JP2004279770A - Speaker authentication device and discriminant function setting method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speaker authentication technique ensuring preset authentication accuracy. <P>SOLUTION: A pattern specifying unit specifies mutually different speech patterns for speaker authentication a plurality of up to (n) times. A score calculator 7a calculates scores of a speaker's speech corresponding to (i)th speech pattern specification and a speaker model as (i)th scores. A set including as elements (k) scores corresponding to speech pattern specifications of a plurality of (k) times is regarded as total scores. A total speaker decision unit 7d regards the total scores as an input to a discriminant function of a (k)-dimensional plane and decides whether the speaker is the identical person or not from the sign of the discriminant function. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speaker authentication device that determines whether a speaker is a person or another person based on an input voice of the speaker, and a method of setting a discriminant function used for the device. As the input voice, for example, a plurality of utterances of a continuous word in which a plurality of symbols or words such as numbers and alphabets are continuous are used.
[0002]
[Prior art]
As a first conventional example of the speaker authentication technology, for example, an apparatus described in Japanese Patent Application Laid-Open No. 9-244684 is known. This speaker authentication technology is to authenticate a user based on the input voice of a speaker, and a speaker who has already registered a speech is obtained from feature amounts registered in advance and feature amounts extracted from the input voice. It is determined whether the person is the person or another person (especially an impostor). As the input voice of the speaker, a specific known keyword such as a personal ID and a password is used.
[0003]
However, in the first conventional example described above, in order to use a known keyword such as a personal ID password as an input voice at the time of authentication, the voice of the registered speaker is recorded in advance, for example, at the time of authentication, and the reproduced voice is recorded. Is used by a third party, so that there is a problem that the person is liable to be spoofed.
[0004]
To cope with this problem, there is a method in which the user utters a large number of keywords at the time of authentication and the number of keywords used as input voice is increased, but this method increases the burden on the user. Further, when the number of keywords used as the input voice is increased, there is a problem that the authentication time becomes longer and the authentication efficiency is reduced.
[0005]
On the other hand, as a second conventional example of the speaker authentication technology, a method described in JP-A-2000-99090 is known. In this speaker authentication technology, a plurality of symbols (numerals or alphabets) are uttered for each specific speaker during speaker registration, and a model of the specific speaker (speaker model) is registered. By specifying an arbitrary symbol string in the plurality of symbols from the side and causing the speaker to utter, the registered speaker model and the input voice determine whether or not the speaker is himself (registered speaker). Is determined. According to this, even if a third party records the voice of the registered speaker at the time of authentication, there is an advantage that it is difficult for the third party to spoof the person because the reproduced voice does not utter the specified symbol string. .
[0006]
However, the above-mentioned second conventional example has room for improvement in the authentication efficiency.
[0007]
[Patent Document 1]
JP-A-9-244684
[Patent Document 2]
JP 2000-99090 A
[0008]
[Problems to be solved by the invention]
The present invention has been made in view of such circumstances, and a purpose of the present invention is to prevent false identification of a person, to guarantee a predetermined authentication accuracy, and to efficiently perform a speaker authentication technique. Is to provide.
[0009]
[Means for Solving the Problems]
A first invention is a speaker authentication device for solving the above-mentioned problem, which is a speaker authentication device that performs personal authentication based on a score indicating a similarity between a utterance of a speaker and a speaker model, and is different from each other. Utterance pattern specifying means for specifying a speaker authentication utterance pattern a plurality of times at a maximum of n times, and a score between the utterance of the speaker corresponding to the i-th utterance pattern specification and the speaker model, As an i-th score, a score calculating means for calculating, and a set including k scores corresponding to a plurality of k utterance pattern designations as an integrated score, and inputting the integrated score into a k-dimensional plane discriminant function The present invention is characterized in that an integrated speaker determining means for determining whether the speaker is the person or another person from the sign of the discriminant function is provided.
[0010]
According to a second aspect of the present invention, in the first aspect of the present invention, the first aspect further comprises a speaker determination unit based on a single utterance to determine whether the speaker is the person or another person by comparing the first score with a threshold value alone. Is characterized in that the speaker determination unit determines whether the speaker is a person or another person when the speaker determination unit based on a single utterance cannot determine whether the speaker is the person or another person with a predetermined authentication accuracy. Speaker authentication device.
[0011]
A third invention is the speaker authentication device according to the second invention, wherein the integrated score is a set having at least the first score as an element.
[0012]
According to a fourth aspect of the present invention, in the first aspect of the present invention, the integrated speaker determining means comprises a single utterance determining means for individually comparing the n-th score with a threshold to determine whether the utterer is the person or another person. Is characterized in that the speaker determination unit determines whether the speaker is a person or another person when the speaker determination unit based on a single utterance cannot determine whether the speaker is the person or another person with a predetermined authentication accuracy. Speaker authentication device.
[0013]
A fifth invention is the speaker authentication device according to the fourth invention, wherein the integrated score is a set having at least the (n−1) th score and the nth score as elements.
[0014]
A sixth invention is the speaker according to the first invention, wherein n = 3, k = 2, and the first score alone is compared with a threshold value to determine whether the speaker is the person or the other person. The integrated speaker determination means may determine whether the speaker is a person or another person with a predetermined authentication accuracy by comparing the first score and the threshold value by the speaker determination means based on a single utterance. If it is not possible, the first integrated score, which is a set including the first score and the second score as an element, is used as an input of a discriminant function to determine whether the speaker is a person or another person. If it is not possible to determine whether the speaker is the person or the person with the authentication accuracy, the second integrated score, which is a set having the second score and the third score as elements, is used as an input of the discriminant function to determine whether the speaker is the person. To judge someone else It is a speaker authentication device according to claim.
[0015]
In a seventh aspect based on the first aspect, the utterance pattern designating means designates the i-th specified speaker authentication utterance pattern having a higher authentication accuracy than the (i-1) -th specified speaker authentication pattern. A speaker authentication device characterized by a pattern.
[0016]
In an eighth aspect based on the first aspect, the utterance pattern storage means for storing the speaker registration utterance pattern used for speaker registration is provided, and the utterance pattern designation means is provided as a speaker authentication utterance pattern. A speaker authentication device characterized in that the speaker authentication device is an utterance pattern including at least one set of a chain of a plurality of words included in the speaker registration utterance pattern.
[0017]
In a ninth aspect based on the first aspect, the discriminant function includes a first discriminant function for accepting the person and a second discriminant function for rejecting another person, and the first discriminant function and the second discriminant function. Is a speaker authentication device characterized in that the false acceptance rate and the false rejection rate are set to be constant values.
