JP3631020B2 - Speaker recognition method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speaker recognition method for determining whether a spoken voice belongs to a specific individual.
[0002]
[Prior art]
In the speaker recognition device for determining whether the spoken voice is of a specific individual, the likelihood is calculated from the registered voice feature data of the designated user and the feature data of the spoken voice, and is set in advance. The determination is made by comparing with a threshold value. If the likelihood is greater than the threshold, it is accepted as the designated person, and if it is smaller, it is rejected as a spoofer.
[0003]
However, if the threshold is set strictly (that is, large), the error (FR: False Rejection) that the correct person is rejected increases, and if the threshold is relaxed (that is, small), the false person is accepted (FA: False). (Acceptance) increases, and it is important to set the threshold value to an appropriate value. Document: Japanese Patent Laid-Open No. 9-198086 “Speaker recognition threshold setting method and speaker recognition apparatus using this method” describes a method of setting a threshold in the learning process of a speaker model.
[0004]
[Problems to be solved by the invention]
However, in the method of determining acceptance / rejection with one threshold as represented by the above method, the utterance time at the time of registration or recognition will be increased if the accuracy is further improved no matter how appropriate the threshold is set. There is a problem that it is necessary to increase the number of utterances or to increase the burden on all users.
[0005]
For example, “J. L. Gauvin, “Experiments with Speaker Verification over the Telephone”, EUROSPECH '95, 1995 ”, the longer the utterance at the time of recognition and the more the number of utterances, the higher the accuracy (that is, the above-mentioned FRFA). It is reported that the error rate EER: Equal Error Rate becomes smaller.
[0006]
An object of the present invention is to solve the above-mentioned conventional problems and to provide a speaker recognition method capable of reducing a user's burden when setting a threshold for speaker recognition.
[0007]
Therefore, in the speaker recognition method of the first invention, as the first recognition step, the user is caused to execute the utterance as the first problem with a light burden, and TA> TB for each user. Two thresholds TA and TB are provided, and if the likelihood calculated from the feature data of the pronounced speech and the feature data stored in the specified user database is greater than the threshold TA, the user is accepted as the principal. However, if the likelihood is smaller than the threshold TB, it is rejected as an impersonator, and if it is between the thresholds TA and TB, it is more burdensome for the user than the first task for more precise recognition. A speaker recognition method having a second recognition step of executing utterance as a second task and verifying whether or not the user is the person using the threshold value T. The threshold TA is used for the first task. When the fraudster acceptance rate is fraud in the second recognition step The threshold TB is set to be equal to the person rejection rate in the second recognition step, and the threshold TB is set to be equal to the person acceptance rate. .
[0009]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0010]
[Description of First Embodiment]
[Description of configuration]
FIG. 1 shows a block diagram of an apparatus for realizing the speaker recognition method according to the first embodiment.
[0011]
The feature quantity analysis unit 11 is connected to the learning / recognition control unit 12 and outputs an analysis result of the input voice data. The learning / recognition control unit 12 is selectively connected to the feature amount storage unit 13 or the likelihood calculation unit 14 and outputs the data obtained from the feature amount analysis unit 11 to one of them. The feature quantity storage unit 13 is connected to the likelihood calculation unit 14 and the threshold setting unit 15 and outputs feature data corresponding to the ID output from the learning / recognition control unit 12. The likelihood calculation unit 14 includes likelihood calculation units (1) and (2), each of which is connected to the first determination unit 16 and the second determination unit 17, and is output from the learning / recognition control unit 12. The likelihood is calculated based on the data and the data output from the feature amount storage unit 13 and output to any one. The threshold setting unit 15 is connected to the first determination unit 16 and the second determination unit 17 and outputs a threshold as necessary. The first determination unit 16 is connected to the result output and learning / recognition control unit 12, and determines and outputs the result based on the data output from the likelihood calculation unit 14 and the threshold setting unit 15. Alternatively, the process is returned to the learning / recognition control unit 12. The second determination unit 17 is connected to the result output, and determines and outputs the result based on the data output from the likelihood calculation unit 14 and the threshold setting unit 15.
[0012]
[Description of operation]
First, the flow of processing when registering (learning) a user will be described with reference to FIG.
At the time of registration (learning), the switch of the learning / recognition control unit 12 is connected to the feature amount storage unit 13 (S100). When the user's ID and voice are input, the feature quantity analysis unit 11 analyzes the input voice, extracts personality feature data, and outputs it to the learning / recognition control unit 12 (S101, 102). ). The learning / recognition control unit 12 outputs the input feature data and ID to the feature amount storage unit 13. The feature quantity storage unit 13 stores the input ID and feature data in association with each other (S103). This completes the registration process of the user corresponding to one ID, and repeats the registration process according to the number of users.
