In order to improve the accuracy of abnormal sound recognition of highway tunnel accident and to solve the problem that convolutional neural networks only pay attention to local information. An integrated voice recognition model based on CNN-RNN is proposed. The model used the Stacking integration strategy to combine the strong feature expression ability of CNN and the strong memory ability of RNN.The gated cyclic memory unit was used to reduce the computational complexity of RNN. SIREN sinusoidal periodic function was used as the implicit activation function of RNN to enhance the fitting ability of the model to sound data. The precision of multi-channel convolution refinement feature extraction was designed to achieve global feature extraction. The performance of the proposed sound recognition model was evaluated on the abnormal sound data set. Experimental results show that the proposed sound model has higher recognition performance than other models and is more robust, which can effectively identify the abnormal sound of highway tunnel accidents.