Abstract: Different types of sounds have different effects on the quality of physical and mental health of urban residents. Accurate classification of urban sounds is conducive to effective evaluation of them, thus promoting the management of urban sounds. Deep learning has been applied in speech recognition, among which the recurrent neural network (RNN) is the most prominent. Due to the obvious gradient disappearance, large network loss and low accuracy of the basic RNN, the improved recurrent neural network was employed to classify the urban background noise. The long short-term memory neural network (LSTM) and the gated recurrent unit (GRU) neural network were used to construct a deep-circulating neural network model. The accuracy of the constructed deep neural network was tested and analyzed by the public data set UrbanSound8K. The model was based on the benchmark of the Mel frequency cepstral coefficient and the results were significantly improved compared with the basic RNN.