A Similarity-Guided Aggregation Network for Sound Event Localization and Detection with Source Distance Estimation

Conventional sound event localization and detection (SELD) focuses on sound event detection (SED) and direction of arrival (DOA) estimation, which provide limited spatial information. The 3D SELD task addresses this limitation by incorporating source distance estimation (SDE). In this paper, we propose a similarity-guided aggregation network (SGA-Net) for sound event localization and detection with source distance estimation (3D SELD) that combines the spectrum-based branch and pre-trained model-based branch. We design a multi-scale feature extraction (MSFE) module for the spectrum-based branch to capture rich acoustic information and a similarity-guided feature fusion (SGFF) module that includes a spatial feature transform (SFT) module for effective fusion of intermediate features from both branches. The experimental results demonstrate that our proposed SGA-Net achieves the best performance in all metrics on the STARSS23 dataset, verifying the effectiveness of SGA-Net. The main code will be available at https://github.com/QingJWan/SGANet.