Machine learning is one of the leading directions in digital signal processing. For example, in neutron spectrometry, artificial neural networks are actively used to suppress gamma background when analyzing signals from scintillation detectors. This article describes a method for determining the quality of n/γ-separation by an artificial neural network. The efficiency of the method is demonstrated by analyzing the signals obtained by measuring the prompt neutron spectrum of 252Cf spontaneous fission using a scintillation detector based on a stilbene crystal. The essence of the method is to determine the proportion of falsely identified events for each of the analyzed signal classes using a known reference method. An exemplary gamma-ray source was used to determine the false count of recoil protons. This approach made it possible to estimate the fraction of events from electrons identified as recoil protons and the fraction of recoil protons perceived as electrons, depending on the light yield of the scintillation signal. This, in turn, made it possible to reconstruct the true energy spectra for different types of particles, including for the region of low signal amplitudes, where the separation quality is usually poor. The reconstructing error was less than 8 % for the light yield region of less than 120 keVee.