This paper proposes a convolutional neural network model for environmental sound classification based on channel and frame-level feature attention. The model selects one-dimensional convolution to further extract effective features from the two-dimensional logarithmic mel spectrum features of environmental sounds, and then uses the SE-Res2Net module to achieve a global perception of shallow input features at a fine-grained level and obtain channel attention scores. Attention statistical pooling is added before the full connection layer, and the attention mechanism is applied to the frame-level features of the input weighted channel features to obtain the importancescores ? of the frame-level features at different times for the channel. After weighted calculation, the average value μ and variance σ of the channel are connected in series as output.The experiment uses Urbansounda8k data set to train and test, and finally reaches 94.5 % accuracy, indicating that the channel-level and frame-level feature attention models have higher classification ability for sound. In order to further prove the performance of the model, the ablation experiment was carried out, and the analysis of the experimental results shows that the attention mechanism of applying channel and frame-level features to the sound features can reduce the classification error rate of the model by 43.8%.