A multiscale self-adaptive attentional feature fusion network for sound event localization and detection

Sound event localization and detection (SELD) in real spatial scenes is an emerging and challenging task. Existing deep learning-based SELD methods usually capture feature information through a single receptive field, resulting in insufficient effective features and lacks global contextual information. To address these issues, we propose a multi-scale self-adaptive attentional feature fusion network (MAFFNet). Firstly, the method fuses logarithmic Mel-spectrogram and intensity vector features through a residual module, then uses multiple convolution kernels to obtain features at different scales and introduces an attentional feature fusion method to adaptively integrate the obtained multi-scale features. Finally, a two-layer bidirectional gated recurrent unit (Bi-GRU) module is employed to construct the temporal dependency of features. Extensive experimental results conducted on the Sony-TAu Realistic Spatial Soundscapes 2022 (STARSS22) demonstrate that the MAFFNet exhibits better robustness across 13 target sound events, with an overall performance improvement of approximately 8%, while having only 1.12 million parameters.