PURPOSE: To reduce the work load of a neural network, to easily construct a closed space of the registrant's pattern and to improve a recognition rate by using the neural network which has one output unit for every registrant.

CONSTITUTION: The speaker recognition method performs an evaluation after a learning work. A mode switching section 6 sets a learning mode and a speech input section inputs a sample learning speech. The inputted speech is divided into blocks and frequency characteristics for every specific time duration are obtained. After that, these data are inputted to a neural network 5 and the transformation functions and the connecting weights of the network 5 are varied so that the average value of the error between the output of the network and the target value which is set by a target setting section 9 becomes lower than a certain value. During the evaluation, an evaluation mode is set by the switching section 6 and the evaluation sample is inputted from the speech input section. Since the registrant who is the subject of each neural network is limited to one person, the work load of the neural network 5 is reduced compared with the conventional one and it is easier to construct the closed space of a registrant's pattern.