Abstract
The diversity of spoofing attacks poses challenges to the generalization of spoofed speech detection models across datasets and attack conditions. However, existing single-model countermeasures may capture complementary spoofing cues inadequately, limiting their robustness under heterogeneous evaluation conditions. To address this issue, we propose a feature-level multi-model fusion network that integrates AASIST, RawNet2, and RawGAT-ST. Specifically, branch-specific projections align the frozen backbone embeddings, and the original backbone logits are excluded from fusion. A gated Transformer models the interactions among the three model tokens, while a mixture-of-experts (MoE) module with learned top-2 aggregation refines the aggregated representation through adaptive expert routing. Training combines label-smoothed classification, contrastive regularization, Mixup augmentation, and routing regularization. The fusion network is trained on the ASVspoof 2019 LA training set, with model selection performed on its development set, and the selected model is evaluated without target-dataset adaptation on ASVspoof 2019 LA, ASVspoof 2019 PA, ASVspoof 2021 LA, and ASVspoof 5; the proposed method achieves equal error rates (EERs) of 0.421%, 19.927%, 2.934%, and 7.245%, respectively. These results demonstrate its effectiveness on the ASVspoof 2019 LA benchmark and its transfer performance across evaluation datasets, while revealing substantial challenges under the more difficult PA condition.