JP2000163092A - Method and device for collating speaker - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve a speaker collation performance. SOLUTION: A personal dictionary creating part 13 generates a phoneme dictionary of a new user based on the speech read when the speech is registered, and stores it in a personal dictionary storage part 23. This is a setting of a set of cohort speakers in units of speakers. At the time of registering, a by-phoneme neighbor speaker selection part 17 selects, by phoneme, speaker information near to each phoneme in the acoustic space by each phoneme of a new phoneme dictionary from the phoneme dictionary of the registered speakers already stored in the storage part 23. The speaker information are arranged in a nearest-first order in distance and stored in the by- phoneme neighbor speaker table storage part 21. This is a setting of a set of cohort speakers in units of phonemes. A personal verification part 15 calculates a likelihood of the phoneme dictionary of the person in question to the speech read at the time of authentication. Based on the information in the storage part 21, a background speaker dictionary synthesis part 19 synthesizes a background speaker phoneme dictionary by combining each phoneme of the neighbor speaker with each phoneme in the speaker phoneme dictionary of the person in question. A mean value of the likelihood of the background speakers is calculated to the input speech. Acceptance/rejection is judged by calculating a normalized likelihood and comparing large-and-small relations with a pre-set threshold.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of setting a background speaker for a true speaker, and normalizing a matching score of the true speaker with respect to the input voice by a matching score of the background speaker with respect to the input voice. The present invention relates to improvement of a person verification method and apparatus.

[0002]

2. Description of the Related Art In a speaker verification apparatus, acceptance / rejection is determined based on a magnitude relationship between a verification score of a dictionary (acoustic model) of a person himself / herself with respect to an input voice and a threshold value set in advance. If the acoustic model is an HMM (Hidden Markov Model), the matching score is the likelihood of the HMM for the input speech. For this reason, a change in the matching score (likelihood) due to the difference in the utterance content and the influence of the voice input system greatly deteriorates the matching performance of the device. Therefore, to reduce the likelihood fluctuation,
Several techniques have been studied to set a background speaker for each speaker and normalize the likelihood of the speaker for the input speech by dividing the likelihood of the background speaker for the same input speech. ing.

[0003] Examples of the above methods include Higgins et al., Which uses the speaker with the highest likelihood as the background speaker, Rosenberg et al., Which uses a cohort speaker set as the background speaker, and Liu et al. A normalization method based on the average value and the variance value of the likelihood, and a normalization method based on the likelihood of a multi-speaker model based on the posterior probability by Matsui et al. Each of the above techniques is directed to accurate and efficient representation of a background speaker.

[0004]

By the way, in each of the above-mentioned methods (that is, a method of normalizing a true speaker based on a likelihood ratio between a true speaker and a background speaker), the acoustic model of the background speaker is
It is desirable to represent the acoustic space for all speakers other than the true speaker. However, it is difficult to find an acoustic model for all ideal speakers. Further, even in an actual service, obtaining an acoustic model from a large number of registered speakers is not efficient because a large calculation cost is required. Hereinafter, a method of Rosenberg et al. (That is, a method using a cohort speaker set) as a background speaker will be described in detail.

In designing a cohort speaker set as a background speaker, it is important to accurately and efficiently approximate the entire speaker space with a small amount of calculation. Since the likelihood is expressed as a logarithmic value, it is possible to improve the resolution of the judgment criterion (that is, the likelihood ratio) of the collation by selecting the background speaker as close as possible to the true speaker. It is thought.

Conventionally, in the above-mentioned method, among the registered speakers other than the person himself / herself, as shown in FIG. 1, the upper N persons who are close to the phoneme dictionary (acoustic model) of the person himself / herself in the acoustic space are used. A selection was made (N = 2 in FIG. 1), and the average value of the likelihoods for the N input voices was used as the likelihood of the background speaker. Alternatively, in the above-mentioned method, there is a case where only one other speaker closest to the likelihood value of the true speaker is selected.
The reason is that the acoustic model of a speaker is accurate in the likelihood for input speech near the acoustic model in the acoustic space, but has an error in the likelihood for input speech far from the acoustic model. By big. In other words, the normalized likelihood must be accurate for the true speaker and for an impostor entering the sound space near the true speaker, but at a distance from the true speaker. It doesn't have to be as accurate for impostors entering into distant space.

