JPH11249684A - Method and device for deciding threshold value in speaker collation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decide a threshold value beforehand and to obtain the suitable threshold value even when different registering words are uttered at every registered speaker. SOLUTION: When the threshold value in speaker collation becoming a reference deciding the said person, when an input pattern is collated with a standard pattern, is decided the standard pattern in sub-word unit of plural speakers excepting the registered speaker is stored beforehand in a storage part 20, and the sub-word standard patterns showing relevant utterance contents are connected at a registering time, and a word standard pattern is formed, and a distribution of a similar degree with the registered voice of the said person is obtained by a similar degree calculation part 30, and the threshold value is decided based on the formed distribution.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for determining a threshold value in speaker verification for determining a threshold for personal identification at the time of verification used in a speaker verification apparatus.

[0002]

2. Description of the Related Art In speaker verification, the degree of similarity between an input voice and a registered personal pattern is determined at the time of verification. Is done. Conventionally, a threshold value used as a criterion for determination is often set afterwards based on the results of evaluation experiments using a plurality of words of a plurality of speakers, and there is a problem in that the time required for setting the threshold value is large.

As a method for solving this problem, a method of setting a threshold value in advance has been proposed by Furui. (S.Furu
i, "Cepstral analysis technique for automatic spea
kerverification, "IEEE Trans. Acoust., Speech, Sig
nal processing, ASSP-29,2, pp.254-271,1981. This method examines the distribution of the similarity (distance) between the registered standard pattern of the subject and the voices of a plurality of speakers different from the subject (speech contents are the same as the subject), and the following (1) ) Is a method of setting a threshold according to the expression. Threshold value (k) = 0.6 * (μ (k) −σ (k) +7.0 (1)

Here, μ is the average value of the similarity between the principal standard pattern and the voices of a plurality of speakers different from the principal,
σ is the variance and k is the registered speaker number.

FIG. 5 is a diagram showing the configuration of an apparatus for executing this method. First, the principal standard pattern creating unit 31
At 0, a personal standard pattern (word standard pattern) is created using the registered voice, and stored in the tree-structured speaker standard pattern storage unit 320 for each word. Next, the similarity calculation section 330 performs similarity calculation between the created personal standard pattern and the voices of the multiple speakers stored in the voice storage section 340 of the multiple speakers. Subsequently, the average value and variance calculation unit 350 calculates the average value and the variance of the similarities from the calculated similarity distribution. Finally, the threshold calculator 360
In step (1), a threshold value is calculated according to equation (1) using the calculated average value and variance. When this method is used, the threshold can be determined in advance from the utterance at the time of registration, so that the work of setting the threshold can be reduced.

[0006]

Since the conventional method is registration in word units, it is necessary to prepare in advance the utterances of a number of speakers having the same contents as the registration utterances of the individual, and a plurality of registered speakers are different. It is difficult to prepare all utterances in advance when registering words. For this reason, the conventional method is an effective method when registering a common utterance content for all registered speakers.However, when a different registration word is uttered for each registered speaker, an appropriate threshold value can be obtained. Can not.

An object of the present invention is to provide a threshold value determining method and apparatus which can determine a threshold value in advance and can obtain an appropriate threshold value even when a different registration word is uttered for each registered speaker. Is to do.

[0008]

In order to solve the above-mentioned problems, a threshold value determining method and apparatus according to the present invention employ the following characteristic configuration.

(1) In a method of determining a threshold value in speaker verification, which is a criterion for personal identification when comparing an input pattern with a standard pattern, standard patterns in subword units of a plurality of speakers other than the registered speaker are stored and registered. A speaker characterized in that a word standard pattern is created by concatenating subword standard patterns representing the contents of the utterances at times, a distribution of similarity with the registered voice of the person is obtained, and a threshold value is determined based on the created distribution. A method of determining a threshold in matching.

(2) The threshold value determining method according to (1), wherein the sub-word is a phoneme, a syllable or a semi-syllable.

(3) The threshold value determination method according to (1), wherein the threshold value is determined as the similarity of the Kth percentile of the distribution of the similarity between the registered voice of the subject and the standard pattern of the multiple speakers.

(4) The threshold value is determined as an average value of N similarities in the vicinity of the K th percentile of the distribution of similarities between the registered voice of the subject and the standard patterns of a plurality of speakers. Decision method.

