Abnormal Sound Event Detection Method Based on Time-Spectrum Information Fusion

Abstract

In this paper, we propose an abnormal sound event detection method based on Time-Frequency Spectral Information Fusion Neural Network (TFSIFNN), addressing the problem that the time structure and frequency information of sound events in real environment are widely varied, resulting in poor performance of abnormal sound event detection. First, we construct a TCN-BiLSTM network based on Temporal Convolutional Networks (TCN) and Bidirectional Long Short-Term Memory (BiLSTM) networks to extract the temporal context information from sound events. Next, we enhance the feature learning capability of the MobileNetV3 network through Efficient Channel Attention (ECA), culminating in the design of an ECA-MobileNetV3 network to capture the spectral information within sound events. Finally, a TFSIFNN model was established based on TCN-BiLSTM and ECA-MobileNetV3 to improve the performance of abnormal sound event detection. The experimental results, conducted on the Urbansound8K and TUT Rare Sound Events 2017 datasets, demonstrate that our TFSIFNN model achieved notable performance improvements. Specifically, it reached an accuracy of 93.93% and an F1-Score of 94.15% on the Urbansound8K dataset. On the TUT Rare Sound Events 2017 dataset, compared to the baseline method, the error rate on the evaluation set decreased by 0.55, and the F1-Score improved by 29.69%.