I'm working on a speaker verification system using X-vector on Korean speakers' data. On unseen data, the system achieves fairly good EER (from 10% to 20%), but I'm encountering a challenge with close embedding similarity. For example, many positive and negative pairs have cosine similarities of 0.96 and 0.93, respectively. In this case, decicion threshold can be chosen at 0.94 which helps achieve good EER. Nonetheless, I find the high threshold to be quite unintuitive.
Additionally, I tried using a pretrained ResNet on the Vox trainer, which yielded a low equal error rate (EER). However, the threshold for acceptance was set at 0.94.
Could someone provide advice on how to address this issue? I feel stuck.