[0018]
A tenth invention is a method for setting a discriminant function used in the speaker authentication device according to the ninth invention for solving the above-mentioned problem, wherein a distribution of k scores is approximated by a k-dimensional normal distribution on a k-dimensional plane. Then, determining a discriminant plane on a k-dimensional plane, moving the discriminant plane in parallel, and setting a first discriminant function and a second discriminant function in which the false acceptance rate and the false rejection rate are constant values are set. Features.
[0019]
BEST MODE FOR CARRYING OUT THE INVENTION
According to the present invention, in a text designation type speaker authentication technique using continuous word utterance, a speaker is made to utter a plurality of different continuous words, and a plurality of scores obtained for each utterance are used in an integrated manner. Is a person or another person. At this time, it does not prevent the determination of whether the speaker is the person or another person using the score of each utterance alone. Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 shows a configuration of a speaker authentication device according to an embodiment of the present invention. In FIG. 1, input / output signals of the speaker verification device 1 are digital signals, and the speaker verification device 1 is realized by digital processing. Therefore, when the input / output is an analog signal, the signal is converted by an analog / digital converter (AD converter) and connected to the speaker verification device 1. Further, as an input audio signal to the speaker verification device 1, an input from a telephone line, a microphone, or the like can be used. In addition, a voice signal can be output from the speaker verification device 1 to a telephone line, a speaker, or the like.
[0020]
In the speaker verification device 1 of FIG. 1, a feature extractor 2 extracts an acoustic feature amount from an input voice signal A of a user. The utterance content determiner 3 determines whether or not the utterance content of the speaker matches the utterance content indicated by the utterance prompt C stored in the buffer 10. The switch 4 switches the connection destination of the feature extractor 2 between the time of speaker registration and the time of speaker authentication. In detail, the switch 4 connects the feature extractor 2 to the registration determiner 5 at the time of speaker registration and to the speaker determiner 7 at the time of person authentication.
[0021]
The registration determiner 5 determines whether or not the feature amount extracted by the feature extractor 2 at the time of speaker registration is sufficient for speaker registration. Further, a speaker model is generated by learning based on the feature amount. For example, it is determined that a sufficient number of voice inputs, such as three times, is sufficient as a condition. Alternatively, the model is evaluated while generating the speaker model, and it is determined that the speaker model is sufficient when the quality reaches a certain level. As the speaker model, HMM (Hidden Markov Model) and GMM (Gaussian Mixture Model) can be used.
[0022]
The voice feature information storage unit 6 stores and stores the speaker model generated by the registration determination unit 5.
[0023]
At the time of speaker authentication, the speaker determiner 7 provides a score indicating the similarity between the utterance of the speaker and the speaker model, in other words, the feature amount extracted by the feature extractor 2 and the speaker feature information storage 6. Calculates a score indicating the degree of similarity with the speaker model, and basically uses multiple scores calculated corresponding to multiple utterances in an integrated manner to determine whether the speaker is the person or the impostor. This determines whether or not speaker authentication is possible (other person) and outputs the speaker determination result B. As shown in FIG. 2, for example, the speaker determination unit 7 includes a score storage unit 7a (score storage unit), a score calculator 7b (score calculation unit), and a single-utterance speaker determination unit 7c (single-utterance-based speech). Speaker determination means) and an integrated speaker determination unit 7d (integrated speaker determination means).
[0024]
The score accumulator 7a accumulates and stores the score used for the judgment at the time when it is not possible to judge whether the speaker is the person or another person with a predetermined authentication accuracy.
[0025]
The score calculator 7b calculates the score between the speaker's utterance and the speaker model for each designation of the speaker authentication utterance pattern (output of the utterance prompt C) during speaker authentication.
[0026]
Here, the score between the utterance (i-th utterance) of the speaker and the speaker model corresponding to the i-th utterance pattern designation (output of the i-th utterance prompt C) is referred to as an i-th score.
[0027]
The single-speaker speaker determination unit 7c performs speaker determination by single-speaking at the time of speaker authentication. By comparing an arbitrary i-th score with a threshold value alone, it is determined whether the speaker is the true speaker based on the comparison result. It is determined whether or not the speaker can be authenticated as an impostor.
[0028]
The integrated speaker determiner 7d extracts k scores corresponding to a plurality of k times of pattern designation from the score calculated by the score calculator 7a at the time of speaker authentication from the score accumulator 7a and integrates them. Is used to determine whether or not the speaker is authenticated, such as whether the speaker is the person or the impostor. More specifically, a set having k scores as elements is defined as an integrated score, and the integrated score is input to a discriminant function, and whether or not speaker authentication is possible is determined from the sign of the discriminant function. Here, k is a natural number of 2 or more, and is set to be equal to or less than a plurality n described later. The discriminant function will be described later with reference to FIG.
[0029]
The pattern designator 8 (speech pattern designating means) includes a speaker registration speech pattern (hereinafter referred to as a speaker registration pattern) and a speaker authentication speech pattern (hereinafter referred to as a speaker authentication pattern). Then, an utterance prompt C for prompting the user to utter these designated patterns is generated and output. The pattern designator 8 can designate different speaker authentication patterns for the same speaker a plurality of times at most n times. Therefore, at the time of speaker authentication, different occurrence prompts C are output to the same speaker a plurality of times n at a maximum. Here, n is a natural number equal to or greater than k.
[0030]
The speaker registration pattern storage unit 9 (speaking pattern storage unit) stores and saves the speaker registration pattern specified by the pattern specifying unit 8.
[0031]
The buffer 10 stores the utterance prompts C output from the pattern designator 8 sequentially in an overwriting manner.
[0032]
[Speaker registration process]
Next, the operation of the speaker verification device 1 during speaker registration will be described with reference to FIG. FIG. 3 is a flowchart showing the flow of the speaker registration process performed by the speaker verification device 1 shown in FIG.
[0033]
First, user authentication using the speaker's personal ID and password is performed, and when the user authentication is accepted, the speaker registration process of FIG. 3 is started.
[0034]
In FIG. 3, first, the pattern designator 8 designates a speaker registration pattern composed of a series of a plurality of words (numbers, alphabets, and other words), and stores the pattern in the speaker registration pattern storage 9. Save (step ST1). For example, a six-digit continuous number (“648790” in this example) as shown at the top of FIG. 4 is used as the speaker registration pattern. Next, the pattern designator 8 generates and outputs an utterance prompt C prompting the utterance of the specified pattern (step ST2). The utterance prompt C is stored in the buffer 10 and output from the speaker verification device 1.
