Lightweight Voiceprint Recognition with Ghost Convolution and Hierarchical Residual Splitting

In the context of rapid development in digitization and the Internet of Things, voiceprint recognition has emerged as a secure, convenient, and cost-effective method for identity verification. Given the resource constraints of edge devices, designing a lightweight voiceprint recognition model is of great theoretical and practical significance. In this paper, following an initial evaluation of several mainstream models, the Efficient Channel Attention, Propagation and Aggregation Temporal Difference Neural Network (ECAPA-TDNN) is chosen as the baseline model for further enhancement. At the outset, both the SE-Res2Block and the Fully Connected Layer of the ECAPA-TDNN are enhanced utilizing the ghost convolution approach, with the objective of reducing computational cost and improving recognition performance. Building upon the innovation, the ECAPA-TDNN is further enhanced by integrating hierarchical residual splitting. Validation conducted on the CN-Celeb1 and AIShell-1 datasets demonstrates that the algorithm achieves significant reductions in parameter size and the number of floating-point operations, with only a minor reduction in accuracy. Consequently, the optimized algorithm is particularly suitable for deployment on resource-constrained edge devices.