Vehicle classification based on acoustic signals in urban environment provides valuable perception information for smart city management. In order to improve the accuracy of current vehicle sound classification, we propose a Sound-Convolutional Recurrent Neural Networks (S-CRNN) method. It combines convolutional neural networks (CNN) and recurrent neural network (RNN). By comparing the weighted F1 score (F1) and error rate (ER), it is proved that the proposed S-CRNN method has better classification performance than the original Sound-Convolutional Neural Networks (S-CNN) method, especially in the vehicles level, the weighted F1 value increases to 28.5%. Long short-term memory (LSTM) and Gate Recurrent Unit (GRU) are both used as RNN for comparison. And the S-CRNN model with GRU reduces the training time by 2.65 hours, maintaining the main performance in the meantime.