Attentional multi-feature fusion for spoofing-aware speaker verification

The Spoofing-Aware Speaker Verification (SASV) system is designed to protect automatic speaker verification (ASV) systems from potential speech spoofing attacks by integrating the ASV and countermeasure systems. The optimization of the ASV system can further enhance the resistance of the SASV system to various spoofing methods. Thus, an Attentional Multi-Feature Fusion framework is proposed in this paper to enhance the speech feature content in the ASV system, aiming to mitigate security vulnerabilities. Furthermore, for feature modeling, we introduce the Conformer module, which combines convolutional neural networks and Transformers to effectively capture both local and global features while extracting fixed-dimensional speaker embedding vectors. The experimental results demonstrate that the proposed ASV system achieves a nearly 25% improvement compared to the baseline ASV system. Furthermore, the SASV system, based on the proposed ASV system, demonstrates nearly 33% and 34% improvements in spoofing awareness with score-based fusion and embedding-based fusion strategies, respectively. This implies that the improved SASV system is capable of detecting and recognizing speech-spoofing attacks more effectively.