(5) The threshold value according to (1), wherein the threshold value is determined based on N similarities randomly selected from a distribution of similarities between a registered voice of the user and standard patterns of a plurality of speakers. Decision method.

(6) The threshold value is determined based on N similarities near the median of the similarity distribution between the registered voice of the person and the standard pattern of a plurality of speakers. Method.

(7) A standard pattern of a plurality of speakers is constituted by each node corresponding to a speaker cluster, and the first-stage node corresponds to a cluster of all speakers. The child node corresponds to the second stage node, the upper node has a standard pattern representing the acoustic characteristics of the lower node connected thereto, and the leaf node at the end is a tree structure story composed of the standard pattern of each speaker. The similarity between the registered voice and the standard pattern of each node in the second row is calculated as the standard pattern of the second node, and the standard value of the lower node of the node having the upper M-th similarity of the calculated similarity is stored. A similarity calculation between the pattern and the registered voice is calculated, and such calculation is performed from the upper node to the lower node, and the standard pattern of the lower node of the upper M node having a higher similarity at each stage and the registered voice are calculated. Degree of similarity Calculated degree of similarity is not calculated node substitutes the value of the upper node, the threshold determination method for determining a threshold value based on the final distribution of the calculated similarity similarity leaf node.

(8) The threshold value determining method according to (7), wherein the similarity of the Kth percentile of the similarity distribution is determined as a threshold.

(9) The threshold value determining method according to (7), wherein the similarity value of the average value of the N similarities near the Kth percentile of the similarity distribution is determined as the threshold value.

(10) The threshold value determining method according to (7), wherein the similarity value of the average value of the N similarities near the K th percentile of the similarity distribution is determined as the threshold value.

(11) The input registration voice is analyzed and converted into a feature vector, and the standard patterns of a plurality of speakers are chained according to the input utterance contents of the standard patterns in subword units. A threshold value determining method for calculating a similarity between the feature vector of the subject and a word standard pattern of a plurality of speakers, and obtaining a threshold at the time of matching from the calculated similarity distribution.

(12) The input registration voice is analyzed and converted into a feature vector, and the standard patterns of a plurality of speakers are chained in accordance with the input utterance contents of the standard patterns in subword units. Is created, the similarity between the feature vector of the subject person and the word standard pattern of multiple speakers is calculated, the calculated similarities are sorted in descending order, and the vicinity of the K percentile of the similarity distribution is calculated. N
A threshold value determination method using the average value of the similarities as a threshold value at the time of comparison.

(13) The input registration voice is analyzed and converted into a feature vector, and one node has a tree structure representing the acoustic features of the lower nodes connected to it. The standard pattern of a plurality of speakers having a tree structure having a pattern at the leaf node at the end is linked to the standard pattern of each subword in accordance with the input utterance content, and a word standard pattern of each node is created. When calculating the similarity with the created tree structure speaker standard pattern from the upper node to the lower node, the similarity calculation of the lower node is performed only for the upper M nodes having the higher similarity in each layer. A threshold value determining method in which the similarity of a node that is not calculated is substituted by the similarity of the upper node, and a threshold at the time of matching is obtained from the distribution of the calculated similarity.

(14) An analysis unit that analyzes the input registration voice of the individual and converts it into a feature vector, and a subword-unit speaker standard pattern storage unit that stores a plurality of speaker standard patterns in subword units. The standard patterns in subword units stored in the subword unit speaker standard pattern storage unit are chained in accordance with the input utterance contents, and word standard patterns for each speaker are created. A feature vector and a similarity calculation unit that calculates the similarity between the word standard patterns of a plurality of speakers, a threshold determination unit that obtains a threshold at the time of matching from the distribution of the similarity calculated by the similarity calculation unit, A threshold value determination device comprising:

(15) An analysis unit that analyzes the input registration voice of the individual and converts it into a feature vector, and a subword-unit speaker standard pattern storage unit that stores a plurality of speaker standard patterns in subword units. The standard patterns in subword units stored in the subword unit speaker standard pattern storage unit are chained in accordance with the input utterance contents, and word standard patterns for each speaker are created. And a similarity calculator that calculates the similarity between the feature vector and the word standard pattern of a plurality of speakers. The similarities calculated by the similarity calculator are sorted in ascending order, and the Kth percentile of the similarity distribution is sorted. And a threshold selection calculation unit that uses an average value of the N similarities of the threshold values at the time of matching.

