Sound Event Localization and Detection via Multi-Head Attention Temporal Convolutional Network

In this paper, we propose a new Multi-Head Attention Temporal Convolutional Network (MHA-TCN). By applying multi-head attention to these features, we capture the feature relationships under short-term and long-term dependencies. A channel attention mechanism is introduced to further enhance the system's robustness. The proposed framework leverages the advantages of attention mechanisms and TCN in temporal modeling. The Multi-Head Attention mechanism successfully integrates and recombines long-term and short-term features in the early stage of model training, and the channel attention mechanism adaptively adjusts the weights of different channels, reducing feature redundancy and improving system robustness and generalization. Experimental evaluations on multiple datasets demonstrate that the proposed method has achieved improvement and the convergence speed of the network was increased.