JP2005084613A - Speaker matching device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speaker matching device improved in the ratio (matching ratio) capable of determining whether a speaker is identical person or pretender. <P>SOLUTION: A matching pattern generation part 1 generates a matching pattern of continuous numbers by discarding and replacing a number with which it is difficult to decide whether the person himself or herself is a pretender with a new number on the basis of an evaluation result obtained by a number unit score evaluation part 10, and the presents the matching pattern to a user. A pattern dependency precision prediction/threshold setting part 9 predicts precision and sets a threshold based upon the numbers constituting the matching pattern. The number unit score evaluation part 10 evaluates similarities, found by the numbers between a speech that the user speaks and a previously registered speech, by precision evaluating functions by the numbers. A speaker decision part 8 decides a similarity to the matching pattern by using the threshold set by the pattern dependency precision prediction/threshold setting part 9. Replacement in number chain units is possible. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speaker verification device, and more particularly, to a text designation type speaker verification device using a continuous numeral pattern.

  As speaker verification techniques for determining whether a person is a person or an impersonator by voice, there are text-specific speaker verification using a verification pattern consisting of continuous numeric patterns and speaker verification using a plurality of verification patterns.

  In text-specific speaker verification using a verification pattern consisting of continuous numeric patterns, a verification pattern consisting of continuous digits is presented to the user, and the similarity between the voice uttered by the user and the pre-registered voice is calculated. Perform speaker verification. Since a different verification pattern is specified from the system side every time verification is performed, it is possible to prevent spoofing by the recorded voice of the person.

  In speaker verification using multiple verification patterns, if it is difficult to determine whether the person is a person or an impersonator with a single utterance, a different verification pattern is presented to encourage recurrence, and multiple utterances (more voices) Since the determination can be made using the information), it is possible to increase the ratio (collation rate) at which it is possible to correctly determine whether the person is a person or an impersonator.

  FIG. 8 is a block diagram of a conventional speaker verification device. In this speaker verification device, a verification pattern consisting of continuous numeric patterns is used, and the ratio (verification rate) for correctly determining whether the person is a person or an impersonator is increased. Therefore, a plurality of collation patterns are used. The collation pattern creation unit 1 creates a collation pattern composed of continuous numbers, presents it to the user, and prompts the user to utter the collation pattern. When the user utters (speech input) according to the presented collation pattern, the speech recognition unit 2 recognizes the utterance content using the speech recognition model 3.

  The recognition result collation unit 4 determines whether or not the recognition result in the voice recognition unit 4 matches the collation pattern. Here, if the user mistakenly inputs a voice that differs from the matching pattern, the recognition result is rejected, and even if the recorded person's voice is used to impersonate the voice, the recognition result matches the matching pattern. Since there is little to do, the recognition result is rejected.

  If the recognition result matches the collation pattern, the principal score calculation unit 5 calculates a score (similarity) between the utterance at the collation and the voice of the principal who has been registered as the registered speaker model 6 previously. In calculating the score, the background speaker model 7 is used to remove the influence of the background speaker. As representative methods of score calculation in the principal score calculation unit 5, there are an HMM (Hidden Markov Model) method, a GMM (Gaussian Mixture density Model) method, a DP matching method, and the like.

The speaker verification determination unit 8 determines the score obtained by the principal score calculation unit 5 using two kinds of threshold values for accepting the person and rejecting the other person. That is, if the score obtained by the principal score calculation unit 5 is S, the threshold for accepting the person is S _acc , and the threshold for rejecting the other person is S _rej , the user is accepted as the person if S ≧ S _acc. , S <S _{rej rejects} the user as not being the person, and S _rej ≦ S <S _acc. Instructs to present a verification pattern for verification. At the time of re-collation, the registered numbers are randomly connected to create a new collation pattern.

FIG. 9 shows the relationship between acceptance / rejection / _{relapse voice} in the determination using two types of threshold _values S _acc and S _rej , with the horizontal axis as score S and the vertical axis as score distribution (frequency). As can be seen from the figure, when S <S _rej , the user can be regarded as an impersonator, and when S ≧ S _acc , the user can be regarded as the person himself. Further, when S _rej ≦ S <S _acc , it is not possible to determine with high probability whether the user is the person or the impersonator.