[0035]
Next, when the speaker utters the speaker registration pattern specified by the utterance prompt C and the input voice signal A is input, the feature extractor 2 extracts an acoustic feature from the input voice signal A (step). ST3, ST4). Next, the utterance content determiner 3 determines whether the utterance content of the speaker matches the utterance content indicated by the utterance prompt C stored in the buffer 10 (step ST5). Next, the registration determination unit 5 determines whether or not the feature amount is sufficient to generate a good speaker model (step ST6). If the feature amount is insufficient or if the determination result in step ST5 does not match, the registration determination unit 5 determines whether the feature amount is sufficient. , And instructs the pattern designator 8 to output the utterance prompt C again. This returns to step ST2.
[0036]
On the other hand, if the determination result in step ST5 is a match and a sufficient feature amount is obtained, the process proceeds to step ST7, where a speaker model is generated and stored in the voice feature information storage 6. Thereby, speaker registration is completed.
[0037]
Note that it is preferable to configure the speaker registration pattern so as to include a large number of vowels or nasal sounds, because acoustic characteristics unique to the speaker can be easily extracted and the characteristics can be efficiently extracted with a small number of utterances.
[0038]
[Speaker authentication process]
Next, an example of the operation of the speaker verification device 1 at the time of speaker authentication will be described with reference to FIGS. FIG. 4 is a diagram showing an example of specifying an utterance pattern, FIGS. 5 to 8, and a flowchart showing a flow of speaker authentication processing performed by the speaker verification device 1 shown in FIG. 1, and FIG. It is a figure showing an example of a setting method of a function.
[0039]
Here, as an example of the embodiment of the present invention, it is assumed that n = 3 and k = 2 for convenience. For convenience, it is assumed that the single-speaking speaker determination unit 7c compares only the first score S1 alone with a threshold value to determine whether the speaker is the person himself or an impostor (another person). Further, in the integrated speaker determination unit 7d, in the speaker determination based on the single utterance based on the comparison between the first score S1 and the threshold value by the single utterance speaker determination unit 7c, the speaker is the identity or the impostor with the predetermined authentication accuracy If it is not possible to determine whether the first score S1 and the second score S2 are elements, the first integrated score is used as an input of the discriminant function, and the first integrated score is used as the input of the discriminant function. If the speaker is the identity or the impostor from the positive or negative sign (first integrated speaker determination), and if it is still not possible to determine whether the speaker is the identity or the impostor with the predetermined authentication accuracy, , Using the second integrated score, which is a set having the second score S2 and the third score S3 as elements, using the second integrated score as an input of the discriminant function, Judgment as impostor (second time) Shall be made an integrated speaker determination).
[0040]
Further, in the present embodiment, the pattern designator 8 designates a pattern whose authentication accuracy is gradually increased each time the speaker is prompted to speak by designating a speaker authentication pattern. In the case of n = 3, for example, for one utterance, one containing at least one chain of a plurality of words included in the speaker registration pattern, for example, including two chains of a plurality of words and the two groups being discontinuous Are selected to form a speaker authentication pattern, and this is designated as the first utterance pattern. Then, for the second utterance, another one set of word chains is added to the first word chain, and the number of sets of a plurality of word chains is increased from the first utterance pattern. A pattern is formed, and this is designated as a second utterance pattern. In the speaker authentication pattern for the second utterance, the authentication accuracy is higher than that for the first utterance pattern because of the large number of pairs of a plurality of word chains. Then, for the third utterance, a speaker authentication pattern is configured to include four consecutive words among the words constituting the speaker registration pattern, and this is designated as a third utterance pattern. I do. The pattern for speaker authentication for the third utterance has a smaller number of words than the pattern for the second utterance, but has a higher authentication accuracy than the pattern for the second utterance because of the large number of series of word chains. Become. In this manner, for the first utterance, the authentication accuracy is not so high by designating the utterance pattern so as to include at least one chain of a plurality of words included in the speaker registration pattern. And the number of words to be uttered is suppressed to reduce the burden on the speaker. For the second utterance, by specifying the utterance pattern by increasing the number of sets of multiple word chains than the first utterance pattern, the number of uttered words increases, but the effect of preventing identity spoofing is maintained while the authentication accuracy is maintained. Enhance. Then, for the third utterance, the authentication accuracy is preferentially increased by increasing the number of a series of word chains and specifying an utterance pattern. In addition, by increasing the number of a series of word chains, the effect of preventing false identification is reduced, but at the time of the third utterance, since the first and second authentication rejection judgments have been passed, there is a possibility of false identification. Is small and there is no problem.
[0041]
FIG. 4 shows a specific example of the utterance pattern designation when n = 3. In the example of FIG. 4, as shown in the second row from the top, a speaker registration pattern (“636790”) composed of six consecutive digits exemplified in the speaker registration process for the first utterance And two non-consecutive pairs (in this example, “79” and “63”) are selected, and a speaker authentication pattern (“7963”) is selected. And a four-digit number (“7963”) in which two sets of two-digit numbers are stored is designated as the first utterance pattern.
[0042]
When the speaker registration pattern is a six-digit continuous number “648790”, a set of a chain of a plurality of words, for example, a two-digit number, is “63”, “38”, “87”, “79”, “90”. There are five sets, and if, for example, two sets are selected at random from these sets, the total number of patterns of four-digit numbers in which two sets of two-digit numbers are stored is 205. For example, “7963” is designated as the first speaker authentication pattern among these.
[0043]
For the second utterance, as shown in the third row from the top in FIG. 4, two sets (“79” and “63”) in the first utterance pattern specification are further added to another set (this example). In this example, "38") is added to form a speaker authentication pattern ("387963") in which the number of sets of multiple word chains is increased from that of the first utterance pattern, and three sets of two-digit numbers are stored. The designated six-digit number ("387963") is designated as the second utterance pattern. When the speaker registration pattern is a 6-digit continuous number “638790”, the total number of patterns of 6-digit numbers in which three sets of 2-digit numbers are stored is 1025. For example, “387963” is designated as the user authentication pattern.
[0044]
For the third utterance, as shown at the bottom of FIG. 4, the speaker authentication pattern (“648790”) includes four consecutive numbers among the numbers that constitute the speaker registration pattern (“648790”). In this example, “8790”) is configured, and the four-digit number (“8790”) storing the four-digit number chain is designated as the third utterance pattern. When the speaker registration pattern is a six-digit continuous number “648790”, the total number of patterns of the four-digit number storing the four-digit number chain is three. For example, “8790” is specified as the pattern.
[0045]
[Speaker judgment by single utterance]
First, user authentication using the speaker's personal ID and password is performed, and when the user authentication is accepted, speaker authentication processing is started.
