Semantic-Aware Conformer With Multi-Scale Fusion for Speaker Verification

This paper proposes an enhanced semantic-aware Conformer model for speaker verification. Building on the original multi-scale feature aggregation conformer (MFA-Conformer) architecture, the proposed model introduces multi-scale semantic fusion to achieve improved performance. While the baseline MFA-Conformer effectively combines global and local features through Conformer blocks and a multi-scale aggregation architecture, our model further emphasizes important features through semantic-aware fusion. Specifically, an enhanced semantic feature fusion is proposed, which adaptively fuses features from different Conformer blocks via an attention mechanism to form low-, mid-, and high-level semantic features. These three types of semantic features are then adaptively fused and refined to highlight important semantic information. Experimental results show that our model achieves an equal error rate (EER) of 0.63% and a minimum detection cost function (minDCF) of 0.082 on the VoxCeleb-O set with data augmentation, outperforming the baseline MFA-Conformer. These improvements highlight the effectiveness of structured semantic fusion and attention-guided feature refinement in automatic speaker verification. This work demonstrates the potential of semantic-aware modeling to enhance speaker verification system performance without significantly increasing computational complexity.