Attention based convolutional recurrent neural network for environmental sound classification

Environmental sound classification (ESC) is a challenging problem due to the complexity of sounds. The performance heavily dependent on effectiveness representative features extracted from environmental However, ESC often suffers semantically irrelevant frames and silent frames. In order deal with this, we employ frame-level attention model focus relevant salient Specifically, first propose convolutional recurrent neural network learn spectro-temporal temporal correlations. Then, extend our RNN mechanism discriminative feature representations for ESC. We investigated when using different scaling function applying layers. Experiments were conducted ESC-50 ESC-10 datasets. Experimental results demonstrated proposed method achieved state-of-the-art or competitive accuracy lower computational complexity. also visualized observed that was able lead tofocus parts