A new method was developed to identify γ-ray spectra with neural network based on wavelet packet decomposition. γ-ray spectra are regarded as non-stationary discrete signals. By using wavelet packet decomposition of γ-ray spectra, characteristic vector of the energy over all frequency band is obtained and used as an input variable into neural network to identify γ-ray spectra results of applying the method to γ-ray sources shows that the method can identify accurately γ-ray spectra from standard sources of different species and batches.