A novel acoustic scene classification (ASC) system based on asymmetric convolution and knowledge transfer is proposed to address the problem caused by limited ASC datasets. Compared with the existing methods which trained models from scratch, the proposed system fine-tunes a pretrained model of other tasks to preserve valid information from the source domain. Besides, targeted at the nature of acoustic features, it adopts asymmetric convolutions to enhance the network capability of feature extraction. Experiments shows that the proposed system outperforms the baseline system by 0. 023. Besides, as shown in the visualization results of convolutional filters, textures of the proposed system are more detailed than other methods. The experiment proves that knowledge transfer can boost model ability of feature representation, and it can further improve system performance by combining with asymmetric convolution.