Non-Patent Documents 1 to 4 below describe text-specific speaker verification and speaker verification using a plurality of verification patterns.
Furui Matsui: “Text-specific speaker recognition” Theory of Science D-2 J79-D-2 No.5 pp.647-656 (1996) Inner Kuroiwa Higuchi: “Study of speaker verification using numbers”, IEICE Tech., SP98-68, pp.1-8 (1998) Shimizu Kato “Improvement of speaker verification accuracy by pattern designation method for consecutive numbers”, sound lecture, autumn 2002, pp129-130 (2002) Shimizu Kato, “Improvement of speaker verification accuracy for time difference data by pattern specification method for continuous numbers”, sound lecture, spring 2003, pp105-106 (2003)

As described above, when the speaker verification determination unit in the conventional speaker verification device determines that S _rej ≦ S <S _acc , the verification pattern creation unit 1 is presented with a verification pattern for re-verification to the user. The collation pattern creation unit 1 again creates a continuous numeric pattern and prompts the user to make a second utterance. However, since the collation pattern created for the second utterance is obtained by simply concatenating numbers at random, the collation pattern is presented to the user for the second utterance and the third utterance. There is a problem that even if prompting and re-verification are performed, it cannot be expected that the ratio (verification rate) for correctly determining whether the person is the person or the impersonator will be sufficiently high.

  An object of the present invention is to automatically create a verification pattern that can be expected to have a sufficiently high ratio (matching rate) that can correctly determine whether the person is a person or a fraud according to the characteristics of the user's utterance. It is to improve the ratio (collation rate) that can be correctly determined.

  In order to solve the above-mentioned problem, the present invention presents a verification pattern consisting of continuous numbers to a user, calculates a similarity between a voice uttered by the user and a previously registered voice, and performs speaker verification. In the person verification device, a pattern-dependent accuracy prediction / threshold setting unit that performs accuracy prediction and threshold setting based on a number (or a number chain) constituting the verification pattern, and a voice uttered by the user and a previously registered voice Based on the evaluation result by the numerical unit score evaluation unit and the numerical unit score evaluation unit that evaluates the degree of similarity calculated for each number (or number chain) with the accuracy evaluation function for each number (or number chain) A collation pattern consisting of new consecutive numbers by replacing a difficult number (or number chain) with another number (or number chain) Using the threshold set by the collation pattern creation unit and the pattern-dependent accuracy prediction / threshold setting unit, the similarity obtained for the collation pattern is determined between the voice uttered by the user and the voice registered in advance. And a determination unit.

  Here, the pattern-dependent accuracy prediction / threshold setting unit performs accuracy prediction and threshold setting based on the score distribution of similarity between the person and the impersonator obtained in advance for each number (or number chain) constituting the verification pattern. Can be done.

  According to the present invention, the degree of similarity is obtained for each number (or number chain) between the voice uttered by the user and the voice registered in advance, and it is determined whether the person is the person or the impersonator from the evaluation result of the degree of similarity. Since a new matching pattern is created by replacing a difficult number (or number chain) with another number (or number chain), when the user makes a recurring voice, the person himself or herself is misrepresented by following the specified matching pattern. It can be expected that the ratio (collation rate) for correctly determining whether the person is a person will be sufficiently high.

  The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a speaker verification apparatus according to the present invention. The same or equivalent parts as those in FIG.

  FIG. 1 differs from FIG. 8 in that a pattern-dependent accuracy prediction / threshold setting unit 9 and a numerical unit score evaluation unit 10 are newly added, and the principal score calculation unit 5 obtains a score (similarity) for each number. The score is output to the numerical unit score evaluation unit 10, and the verification pattern generation unit 1 generates a verification pattern according to the evaluation result in the numerical unit score evaluation unit 10. Since the other parts are the same as those in FIG. 7, the collation pattern creation unit 1, the person score calculation unit 5, the pattern-dependent accuracy prediction / threshold setting unit 9, and the numerical unit score evaluation unit 10 will be mainly described below.

  The collation pattern creation unit 1 creates a collation pattern (continuous numeric pattern) to be uttered by the user. When prompting the user to recite, the collation pattern creation unit 1 has the identity or pretenderness obtained by the numerical unit score evaluation unit 10 described later. A new verification pattern is created by replacing a number that is difficult to determine whether the person is a person or an impersonator with another number based on the evaluation result.