[0046]
In FIG. 5, first, the pattern designator 8 reads out the speaker registration pattern stored in the speaker registration pattern accumulator 9 and, based on the speaker registration pattern, the first speaker authentication pattern. A pattern, for example, "7963" is designated (step ST11). This designation includes at least one chain of a plurality of words included in the speaker registration pattern.
[0047]
As described above, by configuring the speaker authentication pattern so as to include at least one set of a plurality of word chains included in the speaker registration pattern, it is possible to prevent false identification by recording and reproducing the registered speaker's voice. The authentication accuracy can be kept higher than the randomness.
[0048]
Next, the pattern designator 8 generates an utterance prompt C prompting the utterance of the speaker authentication pattern (for example, “7963”) specified first, and outputs the utterance prompt C from the speaker authentication device 1 (step ST12). The utterance prompt C is stored in the buffer 10 by overwriting. Next, when the speaker utters a designated speaker authentication pattern (for example, “7963”) and an input audio signal A is input, the feature extractor 2 extracts an acoustic feature amount from the input audio signal A. (Steps ST13 and ST14). Next, the utterance content determiner 3 determines whether the utterance content of the speaker matches the utterance content indicated by the utterance prompt C stored in the buffer 10. If the result of this determination is that they do not match, the pattern designator 8 outputs the same utterance prompt C again (step ST15). Next, the score calculator 7b refers to the speaker model of the speech feature information storage 6, and refers to a score (first score) indicating the degree of similarity between the feature extracted by the feature extractor 2 and the feature of the speaker model. S1) is calculated (step ST16).
[0049]
Next, the single-speaking speaker determiner 7c compares the first score S1 with a predetermined authentication acceptance determination threshold Th1 (step ST17), and the first score S1 is the authentication acceptance determination threshold. If the value is larger than the value Th1, the speaker is determined to be the person himself / herself, and a determination result B for accepting the authentication is output (step ST19). The threshold value Th1 for authentication acceptance determination is set in the speaker determination device 7 in advance. As the threshold value Th1, a value at which the false acceptance rate becomes several%, for example, 1%, for example, in an experiment, is set so that the probability of falsely accepting the impersonator (other) (authorization rate) is reduced. adopt.
[0050]
On the other hand, as a result of the comparison in step ST17, if the first score S1 is not larger than the authentication acceptance determination threshold value Th1, the first score S1 is compared with a predetermined authentication rejection determination threshold value Th2. If the first score S1 is smaller than the authentication rejection determination threshold Th2, the speaker is determined to be an impostor (another), and a determination result B for rejecting the authentication is output (step S1). ST20). The threshold Th2 for authentication rejection determination is set in the speaker determiner 7 in advance. As the threshold value Th2, for example, a value at which the rejection rate becomes several percent, for example, 1%, of the false rejection rate in the experiment so as to reduce the probability of falsely rejecting the user (rejection rate). adopt.
[0051]
As described above, the error rate when the thresholds Th1 and Th2 are set so that the false rejection rate and the false acceptance rate are equal is known as an equal error rate.
[0052]
On the other hand, as a result of the comparison in step ST18, when the first score S1 is between the authentication acceptance determination threshold value Th1 and the authentication rejection determination threshold value Th2, the single-speaking speaker determination unit 7c sets It is determined that it is not possible to determine whether the authentication is accepted or rejected with the predetermined authentication accuracy, and the first score S1 is stored and stored in the score storage 7a (step ST21). An instruction is issued to update the speaker authentication pattern to a different one, for example, a speaker authentication pattern that increases the authentication accuracy, and to output the next utterance prompt C. Thus, the process returns to step ST11, and the integrated speaker determination is performed.
[0053]
[Integrated speaker judgment: 1st]
When the process returns from step ST21 to step ST11 in the speaker determination based on a single utterance, that is, when the speaker determination unit 7c for a single utterance cannot determine the identity or the other person with the predetermined authentication accuracy, the integrated speaker determination is performed. The integrated speaker determination is performed by the device 7d. This processing will be described with reference to FIGS.
[0054]
In FIG. 5, first, the pattern designator 8 reads the speaker registration pattern stored in the speaker registration pattern storage 9, and based on the speaker registration pattern, the second speaker authentication A pattern, for example, "387963" is designated (step ST11). This pattern specification has higher authentication accuracy than the first speaker authentication pattern.
[0055]
Next, the pattern designator 8 generates an utterance prompt C that prompts the utterance of a speaker authentication pattern (for example, “387963”) to be specified secondly, and outputs the utterance prompt C from the speaker authentication device 1 (step ST12). The utterance prompt C is stored in the buffer 10 by overwriting. Next, when the speaker utters the designated pattern ("387963") and the input audio signal A is input, the feature extractor 2 extracts an acoustic feature amount from the input audio signal A (steps ST13 and ST14). Next, the utterance content determiner 3 determines whether the utterance content of the speaker matches the utterance content indicated by the utterance prompt C stored in the buffer 10. If the result of this determination is that they do not match, the pattern designator 8 outputs the same utterance prompt C again (step ST15). Next, the score calculator 7b refers to the speaker model of the speech feature information storage 6, and refers to a score (second score) indicating the similarity between the feature extracted by the feature extractor 2 and the feature of the speaker model. S2) is calculated (step ST16).
[0056]
Next, the first integrated speaker determination is performed by the integrated speaker determination unit 7d using an integrated score (first integrated score) that is a set including the first score S1 and the second score S2 as elements.
[0057]
However, the processing from steps ST17 to ST21 in FIG. 5 is to be read as the processing from steps ST17 to ST21 shown in FIG.
[0058]
That is, as shown in FIG. 6, the integrated speaker determination unit 7d reads the first score S1 from the score storage unit 7a, and uses the first score S1 and the second score S2 calculated by the score calculator 7b as elements. The set is defined as an integrated score (S1, S2), and the integrated score is defined as a predetermined authentication acceptance determination discriminant function f on a two-dimensional plane. ₁ (First discriminant function: details will be described later with reference to FIG. 9), and the resulting authentication acceptance judgment discriminant function f ₁ Is compared with 0 (zero: threshold) (step ST17). Then, the discrimination function f for authentication acceptance determination ₁ Is positive (plus), it is determined that the speaker is the person himself, and a determination result B for accepting the authentication is output (step ST19). Discrimination function f for authentication acceptance determination ₁ And the threshold value 0 are set in the speaker determination unit 7 in advance. This discriminant function f ₁ In order to reduce the false acceptance rate, for example, a value at which the false acceptance rate becomes several percent, for example, 1% in an experiment, is adopted.