However, in the above-described method of selecting a nearby speaker on a speaker-by-speaker basis, to be precise, the neighboring speakers are different for each phoneme (that is, the inter-speaker distance is different for each phoneme). Is not close to the person in all phonemes but includes an error, and thus does not satisfy the condition of the background speaker that the speaker is close to the person in the acoustic space. For this reason, it has been considered that the accuracy of the likelihood normalization is deteriorated and the matching performance is degraded. That is, since the likelihood of the background speaker can be controlled only on a speaker-by-speaker basis, it has been considered that the resolution of the criterion in speaker verification is not necessarily high.

Accordingly, it is an object of the present invention to improve the performance of speaker verification.

[0009]

A speaker verification apparatus according to a first aspect of the present invention sets a background speaker for a true speaker, and obtains a verification score of the true speaker with respect to an input voice, and outputs the verification score of the input voice. Is normalized by the matching score of the background speaker with respect to, and a speaker whose acoustic feature is relatively close to the registered speaker's acoustic model in terms of local features is selected from the acoustic models of each speaker And the selected speaker
By combining the means for retaining the registered speaker as a nearby speaker and the local features of the acoustic model of the neighboring speaker in correspondence with each local feature of the acoustic model of the true speaker, the background speaker Means for synthesizing the acoustic model.

According to the above configuration, the performance of speaker verification can be improved.

[0011] In a preferred embodiment according to the first aspect of the present invention, each phoneme of the acoustic model is a hidden Markov model (H
MM), and the matching score is the HMM for the input speech.
Is the likelihood of The local feature is a phoneme of the acoustic model of each speaker. In the modification of the above embodiment, the local feature is the state of each phoneme HMM. In another variation, the local feature is a respective distribution of each state of each phoneme HMM.

In the above embodiment, the selecting means selects, from the acoustic models of the speakers, a speaker whose phoneme is relatively short in the acoustic space from the acoustic model of the registered speaker. In a modification of the above embodiment, the selection unit selects, from among the acoustic models of the speakers, a speaker whose distance in the acoustic space from the acoustic model of the registered speaker is relatively short with respect to each state of each phoneme HMM. select. In another modified example, the selecting unit may select, from among the acoustic models of the speakers, a talk that has a relatively short distance in the acoustic space from the acoustic model of the registered speaker with respect to the distribution of each state of each phoneme HMM. Choose a person.

In the above embodiment, the synthesizing means synthesizes the acoustic model of the background speaker by combining the phonemes of the acoustic model of the neighboring speaker with the phonemes of the acoustic model of the true speaker. In a modification of the above embodiment, the synthesis means combines the respective states of the phoneme HMMs of the neighboring speaker in correspondence with the respective states of the phoneme HMMs of the true speaker, thereby forming an acoustic model of the background speaker. Are synthesized. In another modification, the synthesis means combines the distribution of each state of each phoneme HMM with the distribution of each state of each phoneme HMM of the true speaker, thereby combining the distribution of the background speaker. Synthesize an acoustic model.

In the speaker verification method according to the second aspect of the present invention, a background speaker is set for a true speaker, and the likelihood of the true speaker with respect to the input voice is determined based on the likelihood of the background speaker with respect to the input voice. The first method is to select a speaker whose acoustic feature is relatively close to the registered speaker's acoustic model in the acoustic space from among the acoustic models of the speakers from among the acoustic models of the speakers.
And the second step of retaining the selected speaker as a neighboring speaker of the registered speaker, and the local feature of the acoustic model of the neighboring speaker is expressed as a local characteristic of each of the acoustic models of the own speaker. And synthesizing the acoustic model of the background speaker by combining them in accordance with the characteristic features.

A program medium according to a third aspect of the present invention sets a background speaker for a true speaker, and calculates a verification score of the true speaker for the input voice by using a verification score of the background speaker for the input voice. Means for selecting, from the speaker acoustic models of each speaker, a speaker whose distance in the acoustic space with respect to the local speaker's acoustic model is relatively short from among the acoustic models of each speaker; Talked speaker
By combining the means for retaining the registered speaker as a nearby speaker and the local features of the acoustic model of the neighboring speaker in correspondence with each local feature of the acoustic model of the true speaker, the background speaker And a computer program for operating a computer as each unit in the speaker verification device.

[0016]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 2 is a functional block diagram of the speaker verification device according to one embodiment of the present invention.

