Environmental Sound Classification Based on Multi-channel Features and Hybrid Attention

Environmental sound classification (ESC) is an important research direction, but it remains challenging due to the wide variety of environmental sounds which are difficult to uniformly represent and can be easily interfered by noise. The ESC has limitations in accurately separating and recognizing sounds in multi-class complex scenarios, especially under conditions of noise interference or overlapping sound sources. To improve the accuracy of ESC recognition, this paper proposes a classification method that utilizes a multi-channel feature and a mixed attention model. Firstly, the time-level and frequency-domain signals of the ESC are extracted, and filters are applied to extract spectral features and reconstruct into a feature graph. The limitations of single-feature representation are addressed by leveraging the complementarity between the three-channel features. Secondly, a hybrid classification model composed of channel and time-frequency concentration module classification model is introduced. The channel attention model calculates the feature map and assigns weights to different channels, and the channel features with more effective information and better resolution of this kind of sound can be allocated more weights. The time-frequency concentration module, however, focuses on more effective information in the time domain and frequency domain. The method proposed in this paper can effectively suppress the background noise, eliminate redundancy, and improve the convergence speed and classification accuracy. The recognition accuracies on ESC-10 and ESC-50 datasets reach 96.25% and 89.56% respectively, and 98.40% on the Urbansound8k dataset.