[0059]
On the other hand, as a result of the comparison in step ST17, the authentication acceptance determination determination function f ₁ Is not positive, the integrated score (S1, S2) is converted to a predetermined authentication rejection determination function f on a two-dimensional plane. ₂ (Second discriminant function: details will be described later with reference to FIG. 9). ₂ Is compared with 0 (zero: threshold) (step ST18). Then, the authentication rejection determination determination function f ₂ If the sign is negative (minus), it is determined that the speaker is an impostor (other person), and a determination result B for rejecting the authentication is output (step ST20). The authentication rejection determination discriminant function f ₂ And the threshold value 0 are set in the speaker determination unit 7 in advance. This discriminant function f ₂ In order to reduce the false rejection rate, for example, a value at which the false rejection rate is the same as the false acceptance rate of several percent, for example, 1% is adopted in the experiment.
[0060]
On the other hand, as a result of the comparison in step ST18, the authentication acceptance determination discriminant function f to which the integrated score (S1, S2) is input. ₁ Is not positive and the discriminant function f for authentication rejection determination which also receives the integrated score (S1, S2) ₂ Is not negative, the integrated speaker determinator 7c determines that the authentication using the integrated score (S1, S2) cannot determine the authentication acceptance or the authentication rejection with the predetermined authentication accuracy. Then, the second score S2 is accumulated and stored in the score accumulator 7a (step ST21), and the next speaker prompt C is output to the pattern designator 8 by updating the speaker authentication pattern to a different one. To instruct. Thus, the process returns to step ST11 in FIG. 5, and the second integrated speaker determination is performed.
[0061]
[Setting method of discriminant function]
Here, a method of setting the discriminant function will be described with reference to FIG. In this example, a discriminant analysis method is used. In FIG. 9, the vertical axis represents the first score S1 (generally, the score of the j-1st utterance) and the horizontal axis represents the second score S2 (generally, the score of the jth utterance).
[0062]
For example, on a plane having the first score S1 and the second score S2 (generally, the score of the (j-1) th utterance and the score of the jth utterance) as two variables, the distribution of the scores when the speaker is the person himself, The distribution of scores in the case of a person (other person) can be approximated by a two-dimensional normal distribution. Then, the variance of the score of the individual in the case where neither the authentication is accepted nor the authentication is rejected for the first utterance (generally, the j-1 utterance) is substantially equal to the variance of the score of the other person. Therefore, assuming that the variance and the covariance have a common value, the discrimination plane has a generally inclined straight line f on a two-dimensional plane as shown in FIG. ₀ Can be sought. This straight line f ₀ Is a basic discriminant function. This discriminant function f ₀ Is compared with a straight line 11 (corresponding to a conventional threshold value) parallel to the vertical axis and perpendicular to the horizontal axis, the discriminant function f ₀ As can be seen from the shaded area 12a between the straight line 11 and the shaded area 12a, the ratio of correctly determining the person as the person and accepting the authentication increases by the amount of the quirk 12a, and as can be seen from the shaded area 12b. The ratio of rejecting authentication by determining the other person as the correct person increases by the area 12b. That is, it can be seen that the authentication accuracy is improved as compared with the related art. Note that the straight line 11 is used only when the speaker is determined to be a person or another person with a predetermined determination accuracy in a speaker determination based on a single utterance using only the first score S1. This represents a threshold value common to the authentication acceptance determination and the authentication rejection determination in the case where speaker determination based on utterance is performed.
[0063]
Therefore, when using an integrated score having the first score S1 and the second score S2 (generally, the score of the (j-1) th utterance and the score of the jth utterance) as elements, a basic discrimination plane on a two-dimensional plane is used. Naive discriminant function f ₀ Is set, and a basic discriminant function f is set so that the false acceptance rate and the false rejection rate become constant small values as necessary. ₀ Are moved in parallel in a two-dimensional plane, so that the authentication acceptance determination discriminant function f ₁ And the authentication acceptance determination discriminant function f ₂ Is determined.
[0064]
The above discussion also holds true when a set having an arbitrary plurality of k scores as elements is used as an integrated score. In this case, the basic discriminant function f of the k-dimensional plane ₀ Is set as a discrimination plane, and a basic discriminant function f is set so that the false acceptance rate and the false rejection rate become constant values of, for example, 1% as necessary. ₀ Are moved in parallel in a k-dimensional plane, so that a discriminant function f for authentication acceptance determination when an integrated score having k scores as elements is used. ₁ And the authentication acceptance determination discriminant function f ₂ Is determined. Also, the discriminant function f ₀ , F ₁ , F ₂ May be set for each speaker (user), but there is no practical problem even if a common discriminant function is set for many speakers.
[0065]
[Integrated Speaker Judgment: Second Time 1]
Returning from step ST21 in FIG. 6 to step ST11 in FIG. 5, that is, in the integrated speaker determination using the integrated scores (S1 and S2), it is possible to determine whether the user is a person or another person with a predetermined authentication accuracy. If not, the integrated speaker determination unit 7d performs the second integrated speaker determination. This processing will be described with reference to FIGS.
[0066]
In FIG. 5, first, the pattern designator 8 reads out the speaker registration pattern stored in the speaker registration pattern storage 9, and based on this speaker registration pattern, the third speaker authentication A pattern such as “8790” is designated (step ST11).
[0067]
Next, the pattern designator 8 generates an utterance prompt C that prompts the utterance of the third speaker authentication pattern (for example, “8790”), and outputs the utterance prompt C from the speaker authentication device 1 (step ST12). The utterance prompt C is stored in the buffer 10 by overwriting. Next, when the speaker utters a designated pattern (“8790”) and the input audio signal A is input, the feature extractor 2 extracts an acoustic feature amount from the input audio signal A (steps ST13 and ST14). Next, the utterance content determiner 3 determines whether the utterance content of the speaker matches the utterance content indicated by the utterance prompt C stored in the buffer 10. If the result of this determination is that they do not match, the pattern designator 8 outputs the same utterance prompt C again (step ST15). Next, the score calculator 7b refers to the speaker model of the voice feature information storage 6, and refers to a score (third score) indicating the similarity between the feature extracted by the feature extractor 2 and the feature of the speaker model. S3) is calculated (step ST16).
[0068]
Next, speaker determination is performed by the integrated speaker determination unit 7d using an integrated score (second integrated score) that is a set including the second score S2 and the third score S3 as elements.
[0069]
In the second integrated speaker determination, the processing from step ST17 to ST21 in FIG. 5 is replaced with the processing from step ST17 to ST20 shown in FIG. 7 (however, there is no step ST18 or ST21).