(16) An analysis unit that analyzes the input registration voice of the user and converts it into a feature vector, and each node has a tree structure in which one node represents the acoustic features of the lower nodes connected to it. Subword-based tree-structured speaker standard pattern storage unit for storing, in subword units, a tree-structured speaker-based standard pattern of a plurality of speakers having a tree structure having the speaker's standard pattern at the leaf node at the end; The standard pattern of each subword stored in the storage unit is chained according to the input utterance content, a word standard pattern of each node is created, and the feature vector of the person converted by the analysis unit and the created tree structure story When calculating the degree of similarity with the standard pattern from the upper node to the lower node, the similarity to the upper-level M nodes in each layer is calculated. The similarity calculation of the lower node is performed only when the similarity of the non-calculated node is substituted by the similarity of the upper node, and the similarity calculation is performed based on the similarity distribution calculated by the similarity calculator. And a threshold value determining unit for determining the threshold value of the threshold value.

[0025]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a threshold value determining method and apparatus according to the present invention will be described below. The present invention relates to the determination of a threshold value used for identity determination at the time of collation, and stores a standard pattern in subword units of a plurality of speakers other than the registered speaker, and a subword standard pattern representing the utterance content at the time of registration. To create a word standard pattern, and obtain a distribution of similarity with the registered voice of the person.

The threshold value is determined by regarding the distribution thus created as the distribution of the similarity between the principal standard pattern and the impostor's voice. It is considered that the amount of information of the same speaker's voice and the standard pattern learned by it is equivalent, so that a large number of speakers who closely cover the acoustic space with the stored standard patterns of multiple speakers are considered. By using the standard pattern, the distribution of the similarity between the principal standard pattern and all the impostor voices can be regarded as the distribution of the similarity between the registered voice of the principal and the standard pattern of a plurality of speakers.

By obtaining a threshold from the similarity distribution between the created registered voice of the person and the standard pattern of a plurality of speakers, an appropriate threshold can be set for any impostor.

In addition, since the present method uses a standard pattern in units of subwords, any word standard pattern can be created, and an appropriate threshold value can be obtained even when a different word is uttered for each registered speaker. That's the way you can.

Subwords, which are units of standard patterns, are composed of phonemes, syllables, and semisyllables (Watanabe et al., "Unspecified speaker speech recognition using HMMs with semisyllable units," IEICE Transactions D -II Vol.J75-D-II No.8, pp.1281-1289,1992
-8, hereinafter referred to as Document 2).

As a method of calculating the threshold value, various determination methods can be used. For example, the similarity of the Kth percentile of the distribution of the similarity between the registered voice of the subject and the standard pattern of the multiple speakers is selected as the threshold. Can be used.

When this threshold value determination method is used, the total K
The value of the similarity corresponding to the percentage is a value higher than the threshold. This can be said to be equivalent to setting the reception error rate to K percent. By using this threshold value determination method, it is possible to set a threshold value that gives a desired reception error rate in advance. When the reception error rate is set to 10%, the similarity of the 10th percentile of the similarity distribution is set as the threshold, and when the reception error rate is set to 0%, the maximum similarity is set as the threshold.

(Equation 1)

Further, a method may be used in which the average value of N similarities near the K th percentile of the similarity distribution between the registered voice of the subject and the standard pattern of a plurality of speakers is used as the threshold value. According to this method, when the number of stored speakers is small and the coverage of the acoustic space by the standard pattern is coarse, the decrease in threshold value calculation accuracy can be reduced by calculating the average value of N similarities. . The calculation of the threshold is performed according to the equation (2).

N is the number of speaker standard patterns for performing similarity calculation. As other methods, a method of obtaining from N similarities randomly selected from the similarity distribution, a method of obtaining from N similarities near the median of the similarity distribution, and the like can be used.

A threshold value determining method according to another invention is a method in a case where a stored standard pattern of a plurality of speakers is a tree-structured speaker standard pattern. FIG. 5 shows a simple structure of the tree structure speaker standard pattern.

The tree-structured speaker standard pattern is constituted by nodes corresponding to speaker clusters. The first-stage node corresponds to a cluster of all speakers. Each of the nodes at the first level has a child node, which corresponds to the node at the second level. The upper node has a standard pattern representing the acoustic characteristics of the lower node connected to it. The leaf node at the end is composed of the standard pattern of each speaker.

