Abstract: Local regional differences can make it difficult to classify environmental sounds accurately. Therefore, an environmental sound classification based on compact bilinear attention network is proposed. First, multi-dimensional time-frequency features are introduced to fully characterize the characteristics of environmental sound. Second, online random erasing data augmentation is introduced to avoid overfitting of the trained model due to lack of dataset and improve sample diversity. Finally, with the unchanged compact bilinear network framework, DensNet-169 is adopted as the feature extraction module, and the channel spatial location attention module is introduced to pay attention to the differences of local regions of environmental sound features. The experimental results show that the accuracy of the proposed method on ESC-10 and ESC-50 datasets can reach 96.0% and 87.9%, respectively, both of which are better than human ear recognition accuracy.

Key words: compactbilinear network, attention module, environmental sound classification, random erasing data augmentation, multi-dimensional time-frequency features