddings of unlabeled far-field speakers, and clustering methods generate pseudo-labels for fine-tuning. The accuracy of pseudo-label generation through clustering and the domain adaptation during fine-tuning are critical to this semi-supervised learning framework. In this paper, we propose a method that combines Uniform Manifold Approximation and Projection (UMAP) with a community detection algorithm for efficient dimensionality reduction and clustering. UMAP preserves the structural distribution of high-dimensional data, facilitating better clustering. We conduct domain alignment in the distance metric space to achieve domain adaptation during fine-tuning, referring to this approach as distance metric domain adaptation (DMDA). We evaluate the effectiveness of the dimensionality reduction community detection agorithm and the on VoxCeleb (pre-trained) and FFSVC (fine-tuned) datasets. Experimental results demonstrate that the accuracy of generating pseudo labels is improved by using the reduced dimensional community detection algorithm. Additionally, the DMDA method effectively reduces domain mismatch issues during the fine-tuning process.