The apparatus has a phoneme-specific neighborhood speaker table, and is configured to be able to selectively set one of a voice registration mode and a voice authentication mode. As shown in the figure, a voice input unit (input unit) 11 , Personal dictionary creation unit (creation unit) 13
, A personal authentication unit (authentication unit) 15, and a phoneme-specific neighborhood speaker selection unit (selection unit) 17. The apparatus further includes a background speaker dictionary synthesizing unit (synthesizing unit) 19 in addition to the above units.
It also includes a phoneme-specific neighborhood speaker table storage unit (storage unit) 21 and a personal dictionary storage unit (storage unit) 23.

The creating unit 13 creates a phoneme dictionary (acoustic model) of a new user (speaker) based on the voice information read through the input unit 11 during the voice registration mode of the apparatus.
It is stored in the storage unit 23. For example, as shown in the figure, when the new user is speaker 1, speaker 1 including phonemes 1 to P
When the phoneme dictionary of the speaker 2 is the speaker 2, the phoneme dictionary of the speaker 2 including the phonemes 1 to P in the same manner as described above. The phoneme dictionary of the speaker S including the phonemes P is created by the creating unit 13. The phoneme dictionaries of the speakers 1 to S are stored in the storage unit 23 by the creation unit 13.

The above processing operation in the creating unit 13 is generalized as setting of a cohort speaker set for each speaker. That is, when a cohort speaker set is set for each speaker, the likelihood ratio L (i) for the speaker i is expressed by the following equation (1).

[0021]

(Equation 1) In equation (1), ο is the observation vector, λ (i) is the phoneme HMM set of speaker i, K is the cohort size, and ck
(I) is the Kth cohort speaker of speaker i, P is the phoneme number,
Show each. For ck (i) (k = 1, 2,..., K), the top K persons near the speaker i from the registered speakers are selected based on the inter-speaker distance.

The selecting unit 17 performs a phoneme dictionary creation process (setting of a cohort speaker set in units of speakers) by the creating unit 13 in the voice registration mode of the apparatus, in parallel with the newly created phoneme dictionary. For each phoneme, the speaker information whose distance in the acoustic space is short from each phoneme is selected for each phoneme from the phoneme dictionaries of a plurality of registered speakers stored in the storage unit 23. Then, in a state where the speaker information is arranged in ascending order of distance, the storage unit 2 stores a phoneme-based neighbor speaker table.
1 is stored. For example, when the new user is speaker 1, the speakers closest to phoneme 1 of speaker 1 are speaker 9, speaker 7,... if,
As shown in the drawing, the selection unit 17 stores the speaker 9, the speaker 7,..., And the speaker 11 in the phoneme-based neighboring speaker table relating to the phoneme 1 of the speaker 1.
Is stored. Assuming that the speakers closest to the phoneme P of the speaker 1 are the speakers 8, the speakers 21,..., And the speakers 14 in the order from the speaker having the shortest distance, the selecting unit 17 selects Speaker 1
Speaker 8 and Speaker 2 in the phoneme-specific neighborhood speaker table for phoneme P
1,..., The speaker 14 is stored. Further, when the new user is speaker S, the speakers closest to phoneme 1 of speaker S were speaker 2, speaker 30,..., And speaker 19 in order from the speaker having the shortest distance. Then, the selector 17 stores the speaker 2, the speaker 30,..., And the speaker 19 in the phoneme-by-phoneme neighboring speaker table relating to the phoneme 1 of the speaker S as illustrated. The speakers closest to the phoneme P of the speaker S are speaker 24, speaker 13,
..., If the speaker 18 is selected, the selection unit 17 causes the speaker 24, the speaker 13,... 18 is stored.

The above-described processing operation in the selection unit 17 is generalized as a processing operation for setting a cohort speaker set in units of phonemes. That is, when the cohort speaker set is set in units of phonemes, the second term on the right side of the above equation (1) is expressed by the following equation (2).

[0024]

(Equation 2) In equation (2), #ck (i) is the K-th virtual cohort speaker of speaker i, and λ (#ck (i)) is the virtual cohort speaker #
Each represents a phoneme HMM set of ck (i).

Next, the phoneme set of the speaker i is expressed by the following (3)
It is expressed by an equation.

[0026]

(Equation 3) Next, from equation (3), the phoneme HMM set of the virtual cohort speaker is represented by the following equation (4).

[0027]

(Equation 4) In equation (4), ck (i, p) is the phoneme p of speaker i.
Represents the kth cohort speaker. Where ck (i, p)
Is a cohort speaker in phoneme units, and the top K people near speaker i are selected based on the distance between phoneme models between phoneme p of speaker i and phoneme p of another registered speaker. is there.

