Abstract: A signal pattern recognition system includes a reference pattern memory storing plural reference data patterns against which input patterns are compared for recognition. The reference data patterns are formed by merging training patterns together. Each training pattern, to be accepted for merging, must match the previously merged patterns by a threshold amount. The threshold is automatically varied as the number of previously merged training patterns increases. If a predetermined number of successive training patterns is below the threshold, the training process is repeated, from the beginning, for the reference pattern which generates the errors. The system automatically trains each reference pattern with the same number of training patterns to assure uniformity when input data patterns are compared for recognition.