Graph Neural Network of Multiple Diagonal Constraints for Sound Event Detection
Kibae Lee, Jongkwon Choi, Keunhwa Lee, C A Lee · IEEE Access · 2025
Acoustic signals originating from distinct sources inherently exhibit diverse spectral and temporal characteristics. In this study, we propose a novel framework, termed the graph neural network of multiple diagonal constraints (GMDC), which is designed to learn class-specific representations of sound events within the graph domain. The proposed method initially encodes an input audio signal as a graph structure, which is subsequently extended into multiple class-specific graphs to enable the detection of overlapping sound events. To construct these class-specific graphs, the model learns a set of multi-graph structures represented by an adjacency tensor, which is derived based on temporal annotations for each sound class. The learning process is guided by two distinct loss functions: the first enforces diagonal constraints to emphasize features in event regions, while the second further refines this by suppressing off-diagonal activations in non-event regions. Furthermore, the GMDC framework integrates graph learning with event detection through both sequential and joint learning strategies, enabling more effective modeling of complex acoustic scenes. Experimental results on two publicly available datasets demonstrate that the proposed GMDC framework improves the F1-score by 8.0% to 12.1% and reduces the error rate (ER) by 0.02 to 0.04 compared to the baseline convolutional recurrent neural network (CRNN). Furthermore, on the synthetic underwater dataset, the GMDC consistently outperforms the CRNN baseline across various signal-to-noise ratio (SNR) and signal-to interference ratio (SIR) conditions. Even under severely degraded conditions such as -12 dB SNR and SIR, the GMDC achieves improvements of 12.6% and 10.3% in F1-score and reductions of 0.16 and 0.21 in ER, respectively.