To perform a speaker collation with a high matching rate even to short time speech signal data by a speech collating device for collating a person himself/ herself by speech.

Plural two-dimensional templates arranged in a registered spectrogram and each area on a collating speech spectrogram from which a maximum correlation coefficient is calculated are detected. Moreover, respective summed frequency spectrum distributions SumA and SumB are derived from the registered speech spectrogram and the matching speech spectrogram, and plural one-dimensional templates arranged on the SumA and each area on the SumB from which the maximum correlation coefficient is calculated are detected. Then, a difference between a mutual positional relation among the two-dimensional templates and a mutual positional relation among the maximum correlation detection area is comparatively evaluated, and also a difference between the mutual positional relation among the one-dimensional templates and the mutual positional relation among the maximum correlation detection area is compared and evaluated, the identity of the registered speech with the collated speech is discriminated according to the overall difference degree.