Enhancing Vietnamese speech anti-spoofing with Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks (AASIST) and data augmentation

Abstract

The rapid advancement of speech synthesis and spoofing technologies has increasingly challenged the security of Automatic Speaker Verification (ASV) systems. However, most existing studies focus on English, while research on Vietnamese remains limited. In this work, we apply the Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks (AASIST) model to the Vietnamese dataset for Spoofing-Aware Speaker Verification (VSASV) dataset. To improve robustness, we introduce several data augmentation techniques, including additive noise, MP3 compression, loudness normalization, and Rawboost. We further fine-tune the model by freezing lower layers and adapting higher layers to Vietnamese speech characteristics. Experimental results indicate that our retrained AASIST model improves in detecting Vietnamese audio forgeries, achieving EERs of 5.54% on the development set and 5.28% on the test set. These results demonstrate that combining AASIST with data augmentation provides an effective solution for Vietnamese speech anti-spoofing.