Due to the unique "fingerprint spectrum" characteristic, terahertz （THz） spectrum can be used to recognize the materials. With the development of artificial intelligence, deep learning is widely used in the field of THz spectrum recognition. However, the acquired THz spectral data are not always on a large scale due to the influence of experimental equipment, conditions and environment, which cannot meet the data size requirements of the deep learning algorithm. In order to solve this problem, we proposed a method of THz spectrum recognition based on generative adversarial networks （GAN） in this paper. Firstly, an S-G filter and a cubic spline interpolation method were employed to pre-process the THz spectral data. Secondly, the simulation data with the distribution of real THz spectral data were generated by the GAN. Finally, the generated data and real spectral data were taken as the training samples to train the deep neural networks （DNN）, thus obtaining the recognition results of the materials. The experimental results show that the THz spectral data generated by the GAN model can effectively simulate the overall characteristics of real THz spectral data and expand the THz spectral data samples, greatly elevating the spectral recognition accuracy.