Sound Event Detection Based on Bidirectional Temporal Convolutional Network and Gated Recurrent Unit

To address the problem of “gradient vanishing” and “gradient explosion” in sound event detection networks and the difficulty of capturing long-term dependencies, a bi-directional temporal convolutional network with a channel attention mechanism and a bi-directional gated recurrent unit (ABTCN-BGRU)is first introduced in this paper to improve the performance of sound event detection. Firstly, we design a bi-directional temporal convolutional network (CA-BTCN) model with attention mechanism to enhance feature selection in specific regions and achieve the rich features; then, the extracted feature sequences are fed to the Bi-GRU model for training to obtain long-time sequence information and improve the model feature representation; Finally, a fully connected layer is used for information integration and output the classification results. The detection capability of the ABTCN-BGRU model was verified based on the TUT2018 data set, we further compare the proposed method with other deep learning based models, it is verified that this model has stronger sound detection performance.