[0070]
As shown in FIG. 7, the integrated speaker determiner 7d reads the second score S2 from the score accumulator 7a, and sets a set having the second score S2 and the third score S3 calculated by the score calculator 7b as elements. The integrated score (S2, S3) is defined as a predetermined discriminant function f common to authentication acceptance and authentication rejection. ₃ , And the resulting discriminant function f ₃ Is compared with 0 (zero: threshold) (step ST17). And the discriminant function f ₃ Is positive (plus), it is determined that the speaker is the person himself, and a determination result B for accepting the authentication is output (step ST19). On the other hand, as a result of the comparison in step ST17, the discriminant function f ₃ Is not positive, it is determined that the speaker is an impostor (another), and a determination result B for rejecting the authentication is output (step ST20). Discriminant function f ₃ Is a predetermined discriminant function of a two-dimensional plane, and is set in the speaker discriminator 7 in advance together with the threshold value 0.
[0071]
The above common discriminant function f ₃ Is the above-described authentication acceptance determination determination function f ₁ Or the authentication rejection determination discriminant function f ₂ Or these two authentication acceptance discriminant functions f ₁ Or authentication rejection determination function f ₂ Basic discriminant function f ₀ Or the basic discriminant function f ₀ Can be used by appropriately moving the discriminant function in parallel. Discriminant function f ₃ As the authentication acceptance determination discriminant function f ₁ In the case of using, the false acceptance rate becomes small, but the false rejection rate becomes slightly large. Authentication rejection determination function f ₂ On the other hand, in the case of using, the false rejection rate decreases, but the false acceptance rate slightly increases. Also, a basic discriminant function f ₀ Is used, the rejection rate is determined by the authentication acceptance determination function f ₂ Is larger than in the case of using the authentication rejection determination function f ₁ Is larger than in the case of using. In general, the basic discriminant function f ₀ Is used, the rejection rate and the false acceptance rate change according to the direction and amount of the parallel movement. Which discriminant function should be adopted may be determined by experiments or the like according to the purpose of the speaker authentication system and the authentication accuracy required by the system.
[0072]
Here, the speaker determination of the above-described embodiment (the speaker determination using the score s1 of the first utterance alone, and if the determination is not possible, the score S1 of the first utterance and the score S2 of the second utterance are integrated. This shows the effect of the speaker determination used in (1), and if the determination is still impossible, the speaker determination using the score S2 of the second utterance and the score S3 of the third utterance). As a comparative example speaker determination, the speaker determination is performed using the score s1 of the first utterance alone, and if the determination is impossible, the speaker determination is performed using only the score S2 of the second utterance alone. When the judgment was impossible, the speaker judgment was made using only the score S3 of the third utterance alone. However, in both cases of low implementation and the comparative example, the experiment was performed when two months had passed since the speaker registration. Further, the equal error rate was set to 3%.
[0073]
In the speaker determination according to the embodiment, in the speaker determination using the score s1 of the first utterance alone, the percentage of utterances that correctly determined the individual as the individual was 80%, and the percentage of utterances that correctly determined the other person as the other person was 80%. In the speaker determination using the score S1 of the first utterance and the score S2 of the second utterance for the pair of the remaining persons who could not be determined in this way, The ratio is 10%, and the ratio of utterances that correctly judged another person to be another person is 8%, and the score S2 of the second utterance and the score S3 of the third utterance are used in an integrated manner for the remaining persons who still cannot be judged. In the speaker determination, the proportion of utterances that correctly determined the person as the person was 4%, and the percentage of utterances that correctly determined the other person as the other person was 6%. In addition, with a maximum of three utterances, 94% of the person and 94% of the impostor (others) could be judged with an equal error rate of 3%, and the average number of utterances was 1.32.
[0074]
In the speaker determination of the comparative example, in the speaker determination using the score s1 of the first utterance alone, the percentage of utterances that correctly determined the person as the person was 80%, and the percentage of utterances that correctly determined the other person as the other person was 80%. In the speaker determination using only the score S2 of the second utterance alone for the pair with respect to the remaining persons who could not be determined by this, the percentage of utterances that correctly determined the person as an individual was 10%, and the The percentage of utterances correctly determined to be another person is 6%, and the remaining persons for whom the determination is still impossible cannot be determined in the speaker determination using only the third utterance score S3 alone. The proportion was 4%, and the proportion of utterances that correctly judged others as others was 7%. In addition, with a maximum of three utterances, 94% of the person and 93% of the impostor (others) could be judged with an equal error rate of 3%, and the average number of utterances was 1.34.
[0075]
From this result, in the speaker determination of the embodiment, the ratio of the other person determined by the second utterance is increased from 6% to 8% as compared with the comparative example, the authentication accuracy for the whole is improved, and the average utterance is improved. It can be seen that the number has decreased. In the speaker determination of the embodiment, the ratio of others judged in the third utterance is reduced from 7% to 6%. There is no problem because the population was reduced by a small amount.
[0076]
[Integrated speaker determination: 2nd time 2]
In the above-described second integrated speaker determination, two conclusions, that is, authentication acceptance and authentication rejection, have been concluded. However, it is also possible to conclude when neither can be determined (addition of determination). In this case, the processing from steps ST17 to ST21 in FIG. 5 is replaced with the processing from steps ST17 to ST22 shown in FIG. 8 (no step ST21), instead of FIG.
[0077]
As shown in FIG. 8, the integrated speaker determiner 7d reads the second score S2 from the score accumulator 7a, and sets a set having the second score S2 and the third score S3 calculated by the score calculator 7b as elements. The integrated score (S2, S3) is used, and the integrated score is defined as a predetermined authentication acceptance determination discriminant function f on a two-dimensional plane. ₁ (1st discriminant function) as an input, the resulting authentication acceptance judgment discriminant function f ₁ Is compared with 0 (zero: threshold) (step ST17), and the authentication acceptance determination discriminant function f ₁ Is positive (plus), it is determined that the speaker is the person himself, and a determination result B for accepting the authentication is output (step ST19). Discrimination function f for authentication acceptance determination ₁ And the threshold value 0 are set in the speaker determination unit 7 in advance. This discriminant function f ₁ In order to reduce the false acceptance rate, for example, a value at which the false acceptance rate becomes several%, for example, 1% in an experiment is adopted.
[0078]
On the other hand, as a result of the comparison in step ST17, the authentication acceptance determination determination function f ₁ Is not positive, the integrated score (S2, S3) is converted to a predetermined authentication rejection determination discriminant function f on a two-dimensional plane. ₂ As an input of (second discriminant function), the authentication rejection discriminant function f obtained as a result is ₂ Is compared with 0 (zero: threshold) (step ST18), and the authentication rejection determination function f ₂ If the sign is negative (minus), it is determined that the speaker is an impostor (other person), and a determination result B for rejecting the authentication is output (step ST20). The authentication rejection determination discriminant function f ₂ And the threshold value 0 are set in the speaker determination unit 7 in advance. This discriminant function f ₂ In order to reduce the false rejection rate, for example, a value at which the false rejection rate is the same as the false acceptance rate of several percent, for example, 1% is adopted in the experiment.
