A deep learning framework for environmental sound classification by fusing linear and nonlinear features

Environmental Sound Classification (ESC) is a crucial research direction in audio signal processing, aiming to identify and classify specific events in ambient sounds. Traditional methods typically rely on time–frequency features (e.g., inverse Mel spectrogram) fail to fully capture complex temporal patterns in sound signals. To address this critical issue, this paper employs Recurrence Plot (RP) to compensate for the deficiency of nonlinear features in time–frequency features. To validate the fused features and the RP of performance in ESC, comparative experiments were conducted on four different models (e.g., CNN and GRU) and two benchmark datasets. The model’s performance based on deep learning is enhanced through integrating linear and nonlinear feature. Furthermore, a novel SResNet architecture is proposed, which embeds an attention mechanism into the feature fusion process of ResNet-18 and incorporates the SwiGLU activation function to optimize residual blocks. The smoothing property of SwiGLU contributes to stabilize residual networks and accelerate convergence, enabling the capture of more intricate patterns. Experimental results demonstrate that the proposed feature fusion outperforms traditional linear feature fusion methods on both ESC-50 and UrbanSound8K datasets, thereby validating the robustness of RP in ESC tasks. Concurrently, SResNet also exhibits superior performance compared to direct feature fusion. This innovative approach of parallel feature fusion and model optimization advances environmental sound analysis, enabling more comprehensive and accurate representation of ambient sound data.