Abstract: Energy spectrum data analysis is the main source of nuclide identification. Aiming at the slow recognition speed and low accuracy of the emerging energy spectrum nuclide identification method in the noisy environment of mixed radionuclides, an energy spectrum nuclide recognition method based on long short-term memory neural network (LSTM) is proposed. In the experiment, a LaBr₃ crystal detector was used to measure the ⁶⁰Co and ¹³⁷Cs radioactive sources in the environment to obtain a gamma spectrum data set. First, the experiment used data smoothing and normalization methods for data preprocessing. Then, the energy spectrum data was grouped in time series to obtain a usable input sequence array. Finally, the prediction results were obtained through the LSTM model. By comparing two energy spectrum recognition models based on BP neural network and convolutional neural network (CNN), the average recognition rates in the test set are 83.45% and 86.21% respectively, while the average recognition rate of the LSTM model is 93.04%. The experimental results show that the energy spectrum model has performed well in the nuclide identification and can be used in fast energy spectrum nuclide identification equipment.