  The principal score calculation unit 5 outputs the score obtained for each number to the number unit score evaluation unit 10. The score for each number can be obtained by matching for each number break using the registered speaker model 6 and the background speaker model 7, and the score for the matching pattern is different only in the unit of the break. is there.

  The pattern-dependent accuracy prediction / threshold setting unit 9 outputs, to the numerical unit score evaluation unit 10, the parameters of the score distribution of the person and the impersonator for each number constituting the verification pattern. The score distribution of the person and the impersonator for each number can be obtained in advance.

FIG. 2 is a characteristic diagram showing an example of the score distribution of the person and the impersonator who are tabulated for each number. μ _{i, d} and σ _{i, d} are the mean value and variance in the individual's score distribution, respectively, μ _{c, d} , σ _{c, d} are the mean value and variance in the spoofer's score distribution, and S _{EER, d} is a threshold value at which the equal error rate EER _d equals the error acceptance rate FAR _d and the error rejection rate FRR _d . These parameters (average value, variance) and the equal error rate and the threshold value for giving the equal error rate differ depending on the number d, and are therefore indicated with a subscript d.

When the score distributions of the person and the impersonator are approximated as normal distributions, the threshold S _{EER, d} and the equal error rate EER _d are given by the following equations (1) and (2).

Further, the pattern-dependent accuracy prediction / threshold setting unit 9 estimates the score distribution of the person and the impersonator regarding the collation pattern by combining the score distribution of the person and the impersonator for each number obtained in advance, and collates from the score distribution. A threshold value S _{EER, z} depending on the pattern and an equal error rate EER _z are estimated.

The numerical unit score evaluation unit 10 inputs a score for each number and parameters (μ _{c, d} , μ _{i, d} , σ _{c, d} , σ _{i, d} ) for each number, and forms a collation pattern For each number you do, evaluate your identity or your impersonator. As described above, this evaluation result is used when the collation pattern creation unit 1 creates a collation pattern.

The speaker verification determination unit 8 performs determination using two types of threshold _values S _acc and S _reg for accepting the person and rejecting the other person. The threshold _values S _acc and S _reg are also determined by the numbers used in the verification pattern. Different. The thresholds S _acc and S _rej are set according to the score distribution related to the collation pattern estimated by the pattern-dependent accuracy prediction / threshold setting unit 9.

  Next, specific configurations of the collation pattern creation unit 1, the pattern-dependent accuracy prediction / threshold setting unit 9, and the numerical unit score evaluation unit 10 will be described.

  FIG. 3 is a block diagram illustrating an example of the collation pattern creation unit 1. The collation pattern creation unit 1 receives an evaluation result of the individuality for each number from the numerical unit score evaluation unit 10. The number (chain) sorting selection unit 11 instructs the collation pattern generation unit 12 to discard a number whose evaluation result is close to 0.5 and replace it with another new number. The verification pattern generation unit 12 picks up a new number from the registered utterance pattern in accordance with this instruction and adds it to the remaining number as a verification pattern at the time of re-verification. As the numbers to be picked up, any number other than the discarded numbers can be selected in any order.

  FIG. 4 is an explanatory diagram of an example of the replacement of numbers (chains) in the number (chain) sorting selection unit 11. First, the collation pattern creation unit 1 presents the collation pattern of the first utterance, for example, “1 2 3 4” to the user. The user inputs voice according to the verification pattern.

The score S for the collation pattern of the first utterance is S _rej ≦ S <S _acc (see FIG. 9), and the individuality evaluation result for each number obtained by the numerical unit score evaluation unit 10 is 0.45, If it is 0.55, 0.65, or 0.60, the number (chain) selection selection unit 11 has difficulty in discriminating whether the identity evaluation result is a value close to 0.50, that is, the person or the impersonator, as the collation pattern at the time of re-collation. Discard the numbers and replace them with new numbers to create consecutive numbers.

  For example, “1 2” having an evaluation result of authenticity of 0.45 and 0.55 is discarded, and “5 6” is newly added to create a collation pattern “3 4 5 6” of the second voice. Therefore, the verification pattern creation unit 1 presents “3 4 5 6” to the user as the verification pattern of the second utterance. The same applies to recurrent voices after the third utterance.

