Voiceprint Recognition under Cross-Scenario Conditions Using Perceptual Wavelet Packet Entropy-Guided Efficient-Channel-Attention–Res2Net–Time-Delay-Neural-Network Model

Abstract

(1) Background: Voiceprint recognition technology uses individual vocal characteristics for identity authentication and faces many challenges in cross-scenario applications. The sound environment, device characteristics, and recording conditions in different scenarios cause changes in sound features, which, in turn, affect the accuracy of voiceprint recognition. (2) Methods: Based on the latest trends in deep learning, this paper uses the perceptual wavelet packet entropy (PWPE) method to extract the basic voiceprint features of the speaker before using the efficient channel attention (ECA) block and the Res2Net block to extract deep features. The PWPE block removes the effect of environmental noise on voiceprint features, so the perceptual wavelet packet entropy-guided ECA–Res2Net–Time-Delay-Neural-Network (PWPE-ECA-Res2Net-TDNN) model shows an excellent robustness. The ECA-Res2Net-TDNN block uses temporal statistical pooling with a multi-head attention mechanism to weight frame-level audio features, resulting in a weighted average of the final representation of the speech-level feature vectors. The sub-center ArcFace loss function is used to enhance intra-class compactness and inter-class differences, avoiding classification via output value alone like the softmax loss function. Based on the aforementioned elements, the PWPE-ECA-Res2Net-TDNN model for speaker recognition is designed to extract speaker feature embeddings more efficiently in cross-scenario applications. (3) Conclusions: The experimental results demonstrate that, compared to the ECAPA-TDNN model using MFCC features, the PWPE-based ECAPA-TDNN model performs better in terms of cross-scene recognition accuracy, exhibiting a stronger robustness and better noise resistance. Furthermore, the model maintains a relatively short recognition time even under the highest recognition rate conditions. Finally, a set of ablation experiments targeting each module of the proposed model is conducted. The results indicate that each module contributes to an improvement in the recognition performance.