Deep Discriminative Embedding with Ranked Weight for Speaker Verification

Deep speaker-embedding neural network trained with a discriminative loss function is widely known to be effective for speaker verification task. Notably, angular margin softmax loss, and its variants, were proposed to promote intra-class compactness. However, it is worth noticing that these methods are not effective enough in enhancing inter-class separability. In this paper, we present a ranked weight loss which explicitly encourages intra-class compactness and enhances inter-class separability simultaneously. During the neural network training process, the most attention is given to the target speaker in order to encourage intra-class compactness. Next, its nearest neighbor who has the greatest impact on the correct classification gets the second most attention while the least attention is paid to its farthest neighbor. Experimental results on VoxCeleb1, CN-Celeb and the Speakers in the Wild (SITW) core-core condition show that the proposed ranked weight loss achieves state-of-the-art performance.