A new AE source analysis using a neural network was investigated. A two-layer network and the artificial AE source signals caused by pencil lead break were used. Detected waveform data obtained by capacitive displadement transducer were processed by the personal computer system. For calculation of the unknown source waveform, the previously trained memory matrix of the network corresponded to the input/output relation. In association with the number of learning data, a small number of data should have estimated incorrect waveforms. But, when additional restrictions were imposed on the memory matrix, appropriate waveforms could be obtained irrespective of the number of learning data, which could be as small as one. The analysis also had such advantages as that relatively low noise data might be calculated if the detected data should include a lot of noise.