Regarding speaker clustering, see the following paper. For example, Kai-Fu Lee, "Large-Vocabulary S
peaker-Independent Continuous Speech Recognition:
TheSPHINX System, "CMU-CS-88-148, pp.103-107,1988.4
(Referred to as Reference 3). Kosaka et al., "Tree-structured speaker clustering for speaker adaptation," IEICE Technical Report, SP93-110, pp.49-54,119.
3-12 (referred to as Reference 4).

A threshold value calculation method using a tree-structured speaker standard pattern will be described with reference to FIG.

First, the similarity calculation between the registered voice and the standard pattern of each node in the second stage is calculated, and then, the lower order of the node having the M highest similarity of the calculated similarity is calculated. The similarity calculation between the standard pattern of the node (third stage, the lowest node) and the registered voice is calculated. As described above, from the upper node to the lower node, the similarity between the standard pattern of the lower node and the registered voice of the upper M node having the higher similarity at each stage is calculated.

For the similarity of a node that is not calculated, the value of the upper node is used instead. Finally, the similarities calculated at the leaf nodes are sorted, and the similarity of the Kth percentile of the distribution of the similarities becomes a threshold.

Alternatively, the average value of the N similarities near the Kth percentile of the similarity distribution is used as the threshold. Here, it should be noted that the similarity calculated at the leaf node is equal to or more than K percent of the whole. For example, the tree structure has a structure in which each layer is 4 and each child node connected to each node is 4 (the number of speaker standard patterns of leaf nodes is 64).
When the similarity of the 10th percentile is set as the threshold, the similarity calculated at the leaf node may be 7 or more, so that M> = 2. When a tree-structured speaker standard pattern is used, it is only necessary to calculate the similarity for the lower nodes of the nodes with high similarity at each stage, so the threshold calculation for registration can be performed very quickly. It is.

FIG. 1 is a block diagram showing a first embodiment of a threshold value determining method and apparatus according to the present invention.

The input speech at the time of registration is sent to the analyzer 10, and the utterance content is sent to the similarity calculator 30.
The input voice sent to the analysis unit 10 is converted into a feature vector and sent to the similarity calculation unit 30. As feature vectors, cepstrum, cepstrum, (Furui, "Digital Speech Processing," Tokai University Press, pp.44-47,1985,
This is hereinafter referred to as Reference 5. ) Etc. are used.

In the similarity calculating section 30, the standard patterns in subword units stored in the speaker standard pattern storage section 20 in subword units are chained in accordance with the sent utterance contents, and a word standard pattern for each speaker is created. You. continue,
A similarity calculation between the created word standard pattern of each speaker and the feature vector sent from the analysis unit 10 is performed.

The similarity calculation method includes the Viterbi algorithm (Nakagawa, "Speech Recognition by Stochastic Model," IEICE, 1988, hereinafter referred to as Document 6), DP.
Matching (Sakoe et al., "Similarity Evaluation of Speech Patterns Using Dynamic Programming," Proc. Of the IEICE General Conference,
p.136, 1970.8, hereinafter referred to as Reference 7). The calculated similarity is sent to the threshold determination unit 40.

The threshold determining section 40 obtains a threshold from the similarity distribution transmitted from the similarity calculating section 30 and outputs the threshold.

FIG. 2 is a block diagram showing a second embodiment of the threshold value determining method and apparatus according to the present invention.

The input speech at the time of registration is analyzed by the analyzing unit 110.
And the utterance content is sent to the similarity calculation unit 130.

The input speech sent to analysis section 110 is converted into a feature vector and sent to similarity calculation section 130. The feature vector is the same as in the first embodiment.

In the similarity calculation section 130, the speaker standard pattern storage section 12 in subword units is sent in accordance with the sent utterance contents.
The standard patterns in subword units stored in 0 are chained, and a word standard pattern for each speaker is created. Next, the created word standard pattern of each speaker and the analysis unit 11
A similarity calculation is performed with the feature vector sent from 0. The method of calculating the similarity is the same as in the first invention. The calculated similarity is sent to threshold calculating section 140.

In the threshold value calculation unit 140, the similarity calculation unit 13
The similarities sent from 0 are sorted in descending order of similarity, and the average value of N similarities near the Kth percentile of the similarity distribution is calculated. The calculated average value is output as a threshold.

FIG. 3 is a block diagram showing a third embodiment of the threshold value determining method and apparatus according to the present invention.