[0079]
On the other hand, as a result of the comparison in step ST18, the authentication acceptance determination discriminant function f to which the integrated score (S2, S3) has been input. ₁ Is not positive, and the authentication rejection determination discriminant function f is input with the integrated score (S2, S3). ₂ Is not negative, the integrated speaker determiner 7c determines that the authentication using the integrated score (S2, S3) cannot determine the authentication acceptance or the authentication rejection with the predetermined authentication accuracy. (Step ST22), and the determination ends.
[0080]
In the speaker determination of the above embodiment, the speaker determination is performed by using the score s1 of the first utterance alone, and the score S1 of the first utterance and the score of the second utterance for the pair of the remaining persons who cannot be determined. Speaker determination is performed by using the score S2 of the second voice in an integrated manner, and the speaker determination is performed by using the score S2 of the second utterance and the score S3 of the third utterance integratedly for the remaining persons who still cannot be determined. However, there are various ways of setting the integrated score or combining the integrated speaker determination and the speaker determination based on a single utterance, and some of them are exemplified below.
(1) The speaker determination is performed using the score s1 of the first utterance alone, and the score S1 of the first utterance and the score S2 of the second utterance are integratedly used for the remaining persons who cannot be determined. Speaker determination is performed, and speaker determination is performed using the score S1 of the first utterance, the score S2 of the second utterance, and the score S3 of the third utterance for the remaining persons who cannot be determined. In this case, the second integrated score is a set (S1, S2, S3) having S1, S2, and S3 as elements.
(2) The speaker determination is performed using the score s1 of the first utterance alone, and the score S1 of the first utterance and the score S2 of the second utterance are integratedly used for the remaining persons who cannot be determined. Speaker determination is performed, and speaker determination is performed using the score S1 of the first utterance and the score S3 of the third utterance in an integrated manner with respect to the remaining persons for whom determination is impossible. In this case, the second integrated score is a set (S1, S3) having S1 and S3 as elements. In other words, the element of the integrated score does not necessarily have to be a score of a continuous utterance.
(3) The first and second utterances are continuously performed without performing the speaker determination using the score s1 of the first utterance alone, and the score S1 of the first utterance and the score S2 of the second utterance are integrated. For the remaining persons for whom the determination cannot be made using the score S2 of the second utterance and the score S3 of the third utterance, or the scores S1 and S3 of the first utterance. Speaker determination is performed by using the utterance score S3 in an integrated manner (the element of the integrated score does not necessarily need to be a continuous utterance score). As in this case, it may be possible to determine whether the person is a person or another person only by the integrated speaker determination.
(4) The first, second, and third utterances are continuously performed without performing the speaker determination using the score s1 of the first utterance alone, and the score S1 of the first utterance and the score S2 of the second utterance are performed. And the score S3 of the third utterance are used collectively to perform the speaker determination collectively.
(5) The speaker determination is performed using the score s1 of the first utterance alone, and the speaker determination is performed using the score S2 of the second utterance alone for the remaining persons who cannot be determined. For the remaining persons who cannot be determined, the score S1 of the first utterance, the score S2 of the second utterance, and the score S3 of the third utterance are integrated, or the score S2 of the second utterance and the score of the third utterance. The speaker determination is performed using S3 in an integrated manner or using the score S1 of the first utterance and the score S3 of the third utterance in an integrated manner.
(6) The speaker determination is performed using the score s1 of the first utterance alone, and the speaker determination is performed using the score S2 of the second utterance alone for a person who cannot be determined as a result. The speaker is determined using the score S3 of the third utterance alone for the person who became, and the score S1 of the first utterance, the score S2 of the second utterance, and the third utterance of the remaining persons for whom the determination was still impossible Of the second utterance, the score S3 of the third utterance, and the score S1 of the first utterance and the score S3 of the third utterance. Is used for speaker determination.
[0081]
In short, when n = 3 as in the above-described embodiment, different speaker authentication utterance patterns are specified at most three times for the speaker, and scores corresponding to two or three utterance pattern specifications are given. Is set as an integrated score, and the integrated score is used as an input of a discriminant function on a two- or three-dimensional plane, and it is determined from the sign of the discriminant function whether the speaker is the person or another person. At this time, the integrated speaker determination may be applied in one or two stages, or the speaker determination based on a single utterance may be used in combination as necessary.
[0082]
In general, in the present invention, a plurality of n different speaker authentication utterance patterns are specified for a speaker at a maximum, and k utterance patterns corresponding to a plurality of k (2 ≦ k ≦ n) utterance patterns are specified. Is used as an integrated score, and the integrated score is used as an input of a discriminant function on the k-dimensional plane, and a speaker authentication method is used in which it is determined whether the speaker is the person or the other person from the sign of the discriminant function. be able to. At that time, speaker determination based on a single utterance is used together as necessary. As the integrated score, for example, two or three or more scores such as a set having at least the first utterance score as an element, a set having n−1 utterance scores and an n-th utterance score as elements, and the like. Use a set as an element. In addition, it is preferable to apply such integrated speaker determination in one step or in two or more steps.
[0083]
Note that the speaker authentication technology of the present invention can be applied to a voice authentication device for telephone services such as telephone banking, a voice authentication device for various Web services, and the like.
[0084]
Also, a program for realizing each step shown in FIGS. 5 to 9 is recorded on a computer-readable recording medium, and the program recorded on this recording medium is read into a computer system and executed. Authentication processing may be performed. Here, the “computer system” may include an OS and hardware such as peripheral devices. The “computer system” includes a homepage providing environment (or a display environment) if a WWW system is used. The “computer-readable recording medium” refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, and a CD-ROM, and a storage device such as a hard disk built in a computer system.
[0085]
Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that serves as a server or a client when a program is transmitted through a network such as the Internet or a communication line such as a telephone line. As described above, a program holding a program for a certain period of time is also included.
[0086]
Further, the above program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the "transmission medium" for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line.
[0087]
Further, the program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.
[0088]
As described above, the embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, for example, the continuous word may be a continuous word or symbol such as a continuous alphabet or a continuous alphabet. The present invention is not limited to this, and includes design changes and the like within a range not departing from the gist of the present invention.