FIG. 5 is a block diagram illustrating an example of the pattern-dependent accuracy prediction / threshold setting unit 9. The pattern-dependent accuracy prediction / threshold setting unit 9 receives the collation pattern created by the collation pattern creation unit 1. The person / spoofer score distribution prediction unit 91 uses the parameters (μ _{c, d} , μ _{i, d} , σ _{c, d} , σ _{i, d} ) 92 for each number obtained in advance to identify the person regarding the matching pattern. As well as the score distribution of fraudsters. The accuracy prediction / threshold setting unit 93 obtains a threshold for the collation pattern based on the score distribution of the person and the impersonator regarding the collation pattern estimated by the person / spoofer score distribution prediction unit 91, and determines the threshold for speaker collation. Output to unit 8. Although not shown in the drawing, the pattern-dependent accuracy prediction / threshold setting unit 9 sets the parameters (μ _{c, d} , μ _{i, d} , σ _{c, d} , σ _{i, d} ) of the person and the spoofer for each number. Is output to the numerical unit score evaluation unit 10.

The score distribution relating to the verification pattern is obtained from the reproducibility of the normal distribution. For example, when the collation pattern z is composed of four types of variables s, t, u, and v (four-digit numbers), the score distribution s _z regarding the collation pattern _z is the components s _s , s _t , s _u , s _v is approximated by the arithmetic mean of, obtained by the formula described below (3).

At this time, the average value mu _{c, z} and variance _{V c} of score distribution _{s z} of the person about the collation pattern _{z, z} is represented by the following formula (4) obtained in (5).

Further, the average value μ _{i, z} and the variance V _{i, z} of the score distribution s _z of the impersonator regarding the collation pattern z are obtained by the following formulas (6) and (7).

Using the above variables, the equal error rate EER _z and the threshold S _{EER, z} that gives the equal error rate in the score distribution s _z regarding the collation pattern are given by the following equations (8) and (9).

FIG. 6 is a block diagram illustrating an example of the numerical unit score evaluation unit 10. The numerical unit score evaluation unit 10 calculates the score S for each number from the principal score calculation unit 5 and the score distribution parameters (μ _{c, d} , μ _{i, d)} for each number from the pattern-dependent accuracy prediction / threshold setting unit 9. , Σ _{c, d} , σ _{i, d} ) 92 are input, and the identity or the impersonator is evaluated for each number constituting the collation pattern.

Since the range of distribution (average value and variance) of the score S differs for each number, an evaluation scale of personality that normalizes the variation in the range of score distribution (average value and variance) for each number is determined as follows.
When the probability density function of the normal distribution with the mean value μ and the variance σ ^{2 is represented by} the following equation (10), a new evaluation measure Q _d of personality can be defined as the following equation (11).

Here, Q _d satisfies 0 <Q _d <1, and the closer to 0, the more impersonator, the closer to 1, the more likely the person is. In addition, it can be said that it is difficult to determine whether the person is a person or an impersonator near 0.5. The numerical unit score evaluation unit 10 calculates Q _d for each number and passes the result to the collation pattern creation unit 1.

  In the embodiment described above, the identity is evaluated for each number, and a new collation pattern is created by replacement in units of numbers. However, for example, a score is obtained in units of a two-digit number chain, It is also possible to evaluate the identity based on the score distribution in the numerical chain unit calculated from the obtained score distribution in the numerical unit, and create a new matching pattern by replacing the numerical chain unit. it can.

  Further, in the above embodiment, the identity is evaluated by the expression (11). However, the identity can be evaluated by another evaluation function, and the number (or the number) obtained by the principal score calculation unit 5 can be evaluated. It is also possible to use the score for each chain).

7, the conventional method (method of using a second shot th matching pattern does not depend on Q _d) and the present invention (method of using a second shot th matching pattern depending on the Q _d) and de experiments conducted speaker verification The results are shown in comparison.

Here, mobile phone voices of 81 men and 74 women were used. In addition, the threshold S _reg for accepting the principal and the threshold S _acc for _rejecting the other person are set so that the false rejection rate FRR = 1.0% and the false acceptance rate FAR = 1.0%, respectively, and a 6-digit number verification pattern is used. Numbers were exchanged with 2 digits stored in 4 digits. n _{a, (k)} , _{ni, (k)} are the number of occurrences of the person, the impersonator, the k-th occurrence, and na _{, (k)} , n _{r (k)} are the persons who are accepted in the k-th utterance, rejected The number of occurrences of fraudsters.