The authentication unit 15 calculates the likelihood LS of the speaker's phoneme dictionary (acoustic model) for the voice information read through the input unit 11 in the authentication mode of the apparatus.

In the synthesizing unit 19, the authentication unit 15 performs the phoneme dictionary (acoustic model) of the speaker on the voice information as described above.
In parallel with the calculation of the likelihood LS of each phoneme (1 to S) in the phoneme dictionary of the speaker (1 to S) based on the information in the storage unit 21.
~ P), the phoneme dictionary of the background speaker is synthesized by combining the phonemes of the neighboring speakers of each phoneme (1 to P). Then, the average value of the likelihood Lb of the background speaker with respect to the input voice information is calculated. Thereafter, the normalized likelihood Ln = Ls / Lb is calculated, the magnitude relationship between the threshold value set in advance and Ln is compared, and the acceptance / rejection is determined.

As described above, according to one embodiment of the present invention, when the selection unit 17 is in the voice registration mode of the device,
For each phoneme in the newly created phoneme dictionary, speaker information that is close to each phoneme in the acoustic space is selected for each phoneme from the phoneme dictionaries of a plurality of registered speakers already stored in the storage unit 23. Then, the speaker information is stored in the storage unit 21 as a phoneme-specific neighborhood speaker table in a state where the speaker information is arranged in ascending order of distance. Therefore, in the acoustic space,
It is possible to synthesize a phoneme dictionary of a background speaker closer to the phoneme dictionary (acoustic model) of the native speaker, and use the likelihood of the synthesized background speaker for normalization of the likelihood of the native speaker. Thus, the likelihood can be normalized with high accuracy. As a result, a speaker verification device with high authentication accuracy can be realized.

FIG. 3 is a functional block diagram of a speaker verification device according to a first modification of the embodiment of the present invention.

The apparatus has a state-specific nearby speaker table and is configured to be able to selectively set one of a voice registration mode and a voice authentication mode. The above-described apparatus has a state-specific neighborhood speaker selection section (selection section) 25 instead of the phoneme-based neighborhood speaker selection section 17 shown in FIG. 2 and a phoneme-based neighborhood speaker table storage section 21. 2 is different from the speaker verification device shown in FIG. 2 in that a state-specific nearby speaker table storage unit (storage unit) 27 is provided. Components other than the above are denoted by the same reference numerals as those shown in FIG. The processing operations of the creating unit 13 and the synthesizing unit 19 shown in the drawing are slightly different from those of the creating unit 13 and the synthesizing unit 19 shown in FIG.

For example, as shown in FIG.
When it is 1, phoneme 1 state 1, ..., phoneme 1 state J, ...
, A phoneme P state 1,..., A phoneme P state J, and a phoneme dictionary of the speaker 1 is created by the creating unit 13. Also, when the speaker is S, the phoneme 1 state 1,...
J, ..., phoneme P state 1, ..., speaker S consisting of phoneme P state J
Is created by the creating unit 13. The phoneme dictionaries of the speakers 1 to S are stored in the storage unit 23 by the creation unit 13.

In the voice registration mode of the apparatus, the selection unit 25 simultaneously executes the phoneme dictionary creation processing performed by the creation unit 13 with each of the states of each phoneme of the newly created phoneme dictionary. Speaker information having a short distance in space is selected for each state from the phoneme dictionaries of a plurality of registered speakers stored in the storage unit 23. Then, the speaker information is stored in the storage unit 27 as a state-specific neighbor speaker table in a state where the speaker information is arranged in ascending order of distance. For example, when the new user is the speaker 1, the speakers closest to the state 1 of the phoneme 1 of the speaker 1 are the speakers 9, the speakers 7,... , The selection unit 25 stores the speaker 9, the speaker 7,..., And the speaker 11 in the phoneme-specific neighborhood speaker table relating to the state 1 of the phoneme 1 of the speaker 1 as illustrated. Is done. The speakers closest to the state J of the phoneme 1 of the speaker 1 are the speakers 20, 4,.
.., Assuming that the speaker 15 is the speaker 15, the selector 20, the speaker 20, the speaker 4,... Speaker 15 is stored. Assuming that the speakers closest to the state 1 of the phoneme P of the speaker 1 are the speakers 14, the speakers 41,... As described above, in the phoneme-specific neighborhood speaker table according to the state 1 of the phoneme P of the speaker 1,
The speakers 41,..., And the speaker 12 are stored. The speakers closest to the state J of the phoneme P of the speaker 1 are the speakers 1 in the order of the closest speaker
7, speaker 21,..., Speaker 32, the selection unit 2
5, as shown in the figure, the speaker 17, the speaker 21,..., The speaker 17 are stored in the phoneme-by-phoneme neighboring speaker table relating to the state J of the phoneme P of the speaker 1.
32 is stored. When the new user is the speaker S, the speakers closest to the state 1 of the phoneme P of the speaker S are the speakers 8, 11,. If so, the selection unit 25 causes the speaker 8 and the speaker 8 to be stored in the phoneme-by-phoneme neighboring speaker table relating to the state 1 of the phoneme P of the speaker S, as shown in the figure.
Speakers 11,..., And speaker 36 are stored. The speakers closest to the state J of the phoneme P of the speaker S are the speakers 1 in order from the closest speaker.
8, speaker 3,..., Speaker 16, if the selection unit 25
As a result, as shown in the figure, the speaker 18, the speaker 3,..., The speaker 16 are stored in the phoneme-specific neighborhood speaker table relating to the state J of the phoneme P of the speaker 1.
Is stored.

