Vocal Detection Using Convolution Neural Networks with Visualization Tools

In this paper, we report the use of convolutional neural networks (CNN) with the aid of visualization tools, such as Grad-CAM and Score-CAM, to detect vocal signals in audio clips. CNN model typically requires relatively long audio segments for prediction. To improve the detection accuracy with short nonvocal gaps, we include information obtained from the Grad-CAM and Score-CAM, along with the original predicted results, to form the input to a post classifier. The experimental results show that the detection accuracy is improved, especially when short nonvocal gaps are present.