This paper proposes a method for selecting the feature parameters in the statistical discriminant analysis of the human letter‐cognition related evoked potential. The paper aims at the quantitative interpretation of the frequency components, latency and the cortex area of the evoked potentials accompanied by the human letter‐cognition process. Models are constructed at three levels considering the awareness of the subjects: (1) classification, (2) retrieval and matching, and (3) attribute recall. To examine the validity of the models, the number of parameters is increased along the time axis to determine which addition of parameters resulted in a remarkable contribution to the discrimination. By examining the recognition rate, the necessary frequency components are determined. From the viewpoint of improving the recognition rate, methods of selecting the parameters are compared to examine the nature of the parameters selected in the method with the highest recognition rate. As a result, it is seen that the latencies in the models at three levels for the letter‐cognition process are approximately 190, 250 to 350, and 730 ms, respectively; and the necessary frequency components extend to at least 60 Hz. By comparing the selection method for the parameters, the recognition rate reached above 95 percent at the highest. It is seen that the tendency of the selected parameters is the same as those of the models.