To correct a feature vector to be decided at a speech recognition time for correcting the effect of a noise by defining the mean value and the standard deviation of the characteristic vector, using these parameters and normalizing the characteristic vector.

The mean value and the standard deviation of the characteristic vector are defined, and the feature vector is normalized using these parameters. In such a case, the feature vector is normalized using a sliding normarization buffer preferably. In a speech recognition device, a front end 30 forms the feature vector Xi (i=1-M) at a prescribed interval. The feature vector X is stored in a normalization buffer 31, and thus, the mean value and the standard deviation related to respective feature vector components Xi are calculated. Thereafter, the feature vector component Xi to be recognized is normalized using a calculated normalization coefficient in the block 31. In such a case, the normalization buffer 31 that e.g. a length N is fixed is slid on the feature vector X.