PURPOSE: To enable high-accuracy identification by finding the cross correlation function sequence between the input sequence of a time discrete voice signal and a signal sequence which is fixed or determined adaptively according to the signal and matching the found cross correlation function sequence with a previously found cross correlation function.

CONSTITUTION: When a speaker is identified by using voice data including repetitive voicing from the same speaker, the cross correlation function sequence of the speaker to be recognized which is found in advance is stored in addresses 1 - N of a memory 30 and a matcher 40 matches the cross correlation function sequence inputted from a cross correlator 20 with the cross correlation function sequence stored in the memory 30. Then the geographic distance between both function sequences is found as to values at corresponding discrete points of time and the sums of squares of the distances at the respective points are integrated from the start end to the final end and sent to a decision device 50, and the minimum error distance is selected and compared with a predetermined recognition rejection threshold value to make a final decision. When there are plural speakers, a start end and fixed terminal end open type nonlinear expansion/contraction pattern matching method is used.