Dual-Path Compensation Transformer Network with Teacher-Free Feature Distillation for Sound Event Localization and Detection

Sound Event Localization and Detection (SELD) aims to simultaneously detect the types and temporal boundaries of sound events while estimating their corresponding directions-of-arrival through the analysis of multi-channel audio signals. In this paper, we propose the Dual-Path Compensation Transformer Network (DPCT-Net), which leverages the strong time-frequency modeling capabilities of the Dual-Path Transformer to construct the DPCT-Net. DPCT-Net consists of two Information Compensation Transformers (ICT). Unlike conventional Transformers, ICT incorporates multi-head cross-attention into its attention modules, establishing an information compensation mechanism that enriches high-level semantic features with high-resolution spatial cues, thereby enhancing the feature representation capability of the network. Additionally, we introduce a Teacher-Free Feature Distillation (TFFD) training strategy, which utilizes deep-layer features to supervise shallow-layer feature, further improving the accuracy of sound event boundaries detection. We also propose a threshold-based activation coefficient to enable the network to fully benefit from the TFFD training strategy. Experimental evaluations on the STARSS22 and STARSS23 real-world datasets demonstrate that DPCT-Net achieves overall performance improvements of 21% and 15% in the SELD score, respectively, compared to baseline models. These results validate the effectiveness of DPCT-Net in complex acoustic environments.