The above processing operation in the selection section 25 is generalized as a processing operation for setting a cohort speaker set in units of states. Assuming that the phoneme HMM is a left-to-right mixed continuous HMM with the number of states S and the number of mixtures M,
The HMMλp of the phoneme p is a transition probability ap, s, s which loops in a certain state s, a transition probability ap, s, s + 1 to the next state,
The distribution weights wp, s, m {m = 1, 2,..., M} are stored as parameters. In the cohort speaker set in units of states, since the neighbor speaker of speaker i is selected for each parameter held in state s as described above, the HMM set of the virtual cohort speaker is represented by the following (5) ) Expression. ck (i, p, s) is a cohort speaker in units of states, and the top K speakers in the vicinity are determined by the phoneme p of speaker i, the state distance between state s, and the same state of the registered speaker. Selected.

[0036]

(Equation 5) In the authentication mode of the device, the authentication unit 15 calculates the likelihood LS of the phoneme dictionary (acoustic model) of the speaker with respect to the voice information read through the input unit 11.

The synthesizing unit 19 determines that the authentication unit 15 has the likelihood L
In parallel with the calculation of S, based on the information in the storage unit 27, for each state (1 to J) of each phoneme (1 to P) in the phoneme dictionary of the speaker (1 to S), Each state (1
To J), the phoneme dictionary of the background speaker is synthesized by combining the parameters. Then, the average value of the likelihood Lb of the background speaker with respect to the input speech information is calculated, and then the normalized likelihood Ln = Ls / Lb, as in the apparatus shown in FIG.
Is calculated, and the magnitude relationship between the threshold value set in advance and Ln is compared to determine the acceptance / rejection.

FIG. 4 is an explanatory diagram showing a configuration of an HMM having a number of states J and a number of mixtures M according to a first modification of the embodiment of the present invention.

In FIG. 4, the parameters of the HMM are composed of a transition probability A, a distribution weight W, and a distribution N, and the parameters included in the state j are Aj, j, Aj, j + 1, W
j, 1... Wj, M and Nj, 1... Nj, M.

FIG. 5 is a functional block diagram of a speaker verification device according to a second modification of the embodiment of the present invention.

The above apparatus has a distribution-specific nearby speaker table, and is configured so that one of a voice registration mode and a voice authentication mode can be selectively set. The above-described apparatus has a point that a neighboring speaker selecting section (selecting section) 29 for each distribution is provided instead of the neighboring speaker selecting section 17 for each phoneme shown in FIG. 2 is different from the speaker verification device shown in FIG. 2 in that a neighboring speaker table storage unit (storage unit) 31 for each distribution is provided. Components other than the above are denoted by the same reference numerals as those shown in FIG. The processing operations of the creating unit 13 and the synthesizing unit 19 shown in the drawing are slightly different from those of the creating unit 13 and the synthesizing unit 19 shown in FIG.

For example, as shown in FIG.
When it is 1, phoneme 1 state 1 distribution 1, ..., phoneme 1 state 1 distribution M, ..., phoneme P state J distribution 1, ..., phoneme P state J distribution M
The phoneme dictionary of the speaker 1 is created by the creating unit 13. Also, when the speaker is S, the phoneme 1 state
, A phoneme 1 state 1 distribution M,..., A phoneme P state J distribution 1,... And those speakers 1 ~ S
Are stored in the storage unit 23 by the creation unit 13.

