Abstract: Using larger scale datasets in the training stage of speaker verification model usually leads to better performance. However, when the speaker number of the training dataset becomes extreme ...