A SURVEY OF ATTENTION-BASED ARCHITECTURES FOR SYNTHETIC SPEECH DETECTION

Аннотация

This paper presents an analytical review of contemporary architectures for synthetic speech detection, emphasizing the central role of attention mechanisms. Three representative models are examined: a modified VGGish enhanced with the CBAM, YAMNet augmented with BAM, and the cascaded capsule network, ABCCapsNet. The analysis is based on mel spectrograms interpreted as two-dimensional images, where attention modules identify the most informative time–frequency regions. The discussion covers architectural design choices, feature extraction strategies, attention mechanisms, datasets, and performance metrics. Comparative evaluation reveals that YAMNet with BAM achieves the lowest EER with moderate computational complexity, while ABC-CapsNet exhibits strong robustness to data variability and generalizes well across heterogeneous datasets. In contrast, VGGish with CBAM provides a balanced trade-off between accuracy and implementation simplicity. The study underscores the pivotal role of attention mechanisms in enhancing the effectiveness of synthetic speech detection systems and outlines promising directions for further research and development