PURPOSE: To improve the recognition performance in noisy environment even when a large vocabulary are registered by using all recognized object vocabulary which are previously recognized as objects of recognition in less noisy environment, and excluding some of the recognized vocabulary which are previously registered from objects of recognition and performing recognition in the noisy environment.

CONSTITUTION: This device and method are equipped with a speech input part 101, an A/D conversion part 102 which quantizes an input speech obtained by the speech input part 101, an analysis part 103 which obtains feature components of the input speech, a standard pattern storage part 104 for feature vectors of the previously registered recognized object vocabulary a collation part 105 which recognizes the input speech by obtaining the similarity degree between the feature vectors and the feature vectors obtained by the analysis part 103, a noise state detection part 106, and a recognized object vocabulary limiting part 107 which makes the number of the recognized object vocabulary to be collated by the collation part 105 less than the number of the previously registered recognized object vocabulary. The recognized vocabulary limiting part 107 limits the recognized object vocabulary corresponding to the noise state detected by the noise detection part 106.