[0089]
【The invention's effect】
As described above, according to the present invention, for utterances in which it is difficult to judge the person and another person (spoofer), the designated pattern is changed to encourage re-speech and judgment is made using scores of a plurality of utterances in an integrated manner. It is possible to guarantee the authentication accuracy set in advance. Further, according to the present invention, a text-designated speaker authentication that is robust and maintains high authentication accuracy against “recording / playback-type spoofing” in which the utterance of the person is secretly recorded and reproduced later. Technology can be realized. Further, in the case of performing a multi-stage speaker determination using a plurality of utterances in an integrated manner to guarantee a preset authentication accuracy, the authentication accuracy is improved and the proportion of users (speakers) who can guarantee the authentication acceptance is increased. can do.
[0090]
In addition, when it cannot be determined that the authentication is accepted or rejected, a new utterance pattern that enhances the authentication accuracy compared to the previously specified speaker authentication pattern can be used to reduce the burden on the user. The authentication efficiency can be improved by gradually increasing the authentication accuracy. In addition, when re-designating the speaker authentication pattern, if the number of pairs of chains of a plurality of words in the speaker registration pattern is increased from the previously specified speaker authentication pattern, the number of uttered words becomes Although it increases, the authentication accuracy can be improved while maintaining the effect of preventing identity fraud.
[0091]
Furthermore, by designating the speaker as a speaker authentication utterance pattern including at least one set of a plurality of word chains included in the utterance pattern used for speaker registration, it is possible to prevent false identification and burden on the user. Can be reduced as much as possible and authentication can be performed efficiently. In this case, when re-specifying the speaker authentication pattern, if the number of a series of word chains in the speaker registration pattern is increased from the previously specified speaker authentication pattern, the authentication accuracy is prioritized. Can be enhanced.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of a speaker authentication device according to an embodiment of the present invention.
FIG. 2 is a block diagram illustrating a configuration example of a speaker determination unit.
FIG. 3 is a flowchart illustrating a flow of speaker registration processing.
FIG. 4 is a diagram showing an example of voice pattern designation.
FIG. 5 is a flowchart showing the flow of speaker authentication processing.
FIG. 6 is a flowchart showing the flow of a first integrated speaker determination portion during speaker authentication processing.
FIG. 7 is a flowchart showing the flow of a second integrated speaker determination portion during speaker authentication processing.
FIG. 8 is a diagram illustrating an example of a setting method of a discriminant function.
FIG. 9 is a flowchart illustrating another example of the second integrated speaker determination during the speaker authentication processing.
[Explanation of symbols]
1 Speaker authentication device
2 Feature extractor
3 utterance content judgment device
4 switch
5 Registration judgment unit
6 Voice feature information storage
7 Speaker judgment device
7a Score accumulator (score accumulating means)
7b Score calculator (score calculating means)
7c Speaker determination unit for single utterance (speaker determination means by single utterance)
7d Integrated speaker judgment device (Integrated speaker judgment means)
8. Pattern designator (voice pattern designator)
9 Speaker registration pattern storage (voice pattern storage means)
10 buffers
11 Straight line equivalent to conventional threshold
12a Area where the rate of authentication acceptance of the individual increases
12b Area where the rate of rejecting others increases

Claims (10)

In a speaker authentication device for performing personal authentication based on a score representing a similarity between a speaker's utterance and a speaker model,
Utterance pattern designating means for designating different utterance patterns for speaker authentication to a speaker at a maximum of n times,
Score calculation means for calculating, as an i-th score, a score between the utterance of the speaker corresponding to the i-th utterance pattern designation and the speaker model;
A set of k scores corresponding to a plurality of k utterance pattern designations is set as an integrated score, and the integrated score is used as an input of a discriminant function on a k-dimensional plane. A speaker authentication device comprising integrated speaker determination means for determining whether a person is another person. In claim 1,
A speaker determination unit based on a single utterance that determines whether the speaker is the person or another person by independently comparing the first score with a threshold value,
The integrated speaker determination means determines whether the speaker is the person or another person when the speaker determination means based on a single utterance cannot determine whether the speaker is the person or another person with predetermined authentication accuracy. A speaker authentication device, characterized in that: 3. The speaker authentication device according to claim 2, wherein the integrated score is a set having at least the first score as an element. In claim 1,
Comparing the n-th score alone with a threshold value, a speaker determination unit based on a single utterance that determines whether the speaker is the person or another person is provided,
The integrated speaker determination means determines whether the speaker is the person or another person when the speaker determination means based on a single utterance cannot determine whether the speaker is the person or another person with predetermined authentication accuracy. A speaker authentication device, characterized in that: 5. The speaker authentication device according to claim 4, wherein the integrated score is a set including at least the (n-1) th score and the nth score as elements. In claim 1,
n = 3, k = 2, and a speaker determination unit based on a single utterance for individually comparing the first score with a threshold to determine whether the speaker is the person or the other person;
The integrated speaker determination means determines whether the speaker is a person or another person with a predetermined authentication accuracy by comparing the first score and the threshold value by the speaker determination means based on a single utterance. The first integrated score, which is a set having the score and the second score as elements, is used as an input of a discriminant function to determine whether the speaker is a person or another person. In the determination based on the first integrated score, a predetermined authentication accuracy is used. When it is not possible to determine whether the speaker is the person or another person, the second integrated score, which is a set having the second score and the third score as elements, is used as an input of the discriminant function to determine whether the speaker is the person or another person A speaker authentication device for performing the following. 2. The speaker authentication utterance pattern designated by the utterance pattern designation means in the i-th order according to claim 1, wherein the utterance pattern designation unit has an authentication accuracy higher than that of the speaker authentication utterance pattern designated by the (i-1) -th order. A speaker authentication device characterized by the above-mentioned. 2. The speaker according to claim 1, further comprising utterance pattern storage means for storing a speaker registration utterance pattern used for speaker registration, wherein said utterance pattern designation means is a speaker authentication utterance pattern. A speaker authentication device for designating an utterance pattern including at least one set of a chain of a plurality of words included in a registration utterance pattern. 2. The method according to claim 1, wherein the discriminant function includes a first discriminant function for accepting a person and a second discriminant function for rejecting another person, wherein the first discriminant function and the second discriminant function are different from each other. A speaker authentication apparatus characterized in that the rejection rate is set to a constant value. A method for setting a discriminant function used in the speaker authentication device according to claim 9,
approximating the distribution of k scores with a k-dimensional normal distribution on a k-dimensional plane to obtain a discrimination plane on the k-dimensional plane;
A discriminant function setting method characterized by moving the discriminant plane in parallel and setting a first discriminant function and a second discriminant function in which the false acceptance rate and the false rejection rate are constant values.

Images