PROBLEM TO BE SOLVED: To provide a speaker likelihood evaluation device that robustly calculates similarity of a speaker feature vector for recognizing a speaker even when a registered utterance and a recognized utterance are different in length.;SOLUTION: A registered utterance division device comprises: an utterance division part 110 which generates divided registered utterance shorter than the length of registered utterance by cutting a part of the registered utterance out; a segment division part 310 which generates a registered utterance segment shorter than the length of the divided registered utterance from the registered utterance; a divided registered utterance generation part 320 which generates divided registered utterance by combining registered utterance segments together to the same length as the length of a section of the divided registered utterance; a divided registered utterance speaker feature vector calculation part 1001 which generates a speaker feature vector of the divided registered utterance; a recognized utterance speaker feature vector calculation part 1002 which generates a speaker feature vector of recognized utterance; and a vector similarity calculation part 850 which calculates a similarity between the speaker feature vector of the divided registered utterance and the speaker feature vector of the recognized utterance.;SELECTED DRAWING: Figure 7;COPYRIGHT: (C)2018,JPO&INPIT
展开▼