Acoustic Scene Recognition and Classification-Based Adaptive Convolutional Neural Network

Aiming at the limitations of traditional neural networks with fixed convolutional kernels and difficulty in dynamically responding to input audio changes in acoustic scene recognition and classification(ASRC), this paper proposes an innovative adaptive convolutional module (AConv). The AConv module adaptively adjusts CNN's convolutional kernel size based on the spectrogram feature distribution. It is used to replace the traditional convolutional layers in the network model, employing data augmentation techniques during the data processing stage to improve the generalization ability of the model. On the UrbanSoud8K and ESC-50 dataset, firstly, we verified the performance changes of the network model within the adjustable range of different convolutional kernel sizes; Subsequently, we further explored the replacement network while keeping the adjustable range of the convolutional kernel size, the impact of different numbers of convolutional layers on the performance of the network model; and then experiments on the impact of different input features on AConv. Experimental results show that the replacement of the AConv module significantly improves the performance of the network model. In the end, compared with other models, the structure proposed in this paper showed obvious advantages on different data sets, verifying its effectiveness and potential in ASRC tasks.