  Figure 7 shows that 11.5% of the person who made a recurrence without being judged by the first utterance and 12.0% of the spoofer were correctly judged by the second utterance. In contrast to 3.6%, in the present invention, it is 6.5% for the person and 4.3% for the fraud.

  In the present invention, by creating a collation pattern in consideration of the identity of the person, it is possible to improve the ratio (collation rate) that can correctly determine whether the person is a person or an impersonator. Therefore, the voice for telephone services such as telephone banking The present invention can be applied to an authentication system, a voice authentication system on the Internet or a mobile Internet, and the like.

It is a block diagram which shows one Embodiment of the speaker collation apparatus which concerns on this invention. It is a characteristic view which shows an example of the score distribution of the person who is totaled for every number, and an impersonator. It is a block diagram which shows an example of a collation pattern creation part. It is explanatory drawing of an example of replacement | exchange of the number (chain) in this invention. It is a block diagram which shows an example of a pattern dependence precision prediction and threshold value setting part. It is a block diagram which shows an example of a numerical unit score evaluation part. It is a figure which compares and shows the experimental result which performed speaker collation according to the conventional system and this invention. It is a block diagram of the conventional speaker verification apparatus. It is a characteristic view which shows the relationship of acceptance / rejection / recurrence voice in the determination using two types of threshold values.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Collation pattern creation part, 2 ... Speech recognition part, 3 ... Speech recognition model, 4 ... Recognition result collation part, 5 ... Personal score calculation part, 6 ... Registered speaker Model: 7 ... Background speaker model, 8 ... Speaker collation determination unit, 9 ... Pattern dependence / accuracy prediction threshold setting unit, 10 ... Number unit score evaluation unit, 11 ... Number ( Chain) selection selection unit, 12 ... collation pattern generation unit, 13 ... registered utterance pattern, 91 ... person / spoofer score distribution estimation unit, 92 ... parameter of score distribution for each number, 93 ..Accuracy prediction / threshold setting section

Claims (3)

In a speaker verification device that presents a verification pattern consisting of continuous numbers to a user, calculates a similarity between a voice uttered by the user and a previously registered voice, and performs speaker verification,
A pattern-dependent accuracy prediction / threshold setting unit for performing accuracy prediction and threshold setting based on the numbers constituting the verification pattern;
A numerical unit score evaluation unit that evaluates the degree of similarity calculated for each number between the voice uttered by the user and the previously registered voice, with an accuracy evaluation function for each number;
Based on the evaluation result by the numerical unit score evaluation unit, a collation pattern creation unit that creates a collation pattern composed of new continuous numbers by replacing a number that is difficult to identify whether the person or the impersonator with another number;
A determination unit that determines the similarity obtained for the matching pattern between the voice uttered by the user and the voice registered in advance using the threshold set by the pattern-dependent accuracy prediction / threshold setting unit; A speaker verification device characterized by that. In a speaker verification device that presents a verification pattern consisting of continuous numbers to a user, calculates a similarity between a voice uttered by the user and a previously registered voice, and performs speaker verification,
A pattern-dependent accuracy prediction / threshold setting unit that performs accuracy prediction and threshold setting based on a number chain that constitutes a matching pattern;
A numerical unit score evaluation unit that evaluates the degree of similarity obtained for each number chain between the voice uttered by the user and the previously registered voice with an accuracy evaluation function for each number chain;
Based on the evaluation result by the numerical unit score evaluation unit, a collation pattern creation unit that creates a collation pattern composed of new continuous numbers by replacing a number chain that is difficult to determine whether the person or the impersonator with another number chain, and
A determination unit that determines the similarity obtained for the matching pattern between the voice uttered by the user and the voice registered in advance using the threshold set by the pattern-dependent accuracy prediction / threshold setting unit; A speaker verification device characterized by that. The pattern-dependent accuracy prediction / threshold setting unit is configured to perform accuracy prediction and threshold setting based on a score distribution of similarity between a person and an impersonator obtained in advance for each number or number chain constituting a collation pattern. The speaker verification device according to claim 2 or 3.

Images