PURPOSE: To prevent recognition due to an omission of a fall of frequency analytic data of a consonant, etc., of an input speech from its resampled data by checking whether or not frequency analytic data of a specific frequency band in a specific speech part section of the input speech has a minimum value.

CONSTITUTION: A frequency analysis of the input speech is taken, and the frequency analytic data in a speech section 2 is resampled 4 to generate an input speech pattern, which is matched against a previously generated dictionary patterns 6 to recognize the input speech. At this time, the frequency analytic data 3 in the specific frequency band in the specific partial section of the input speech is searched for according to the recognition result of the input speech when necessary. Then it is checked whether or not the frequency analytic data has a minimum value and according to the check result, the final recognition result of the input speech is found. Consequently, even if the input speech has features lost owing to the resampling, it can correctly be recognized and a speech having a long speech section need not be replaced with another speech having a short speech section for speech input.