[0013]
Next, the flow of processing when recognizing whether or not the user is the user will be described with reference to FIG.
[0014]
The user executes a task (issue 1) with a small burden (for example, the utterance time is short or the number of utterances is small). At this time, the switch of the learning / recognition control unit 12 is connected to the likelihood calculating unit (1) 14 (S200). When the user ID and the voice of the task 1 are input, the feature amount analysis unit 11 analyzes the input voice, extracts the personality feature data, and outputs it to the learning / recognition control unit 12 (S201). 202). The learning / recognition control unit 12 outputs the input ID and feature data to the likelihood calculation unit (1) 14. The likelihood calculation unit (1) 14 searches for feature data corresponding to the ID input from the feature amount storage unit, calculates the likelihood by comparing with the feature data output from the learning / recognition control unit 12, It outputs to the 1st determination part 16 (S203,204). Here, the likelihood indicates that the larger the value, the more similar the feature data. The first determination unit compares the output likelihood with thresholds TA and TB set by the threshold setting unit 15 (TA> TB: a setting method will be described later), and if the likelihood is greater than TA, the user Is accepted as the principal of the ID, and the process is terminated (S205, 213). If the likelihood is smaller than TB, the user is rejected as an ID spoofer, and the process is terminated (S206, 214). In other cases, the learning / recognition control unit 12 is notified, and a task (task 2) with a heavy burden on the user (for example, a long utterance time or a large number of utterances) is executed.
[0015]
In Task 2, the switch of the learning / recognition control unit 12 is connected to the likelihood calculating unit (2) 14 (S207). When the voice of the task 2 is input, the feature data extracted by the feature quantity extraction unit 11 passes through the learning / recognition control unit 12 as in the case of the task 1, and this time the likelihood calculation unit (2) 14 together with the ID. (S208, 209). The likelihood calculating unit (2) 14 searches the feature data for the input ID, calculates the likelihood, and outputs the likelihood to the second determining unit 17 (S210, 211). The second determination unit 17 compares the output likelihood with a threshold T set by the threshold setting unit 15 (the setting method will be described later). If the likelihood is greater than T, the user is identified as the identity person. Acceptance, otherwise, the user is rejected as an ID spoofer and the process is terminated (S212, 214).
[0016]
Next, an example of a threshold setting method in the threshold setting unit 15 will be described.
[0017]
The threshold setting unit 15 sets three thresholds T, TA, and TB for each user ID based on the user feature data stored in the feature amount storage unit 13. T can be obtained by using speech data (feature data) used for learning as disclosed in the above-mentioned document: Japanese Patent Laid-Open No. 9-198086. However, here, audio data (feature data) corresponding to the problem 2 is used. Further, in this document, the threshold value T is set so that the principal rejection rate and the impersonator acceptance rate are equal, but either may be set to be smaller (good).
[0018]
TA and TB are set depending on the rejection rate of the person, the rate of acceptance of the impersonator, and the determination rate in task 1 (the ratio at which acceptance / rejection can be determined in task 1 regardless of correctness). Here, a setting method in the case where the subject rejection rate and the spoofer acceptance rate in Task 1 are made equal to the subject rejection rate and the spoofer acceptance rate in Task 2 will be described with reference to FIG. When the threshold value is T in Task 2, the impersonator acceptance rate is EA, and the principal rejection rate is EB. In general, the smaller the user burden (the shorter the utterance time and the fewer the number of utterances), the lower the reliability of the likelihood and the worse the recognition rate. In other words, both the subject rejection rate and the fraudster acceptance rate, the curve (dotted line) of task 1 is above the curve (solid line) of task 2 (the one with the lower recognition rate), and the threshold is T. The person acceptance rate and the principal rejection rate are larger (bad) than EA and EB, respectively. Here, with respect to the problem 1, if the threshold value is made larger or smaller than T, there may be a point where the spoof acceptance rate becomes EA and a person rejection rate becomes EB. The thresholds at this time are set as TA and TB, respectively. However, depending on the recognition rate of Task 1, there may be no threshold at which the false actor acceptance rate becomes EA or the threshold at which the person rejection rate becomes EB. In that case, only the existing threshold value is enabled or the task is changed.
[0019]
As described above, according to the first embodiment, two threshold values TA and TB (TA> TB) are prepared for the task 1 with a small user burden, and the user is more likely when the likelihood is larger than TA. If the user is smaller than TB, the user can be rejected as an ID spoofer, so that many of the users can determine whether to accept or reject the task with a small burden. At this time, depending on the problem and the threshold value, it is possible to maintain the same person rejection rate and fraudster acceptance rate as the problem 2 with a large user burden.
