Sound Event Detection Method for Home Scenarios Based on Self-Knowledge Distillation

This study addresses the task of Sound Event Detection (SED) in home environments, aiming to improve detection performance while achieving model lightweighting to meet the deployment needs of resource-constrained devices. A lightweight acoustic modeling framework named SKD-CRNN based on Self-Knowledge Distillation (SKD) is proposed. The method constructs a dual-branch Convolutional Recurrent Neural Network (CRNN) and introduces a self-distillation loss to guide the student branch to learn from the teacher branch, enabling knowledge transfer without an external teacher model. Experiments on the DCASE2023 Task4 dataset demonstrate that SKD-CRNN reduces the number of parameters to 80.7% of the baseline system while achieving consistent performance improvements in sound event detection across eight household sound event categories. Notably, the detection of non-instantaneous events such as electric shaver/toothbrush and dishes shows significant enhancement in onset and offset identification. Event-based, intersection-based, and segment-based F1 scores increase by 7.8%, 7.1%, and 6.2%, respectively, with PSDS1 and PSDS2 metrics improving by 2.5% and 3.2%. Experimental results demonstrate that SKD-CRNN improves SED detection capability while effectively reducing model complexity, making it suitable for edge deployment in home environments with limited computational resources.