Improving ECAPA-TDNN Performance with Coordinate Attention

The current mainstream networks, such as squeeze and excitation residual neural network (SE-ResNet) and emphasized channel attention, propagation and aggregation based time delay neural network (ECAPA-TDNN), enhance the capability of speaker embedding extractors to extract more discriminative speaker embeddings by incorporating squeeze and excitation (SE) attention within the convolutional blocks. However, the SE attention focuses solely on encoding inter-channel information, overlooking the importance of spatial positional information and time-frequency information, which are crucial for the model’s performance. In this paper, we first experimentally compare the effectiveness of several mainstream attention mechanisms in the computer vision domain for the ECAPA-TDNN model. Next, we focus on the substantial improvements that coordinate attention (CA) brings to the ECAPA-TDNN model. The introduction of CA can help the model embed time-frequency information into the channel representation. Even without using AS-Norm, our proposed model achieves relative reductions of about 5.3% equal error rate (EER) and 5.5% minimum detection cost function (minDCF) on both the Voxceleb-O and Voxceleb-H test sets compared to the ECAPA-TDNN baseline model. In addition, the EER is relatively reduced by 9.46% on the CN-Celeb1 test set. This result strongly demonstrates that the CA module can effectively improve the generalization ability of the ECAPA-TDNN model.