ECAPA-TDNN WITH DUAL-BRANCH ATTENTION MECHANISM FOR SYNTHETIC SPEECH DETECTION

This study presents a modified ECAPA-TDNN model with a dual-branch attention mechanism for the task of synthetic speech detection aimed at countering audio spoofing. The architecture integrates two parallel spectral feature branches (LFCC and LogMel) with multi-level attention through Attentive Statistics Pooling followed by a fusion encoder. The proposed approach demonstrates a significant improvement in the accuracy of synthetic speech detection, achieving an EER (LA) of 0.08%, which confirms the effectiveness of the dual-branch design and the fine-tuning of hyperparameters to enhance detection performance.