The analysis using a neural network was carried out to determine the acoustic emission (AE) source waveform from the artificial AE signal detected by a piezoelectric AE sensor. We assume that each sampling point of the waveform corresponds to one unit of a layer in the network. It was found that the two-layer network assuming convolution form would be effective to determine source waveforms. Since linearity of the input/output relationship of the system would be limited, two-layer networks might not be available to deal with a general problem in AE source analysis. Hence, the network structure was extended to three layers, in which was used the back-propagation algorithm for learning. It was demonstrated that if appropriate waveform data were provided for learning, analogous source waveforms could be exactly determined. It was also revealed that the number of intermediate layer units might be of greater necessity than the number of learning patterns, but the number of units exerted little influence on the calculations.