Dual Path Residual Network for Speaker Verification
Yibin Zhan, Rui Wang, Zhihua Wei, Yitong Pang · 2021
Current speaker verification technology relies on the use of neural networks to extract the representation of the speaker. In this work, we present a novel dual path network, SE-Shrinkage-ResNet. The basic module of the network, SE-Shrinkage, is a parallel connection of the SE-ResNet block and the DRSN block, which jointly contributes to the generation of the speaker-discriminative embedding. SE-ResNet blocks explicitly model channel interdependencies, while DRSN blocks employ channel-shared and channel-wise thresholds to mitigate the effect of high noise signals. Experimental results on the VoxCelebl dataset show that the proposed SE-Shrinkage-ResNet achieves superior performance compared to SE-ResNet and DRSN, and obtains a lower equal error rate compared to the latest ECAPA-TDNN with fewer parameters.