In the voice registration mode of the apparatus, the selection unit 29 performs, in parallel with the operation of the phoneme dictionary creation by the creation unit 13, the distribution of each state of each phoneme of the newly created phoneme dictionary. Speaker information whose distance in the acoustic space is short from each distribution is selected for each distribution from the phoneme dictionaries of a plurality of registered speakers stored in the storage unit 23. Then, in a state where the speaker information is arranged in ascending order of distance, a storage unit 31 as a distribution-specific neighborhood speaker table is arranged.
To be stored. For example, when the new user is speaker 1, speakers close to distribution 1 of state 1 of phoneme 1 of speaker 1 are speakers 9, speaker 7,. Assuming that the speaker 11 is the speaker 11, as shown in FIG.
Speaker 9, speaker 7,..., Speaker 11 are stored in the distribution-specific neighboring speaker table relating to distribution 1 of state 1 of one phoneme 1. If the speakers closest to the distribution M of the state 1 of the phoneme 1 of the speaker 1 are the speakers 20, 4,... , As shown, phoneme 1 of speaker 1
Speaker 2 in the distribution-specific nearby speaker table for distribution M in state 1 of
0, speaker 4,..., Speaker 15 are stored. If it is assumed that the speakers closest to the distribution 1 of the state J of the phoneme P of the speaker 1 are the speakers 14, the speakers 41,. , As shown, state J of phoneme P of speaker 1
Speaker 14 and speaker 4 in the distribution-specific neighborhood speaker table for distribution 1
1,..., The speaker 12 is stored. The speakers closest to the distribution M of the state J of the phoneme P of the speaker 1 are the speakers 1 in the order of the nearest speaker.
7, speaker 21,..., Speaker 32, the selection unit 2
9, the distribution M of the state J of the phoneme P of the speaker 1 as shown in FIG.
Speaker 17, speaker 21, ... in the distribution-specific neighborhood speaker table according to.
..., speaker 32 is stored. When the new user is the speaker S, the speakers closest to the distribution 1 of the state J of the phoneme P of the speaker S are the speakers 8, 11,.
.., If it is speaker 36, as shown in the figure, the selection unit 29 stores the speaker 8, the speaker 11, and the speaker 11 in the distribution-specific neighboring speaker table according to the distribution 1 of the state J of the phoneme P of the speaker S. ..., The speaker 36 is stored. The speakers closest to the distribution M of the state J of the phoneme P of the speaker S are the speakers 18, 3,.
If it is 6, the selection unit 29 causes the speaker 18, the speaker 3,... Person 16 is stored.

The above processing operation in the selection unit 29 is generalized as a processing operation for setting a cohort speaker set in units of distribution. In the cohort speaker set in units of distribution, the HMM set of the virtual cohort speaker is represented by the following equation (6).

[0045]

(Equation 6) Since there are M cohort speaker sets distributed in the same state, the transition probability a is the sum of the transition probabilities from each cohort speaker set, and the sum of the self-loop and the transition probability to the next state is 1. It is calculated by re-normalization so that The transition probability a is represented by the following equation (7).

[0046]

(Equation 7) Similarly, the distribution weight w is calculated by renormalizing the distribution weights from each cohort speaker set so that the sum becomes 1. The distribution weight w is expressed by the following equation (8).

[0047]

(Equation 8) In equations (6), (7) and (8), ck (i, p, s,
m) is the cohort speaker in units of distribution, more specifically, the k-th cohort speaker for the distribution m of the state s of the phoneme p of the speaker i, and the neighboring speakers selected for each distribution. is there.

In the authentication mode of the device, the authentication unit 15 calculates the likelihood LS of the phoneme dictionary (acoustic model) of the speaker with respect to the voice information read through the input unit 11.

The synthesizing unit 19 determines that the authentication unit 15 has the likelihood L
In parallel with the calculation of S, based on the information in the storage unit 31, each state (1 to J) of each phoneme (1 to P) of the phoneme dictionary of the speaker (1 to S) is determined. The phoneme dictionary of the background speaker is synthesized by combining the distributions (1 to M) with the distributions (1 to M) of the neighboring speakers, respectively. Then, the average value of the likelihood Lb of the background speaker with respect to the input voice information is calculated. In this case, the transition probability A of the background speaker is the transition probability A of the speech dictionary of the true speaker.
Is diverted as it is. Then, as in the apparatus shown in FIG. 2, the normalized likelihood Ln = Ls / Lb is calculated, and the magnitude relationship between the threshold value set in advance and Ln is compared to determine acceptance / rejection. The parameters included in the distribution m of the state J described above are Wj, M and Nj, M in FIG.

