Time-Decoupled Spectral Modeling: Wide-Band Recalibration and Adaptive Fusion for Sound Event Detection

Convolutional recurrent neural networks (CRNNs) provide a compact solution for sound event detection (SED), but distant frequency bands interact only through successive local operations in their finite-depth convolutional front-ends. We propose Time-Decoupled Spectral Gating (TDSG), a parameter-efficient module that supplements CNN-based SED with wide-band spectral context. TDSG estimates a segment-conditioned spectral gate from a temporally aggregated channel–frequency descriptor while retaining the original frame-level feature for event localization. Element-wise adaptive fusion controls the strength of recalibration at each channel–time–frequency position. Controlled experiments on DESED and Urban-SED with CRNN and FDC-CRNN front-ends show higher mean detection scores over five runs per setting. On DESED, CRNN+TDSG improves the two Polyphonic Sound Detection Score scenarios (PSDS1/PSDS2) from 0.400/0.598 to 0.441/0.668 with 12.7% more parameters and 29.0% more FLOPs. On Urban-SED, it improves PSDS1/PSDS2 from 0.161/0.244 to 0.276/0.320. CRNN+TDSG also achieves higher mean PSDS than FDC-CRNN on both datasets with 54.9% fewer parameters and 60.4% fewer FLOPs. Ablation and timing analyses characterize the accuracy–complexity trade-off.