Voiceprint Database Construction Method Based on Temporal Dynamic Multi-Scale Channel Attention Fusion

To address the problem of speaker verification accuracy being easily affected by external factors, a speaker verification algorithm based on a Temporal Dynamic Multi-Scale Channel Attention Fusion Time-Delay Neural Network (TDMCATDNN) model was proposed. The model had been improved in two aspects based on the Emphasized Channel Attention Propagation Aggregation Time Delay Neural Network (ECAPA-TDNN), including adding a multi-scale channel attention module to fuse multi-scale information by combining local and global features, and embedding a temporal dynamic convolutional neural network to consider temporal variation which adapts to time bins by applying a weighted sum of trained basis kernels. These improvements enabled the model to better utilize multi-scale temporal and spectral information, enhancing its recognition capability. Experimental results show that, compared to the ECAPA-TDNN model, the TDMCA-TDNN model achieves reductions of 3.9 % in equal error rate and 6.9 % in minimum detection cost function, respectively. The lowest equal error rate is reduced to 9.56 %, and the minimum detection cost function reaches 0.5352.