Next, an evaluation experiment of the apparatus shown in FIGS. 2, 3 and 5 and a conventional apparatus will be described.

First, as experimental conditions, FIG. 6 shows specifications of audio data used in an evaluation experiment for each of the above devices.

The collation method is a text designation type.
To perform a closed experiment (that is, an experiment in which a registered speaker other than the registered speaker is used as a spoofer) and set B to an open experiment (that is, an experiment in which a speaker other than the registered speaker is used as a spoofer). Used as a person. The HMM was a context-independent phoneme HMM having three states and three mixtures, and in speaker registration, all parameters of the HMM were estimated by ML estimation.

Next, a description will be given of the experimental results and their considerations.

In the above-described experiment, the performances of the respective methods were compared based on the EER (equal error rate) based on a threshold value that was given afterward. FIG. 7 shows data (closed data) obtained as a result of the closed experiment, and FIG. 8 shows data (open data) obtained as a result of the open experiment.
In FIG. 7, a curve 33 represents a cohort by speaker (that is, related to the conventional apparatus) EER (unit%), and a curve 35 represents a cohort by phoneme (that is, related to the apparatus shown in FIG. 2) E
The ER (unit%) and the curve 37 indicate the cohort EER (unit%) for each state (that is, according to the apparatus shown in FIG. 3). Further, the curve 39 indicates the cohort by distribution (that is, according to the apparatus shown in FIG. 5) EER (unit%).

On the other hand, also in FIG. 8, a curve 41 which is the same kind of curve as the curve 33 shown in FIG.
R (unit%) is represented by a curve 43 which is the same kind of curve as the curve 35.
Indicates a cohort EER (unit%) for each phoneme, and a curve 45 which is the same kind of curve as the curve 37 indicates a cohort EER (unit%) for each state. Further, a curve 47, which is the same kind of curve as the curve 39, shows a cohort EER (unit%) for each distribution.

As is clear with reference to FIGS. 7 and 8,
In both closed and open experiments, the EER decreases as the size of the cohort speaker set increases, and tends to saturate around 4.5. In both experiments, selecting and designing a cohort speaker set in more detailed units can reduce EER, and in the case of size 5, the method using distribution as a unit is smaller than the method using speaker as a unit. As a result, a high error reduction rate of 67% was obtained in the closed experiment and 35% in the open experiment.

Further, the matching performance in size 2 of the method using distribution as a unit is almost equal to the matching performance in size 5 of the method using speaker as a unit, and the matching performance according to each embodiment of the present invention is smaller. High-resolution likelihood normalization was achieved with the size of the cohort speaker set.

The contents described above relate only to an embodiment of the present invention and modifications thereof, and needless to say that the present invention is not limited to only the above contents.

[0059]

As described above, according to the present invention,
The performance of speaker verification can be improved.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram showing a manner of setting a background speaker.

FIG. 2 is a functional block diagram of the speaker verification device according to the embodiment of the present invention.

FIG. 3 is a functional block diagram of a speaker verification device according to a first modification of the embodiment of the present invention.

FIG. 4 is an explanatory diagram showing a configuration of an HMM having a number of states J and a number of mixtures M according to a first modification of the embodiment of the present invention.

FIG. 5 is a functional block diagram of a speaker verification device according to a second modification of the embodiment of the present invention.

FIG. 6 is an explanatory diagram showing, as experimental conditions, specifications of voice data used in an evaluation experiment on a conventional speaker verification device and each speaker verification device according to the present invention.

FIG. 7 is a diagram showing data obtained as a result of performing closed experiments on a conventional speaker verification device and each speaker verification device according to the present invention.

FIG. 8 is a diagram showing data obtained as a result of conducting an open experiment on a conventional speaker verification device and each speaker verification device according to the present invention.