[0020]
[Description of Second Embodiment]
[Description of configuration]
In the first embodiment described above, the method of authenticating with the voice feature data for both of the tasks 1 and 2 has been described. However, the task 2 may be a means that does not use the voice feature data. Here, an example in which a password is input as problem 2 will be described.
[0021]
FIG. 5 shows a block diagram of an apparatus for realizing the speaker recognition method according to the second embodiment.
[0022]
The feature amount analysis unit 11, the learning / recognition control unit 12, the feature amount storage unit 13, the likelihood calculation unit 14, the threshold setting unit 15, and the first determination unit 16 that perform the task 1 are implemented except for the following points. This is the same as the first embodiment. The learning / recognition control unit 12 is connected to the second control unit 19 in addition to the feature amount storage unit 13 and the likelihood calculation unit 14, and outputs an ID and passes control when it cannot be determined in the task 1. . The second control unit 19 is selectively connected to the password storage unit 18 and the second determination unit 17 and outputs the input password and ID. The password storage unit 18 is connected to the second determination unit 17 and outputs a password corresponding to the ID output from the second control unit 19 as necessary. The second determination unit 17 is connected to the result output, and determines and outputs the result from the ID and password output from the second control unit 17 and the password storage unit 18.
[0023]
[Description of operation]
The learning of voice feature data is the same as in the first embodiment.
When the password is learned, the switch of the second control unit 19 is connected to the password storage unit 18. When the ID and password are input, the second control unit 19 outputs them to the password storage unit 18. The password storage unit 18 stores the ID and the password in association with each other. This completes the learning process for one ID.
[0024]
Next, the flow of processing during recognition will be described.
Problem 1 is the same as in the first embodiment.
In Task 2, the learning / recognition control unit 12 outputs the ID to the second control unit 19 and passes control. When a password is input as the assignment 2, the second control unit 19 outputs the password and the ID output from the learning / recognition control unit 12 to the second determination unit 17. The second determination unit 17 searches the password storage unit 18 for a password corresponding to the ID output from the second control unit 19 and compares it with the password output from the second control unit 19. Accept as the ID person, otherwise reject the user as an ID spoofer and end the process.
[0025]
As described above, according to the second embodiment, a means that does not use voice feature data can be taken as problem 2, and can be easily integrated with other authentication means such as a password.
[0026]
【The invention's effect】
As described above in detail, according to the first invention, as the first recognition step, the user is caused to execute utterance as the first problem with a light burden, and TA> TB for each user. Two thresholds TA and TB are provided, and if the likelihood calculated from the feature data of the pronounced speech and the feature data stored in the specified user database is greater than the threshold TA, the user is selected. If it is accepted as the principal and is smaller than the threshold TB, it is rejected as an impersonator, and if it is between the thresholds TA and TB, it is more burdensome for the user than the first task for more precise recognition. A speaker recognition method having a second recognition step for executing the utterance as the second task and verifying whether or not the user is the person using the threshold T, when the first task is performed with the threshold TA. The rate of accepting an impersonator is accepting an impersonator in the second recognition step Since the threshold TB is set to be equal to the rejection rate in the second recognition step, the threshold TB is set to be equal to the rate. Many of them can accept / reject decisions with less burdensome tasks.
[Brief description of the drawings]
FIG. 1 is a block diagram of an apparatus for realizing a speaker recognition method according to a first embodiment.
FIG. 2 is a flowchart of processing at the time of user registration.
FIG. 3 is a flowchart of processing during recognition in the first embodiment.
FIG. 4 is an explanatory diagram for setting a threshold value.
FIG. 5 is a block diagram of an apparatus for realizing a speaker recognition method according to a second embodiment.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 11 Feature-value analysis part 12 Learning / recognition control part 13 Feature-value preservation | save part 14 Likelihood calculation part 15 Threshold setting part 16 1st determination part 17 2nd determination part 18 Password preservation | save part 19 2nd control part

Claims (1)

As the first task with a light burden on the user, utterance is executed, and two threshold values TA and TB are set so that TA> TB for each user. The likelihood is calculated from the feature data stored in the person's database, and if the likelihood is greater than the threshold TA, the user is accepted as the person, and if the likelihood is less than the threshold TB, the person is rejected as a spoofer A first recognition step;
When the likelihood is between the threshold values TA and TB in the first recognition step, the second problem is more burdensome for the user than the first problem for more precise recognition. A speaker recognition method having a second recognition step of executing an utterance and verifying whether or not the user is authentic using a threshold T,
The threshold TA is set so that the impersonator acceptance rate when the first problem is implemented is equal to the impersonator acceptance rate in the second recognition step, and the threshold TB is implemented as the first problem. A speaker recognition method, wherein the person rejection rate when set to be equal to the person rejection rate in the second recognition step is set.

Images