<p>A Transformer-based convolution neural network with potent self-attention and bidirectional cross-attention is proposed to improve the classification accuracy of acoustic scene classification (ASC). In this method, the audio signals from different scenes are obtained by the recording devices, and Mel Frequency Cepstral Coefficients features are first extracted as the inputs of the network. Then, the extracted feature maps are input into CNN and Transformer branches through 3 × 3 convolution layers and token setting layers, and the cross-attention blocks are used to link these branches. Finally, the output feature vectors of both branches are merged and passed through a multi-layer perceptron layer to obtain the precise classification of multiple acoustic scenes. In the proposed method, CNN branch that predominantly captures the local features, and Transformer branch takes a broader perspective by analyzing the global information. The cross-attention block and Transformer branch improve the network’s representation capability. Meanwhile, the potent self-attention block can encode positional information using different sizes of convolution kernels and leverages the CNN’s capacity to aggregate local information, which enhances the representation learning for precisely classifying multiple acoustic scenes. The proposed method achieves the classification accuracies of 77.6% and 85.4% on two public datasets, respectively, surpassing those of several previous deep learning-based ASC methods. The experimental results prove that the proposed method accurately classifies multiple acoustic scenes, and its classification accuracy is better than those of several previous deep learning-based ASC methods. Meanwhile, several ablation experiments including Transformer branch, token number, token dimension and attention strategy are implemented to further demonstrate that the stability and robustness of the proposed method are stronger.</p>