This paper proposes a new feature aggregation mechanism for deep neural network based speaker embedding for text-independent speaker verification. In speaker verification models, frame-level features are fed into the pooling ...