The input voice at the time of registration is analyzed by the analysis unit 210.
And the utterance content is sent to the similarity calculating section 230. The input speech sent to the analysis unit 210 is converted into a feature vector and sent to the similarity calculation unit 230. The feature vector is the same as in the first invention.

In the similarity calculation section 230, the standard patterns in subword units stored in the tree-structured speaker standard pattern storage section 220 in subword units are chained in accordance with the sent utterance contents, and a word standard pattern for each node is created. Is done. Next, similarity between the created word standard pattern of each node and the feature vector sent from the analysis unit 210 is calculated.

The method of calculating the similarity between the tree structure speaker standard pattern and the feature vector will be described with reference to the case of the tree structure shown in FIG.

First, the similarity between the feature vector and the standard pattern of each node in the second stage is calculated. Next, the similarity calculation between the standard pattern and the feature vector of the lower node (third stage, the lowest node) of the node having the highest M similarity of the calculated similarity is performed. The similarity of a node that is not calculated is substituted by the similarity of its upper node. Thus, the similarity calculation is performed from the upper node toward the leaf node, and the finally calculated similarity is sent to the threshold value determination unit 240.

In the threshold value determining section 240, the similarity calculating section 23
The similarities sent from 0 are sorted in descending order of similarity, and the similarity corresponding to the Kth percentile of the similarity distribution is output as a threshold. Alternatively, the average value of the N similarities near the Kth percentile of the similarity distribution is output as the threshold.

[0057]

As described above, according to the present invention,
A threshold for obtaining a desired reception error rate can be set in advance, and an appropriate threshold can be obtained even when a different word is uttered for each registered speaker.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing a first embodiment of a threshold value determining device according to the present invention.

FIG. 2 is a configuration diagram showing a second embodiment of the threshold value determining device according to the present invention.

FIG. 3 is a configuration diagram showing a third embodiment of the threshold value determining device according to the present invention.

FIG. 4 is a diagram for explaining a conventional threshold value determining method.

FIG. 5 is a configuration diagram of a conventional threshold value determining device.

[Explanation of symbols]

10, 110, 210 Analysis unit 20, 120 Subword unit speaker standard pattern storage unit 30, 130, 230, 330 Similarity calculation unit 40, 240 Threshold determination unit 140, 360 Threshold calculation unit 220 Subword unit tree structure speaker Standard pattern storage unit 310 Personal standard pattern creation unit 320 Personal standard pattern storage unit for each word 340 Speech storage units for multiple speakers 350 Average value and variance calculation unit 360 Threshold calculation unit

Claims (16)