[Explanation of symbols]

 Reference Signs List 11 voice input unit (input unit) 13 personal dictionary creation unit (creation unit) 15 personal authentication unit (authentication unit) 17 neighborhood speaker selection unit by phoneme (selection unit) 19 background speaker dictionary synthesis unit (synthesis unit) 21 Phoneme-specific neighborhood speaker table storage unit (storage unit) 23 Personal dictionary storage unit (storage unit) 25 State-specific neighborhood speaker selection unit (selection unit) 27 State-specific neighborhood speaker table storage unit (storage unit) 29 Distribution Neighborhood speaker selection unit (selection unit) 31 Neighborhood speaker table storage unit by distribution (storage unit)

Claims (13)

[Claims] 1. A speaker matching device for setting a background speaker for a true speaker and normalizing a matching score of the true speaker with respect to the input voice by a matching score of the background speaker with respect to the input voice. Means for selecting, from among the speaker acoustic models, speakers whose distances in the acoustic space with respect to the registered speaker's acoustic model are relatively short with respect to local features; and Means for holding as a nearby speaker of the speaker, and combining the local features of the acoustic model of the near speaker with the local features of each of the acoustic models of the true speaker to obtain the background speaker. A speaker verification device, comprising: means for synthesizing an acoustic model. 2. The speaker verification device according to claim 1, wherein each phoneme of the acoustic model is a hidden Markov model (HM).
M), and the collation score is HM for the input voice
A speaker verification device characterized by M likelihood. 3. The speaker verification device according to claim 1, wherein the local feature is a phoneme of an acoustic model of each speaker. 4. The speaker verification device according to claim 2, wherein the local feature is a state of each of the phoneme HMMs. 5. The speaker verification device according to claim 2, wherein the local feature is a distribution of each state of each of the phoneme HMMs. 6. The speaker verification apparatus according to claim 1, wherein the selecting unit has a relatively short distance in sound space between the phoneme and the registered speaker's acoustic model from among the acoustic models of each speaker. A speaker verification device for selecting a speaker. 7. The speaker verification device according to claim 1 or 5, wherein the selecting means selects, from among acoustic models of each speaker, an acoustic model of a registered speaker with respect to each state of each phoneme HMM. And a speaker whose distance in the acoustic space is relatively short is selected. 8. The speaker verification device according to claim 1, wherein the selecting unit selects, from among the acoustic models of each speaker, an acoustic model of a registered speaker with respect to a distribution of each state of each phoneme HMM. And a speaker whose distance in the acoustic space is relatively short is selected. 9. The speaker verification device according to claim 3, wherein the synthesizing unit calculates a phoneme of an acoustic model of the nearby speaker.
A speaker verification apparatus characterized in that an acoustic model of a background speaker is synthesized by combining acoustic models of an individual speaker in correspondence with respective phonemes. 10. The speaker verification device according to claim 4, wherein the synthesizing unit associates each state of each phoneme HMM of the neighboring speaker with each state of each phoneme HMM of the own speaker. A speaker verification apparatus characterized in that an acoustic model of a background speaker is synthesized by combining the speaker models. 11. The speaker verification apparatus according to claim 5, wherein said synthesizing means calculates a distribution of each state of each of said phoneme HMMs, and a distribution of each state of each of said phoneme HMMs of a true speaker. A speaker verification device characterized in that an acoustic model of a background speaker is synthesized by combining the speaker models with each other. 12. A background speaker is set for a true speaker,
In a speaker verification method for normalizing the likelihood of a true speaker with respect to an input voice by the likelihood of a background speaker with respect to the input voice, a registered speaker with respect to local features is selected from among acoustic models of the respective speakers. A first step of selecting a speaker whose distance in the acoustic space with the acoustic model is relatively short; a second step of holding the selected speaker as a nearby speaker of the registered speaker; A third process of synthesizing the acoustic model of the background speaker by combining the local features of the acoustic model of the neighboring speaker in correspondence with the local features of each of the acoustic models of the true speaker; A speaker verification method, comprising: 13. A background speaker is set for a true speaker,
In a speaker verification device that normalizes a verification score of a true speaker with respect to an input voice by a verification score of a background speaker with respect to the input voice, the speaker model of a registered speaker regarding a local feature is selected from among acoustic models of the respective speakers. Means for selecting a speaker whose distance in the acoustic space with the acoustic model is relatively short; means for holding the selected speaker as a nearby speaker of the registered speaker; and an acoustic model of the nearby speaker Means for synthesizing the acoustic model of the background speaker by combining the local features of the speaker corresponding to the local features of each of the acoustic models of the speaker. A computer-readable program medium carrying a computer program for operating a computer as each of the means in the device.