A speaker verification method using frame-level self-attention

A new method of applying the self-attention mechanism for text-independent speaker verification is presented in this paper. There are differences between the speech frames, which may affect the performance of the speaker verification system. Generally, to capture the differences, the speaker verification system needs to perform a weighted average for the frame-level output in the pooling stage of extracting the speaker embedding, and the weights are learned through the self-attention mechanism. In this work, the self-attention mechanism is introduced into the frame-level model in the system to extract frame-level embeddings, which provides a new way of capturing the differences of speech frames for speaker verification. A self-attention layer is added between the frame-level layers to directly obtain different frame-level features. These features are then combined to form the more discriminative speaker embeddings so that the system achieves better performance. Results of experiments conducted on the Voxceleb1 dataset show that the proposed system outperforms the baselines. Moreover, the improvement is consistent with different speech duration.