[Claims] In a method for determining a threshold value in speaker verification, which is a criterion for personal identification when comparing an input pattern and a standard pattern, a standard pattern in subword units of a plurality of speakers other than a registered speaker is stored, and Speaker verification characterized by creating a word standard pattern by concatenating subword standard patterns representing the utterance contents, obtaining a distribution of similarity with the registered voice of the person, and determining a threshold based on the generated distribution. Threshold determination method. 2. The method according to claim 1, wherein the sub-word is a phoneme, a syllable or a semi-syllable. 3. The speaker verification according to claim 1, wherein said threshold value is determined as a similarity of a K-percentile of a similarity distribution between a registered voice of the individual and a standard pattern of said plurality of speakers. Threshold determination method. 4. The method according to claim 1, wherein the threshold value is determined as an average value of N similarities near the K th percentile of a similarity distribution between a registered voice of the subject and a standard pattern of a plurality of speakers. Threshold determination method for speaker verification. 5. The method according to claim 1, wherein the threshold value is determined based on N similarities randomly selected from a similarity distribution between a registered voice of the subject and a standard pattern of a plurality of speakers. Threshold determination method for speaker verification. 6. The talk according to claim 1, wherein the threshold value is determined based on N similarities near the median of the similarity distribution between the registered voice of the subject and the standard pattern of a plurality of speakers. Threshold determination method in person verification. 7. A standard pattern of a plurality of speakers is constituted by nodes corresponding to speaker clusters. A first-stage node corresponds to a cluster of all speakers, and each node of the first stage is a child. The node corresponds to the second-stage node, and the upper node has a standard pattern representing the acoustic characteristics of the lower node connected thereto, and the leaf node at the end is a tree-structured speaker composed of the standard pattern of each speaker. The similarity between the registered voice and the standard pattern of each node in the second stage is stored as a standard pattern, and the standard pattern of the lower node of the node having the upper M similarity of the calculated similarity is calculated. The similarity calculation between the registered voice and the registered voice is calculated from the upper node to the lower node. Calculate degree , The similarity of not compute nodes,
A threshold value determination method in speaker verification, characterized in that a threshold value is determined based on a distribution of similarity / similarity finally calculated at a leaf node by using the value of the upper node instead. 8. The method according to claim 7, wherein the similarity of the Kth percentile of the similarity distribution is determined as a threshold. 9. The method according to claim 7, wherein the similarity of an average value of N similarities near the Kth percentile of the similarity distribution is determined as a threshold. 10. The method according to claim 7, wherein the similarity of the average of the N similarities near the Kth percentile of the similarity distribution is determined as the threshold. 11. Analyzing the input registration voice of the person,
After converting the standard pattern of a plurality of speakers into a standard pattern in subword units according to the input utterance contents, a word standard pattern of each speaker is created, A method for determining a threshold value in speaker verification, comprising calculating a similarity with a word standard pattern of the subject and calculating a threshold at the time of verification from a distribution of the calculated similarity. 12. Analyzing the input registration voice of the person,
After converting the standard pattern of a plurality of speakers into a standard pattern in subword units according to the input utterance contents, a word standard pattern of each speaker is created, The similarity with the word standard pattern is calculated, the calculated similarities are sorted in descending order, and the average value of N similarities in the vicinity of the Kth percentile of the similarity distribution is used as a threshold value for comparison. Threshold determination method in speaker verification. 13. Analyzing the input registration voice of the person,
This is a tree structure in which one node represents the acoustic characteristics of the lower nodes connected to it, and the standard pattern of each speaker is a tree structure with the standard pattern of each speaker at the leaf node at the end. Are chained according to the input utterance content of the standard pattern in subword units, a word standard pattern of each node is created, and the similarity between the feature vector of the subject and the created tree structure speaker standard pattern is determined from the upper node. When calculating toward lower nodes,
Only for the top M nodes having high similarity in each hierarchy, the similarity of the lower node is calculated, and the similarity of the non-calculated node is substituted by the similarity of the upper node. A threshold determination method in speaker verification, wherein a threshold at the time of verification is obtained. 14. Analyzing the input registration voice of the person,
An analyzing unit that converts the feature vector into a feature vector; a speaker standard pattern storage unit that stores a plurality of speaker standard patterns in subword units; and a subword unit that is stored in the subword unit speaker standard pattern storage unit. Are chained according to the input utterance contents, and a word standard pattern of each speaker is created, and the similarity between the feature vector converted by the analysis unit and the word standard pattern of a plurality of speakers is calculated. A threshold value determining unit for speaker verification, comprising: a similarity calculating unit that calculates a threshold value for matching based on a distribution of similarities calculated by the similarity calculating unit. 15. An analysis unit for analyzing an input personal registration voice and converting it into a feature vector; a subword-unit speaker standard pattern storage unit for storing a plurality of speaker standard patterns in subword units; The sub-word-unit standard patterns stored in the sub-word-unit speaker standard pattern storage unit are chained in accordance with the input utterance contents, word standard patterns for each speaker are created, and the characteristics of the individual converted by the analysis unit are used. A similarity calculator that calculates the similarity between the vector and the word standard pattern of a plurality of speakers; and the similarities calculated by the similarity calculator are sorted in ascending order, and N near the Kth percentile of the distribution of the similarity is sorted. And a threshold selection calculation unit that uses an average value of the similarities as a threshold at the time of comparison. 16. Analyzing the input personal registration voice,
An analysis unit that converts to a feature vector, and a plurality of speakers with a tree structure in which one node represents the acoustic features of the lower nodes connected to it, with the standard pattern of each speaker at the leaf node at the end A tree-structured speaker standard pattern storage unit that stores the standard pattern of the sub-words in units of sub-words; To create a word standard pattern of each node, and calculate the similarity between the feature vector of the person converted by the analysis unit and the created tree structure speaker standard pattern from the upper node to the lower node. In this case, the similarity of the lower nodes is calculated only for the upper M nodes having a high similarity in each layer, and the calculation is not performed. A similarity calculator that substitutes the similarity of the node with the similarity of the upper node; and a threshold determiner that obtains a threshold at the time of matching from the distribution of the similarity calculated by the similarity calculator. A threshold determination device in speaker verification characterized by the following.