PURPOSE: To evade the problem of partial matching by only one-stage collation even when a talk to be recognized object is changed by letting the cumulation result of similarity degree, obtained for a sequence of feature vectors of a standard pattern, be similarity degree with an input speech.

CONSTITUTION: This system has a speech input part 1 which inputs a speech, a feature extraction part 2 which converts the speech inputted from the speech input part 1 into an input pattern being a time series of feature vectors, and a collation part 3 which calculates the distance between the feature vectors of the input pattern extracted by the feature extraction part 2 and the feature vectors of the standard pattern. Then the difference between a 1st value which is previously given and the distance calculated by the collation part 3 is regarded as the similarity degree between the input pattern and standard pattern and the cumulation result of calculated similarity degree for the sequence of feature vectors of the standard pattern for the word to be recognized object is defined as the similarity degree of the word to be recognized object. Therefore, this system operates even on a recognition device which changes talks like a specific speaker type speech recognition device.