ERANNs: Efficient Residual Audio Neural Networks for Audio Pattern Recognition

We present a new architecture of convolutional neural networks (CNNs) based on ResNet for audio pattern recognition tasks. The main modification is introducing a new hyper-parameter for decreasing temporal sizes of tensors with increased stride sizes which we call "the decreasing temporal size parameter". Optimal values of this parameter decrease the number of multi-adds that make the system faster. This approach not only decreases computational complexity but it can save and even increase (for the AudioSet dataset) the performance for audio pattern recognition tasks. This observation can be confirmed by experiments on three datasets: the AudioSet dataset, the ESC-50 dataset, and RAVDESS. Our best system achieves the state-of-the-art performance on the AudioSet dataset with mAP of 0.450. We also transfer a model pre-trained on the AudioSet dataset to the ESC-50 dataset and RAVDESS and obtain the state-of-the-art results with accuracies of 0.961 and 0.748, respectively. We call our system "ERANN" (Efficient Residual Audio Neural Network).