PURPOSE: To save power consumption and to secure a high speech recognition performance.

CONSTITUTION: Employing a reverse error propagation method, the stored contents of a network connecting storage section 38 are updated by a forward arithmetic section 36 and a learning process section 28 in a learning mode. In a recognition mode, the learning process section 28 is separated by connecter sections 48a to 48f to perform the recognition operation, Furthermore, the information stored in the prepared storage medium is written in the network connecting storage section 38 to perform the recognition operation. Moreover, the least square approximate linear components are eliminated from each frequency parameter of the frequency parameter group, the value which is equivalent to the power of the measured speech is added and thus, the normalized frequency parameter group is generated.