Environmental Sound Classification Based on Knowledge Distillation

With the construction of smart cities, the research on Environmental Sound Classification (ESC) has been further developed, and good results have been achieved in the existing large network models, but due to its large model, it is not conducive to deployment on small and embedded devices. To this end, we apply knowledge distillation to the Environmental Sound Classification (ESC) task, transferring the knowledge learned from audio data through a large network model into a lightweight network model to improve lightweight network training. On this basis, we improved the knowledge distillation method, and the lightweight network model can obtain more information from different layers of the large network model. We found that our model outperformed existing models, achieving 87% accuracy on ESC-50.