The recognition system of spoken words using the restricted number of learning samples is described. The learning samples are composed of a part of the whole words to be recognized and are used to derive the reference patterns of phonemic spectrums needed in the recognition system. Four kinds of algorithms for selecting the optimum set of words constituting the learning samples are proposed and tested. Recognition test of twenty words is done for 20 speakers. By the use of a half of the twenty words as learning samples, the recognition rate of 98.6 % is obtained in the case of the optimum learning sample.