JP4440414B2 - Speaker verification apparatus and method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speaker verification apparatus or method for determining by voice whether or not a user is a user based on a feature amount of voice data registered in advance.
[0002]
[Prior art]
In recent years, with the development of computer technology, the communication environment has been rapidly improved. With the development of such a communication environment, it has become possible to perform computer access (Computer Telephony Integration) by telephone even in ordinary homes.
[0003]
In the field of computer access by telephone, there is a problem in accessing information that should not be communicated to anyone other than the person or a specific person, such as information relating to privacy or information having a confidentiality obligation. In other words, for example, when using a touch-tone phone, it is possible to acquire the authority to access the information by entering the password by operating a button on the phone, but if someone knows the password, Nevertheless, the problem is that the information can be easily accessed. For this reason, there is an increasing need to collate whether or not the user is a person or a specific individual using a voice unique to the person.
[0004]
[Problems to be solved by the invention]
However, speech synthesis technology has also made rapid progress in recent years, and it is no longer impossible to express the personality of a speaker by making full use of such technology.
[0005]
That is, in the prior art, a human voice is assumed as an input for speaker verification, and in order to accurately synthesize a specific human voice, a large amount of the human voice waveform data is used. Because it had to be recorded and edited, it was not feasible.
[0006]
However, in recent years, it has become possible to realize speech synthesis reflecting personality by recording a little of the person's voice, and it has become possible to easily imitate the voice of others.
[0007]
By using such a synthesized speech device that can also express the personality of a speaker, a third party can impersonate a specific individual, and the speaker verification system itself can be abused. Has occurred.
[0008]
In order to solve the above problems, the present invention provides a speaker verification apparatus and method that can accurately perform speaker verification even when using synthesized speech that also expresses the personality of the speaker. The purpose is to provide.
[0009]
[Means for Solving the Problems]
In order to achieve the above object, a speaker verification apparatus according to the present invention is a speaker verification apparatus that determines whether or not the voice of a speaker to be input by voice matches the voice of a registered speaker registered in advance. A speaker input instructing unit that gives instructions to the speaker about the utterance content to be input, a voice input unit that inputs one or more of the voices of the speaker, and a voice analysis unit that analyzes the voice input by the voice input unit And an input voice similarity calculation unit that calculates the similarity between two or more voices that are the same utterance content, and the calculated similarity is a similarity of a predetermined level or higher that is close to perfect match In the case of the degree, the speaker is collated using information that the degree of similarity matches.
[0010]
With such a configuration, it is possible to identify artificial voices as well as judgments as to whether or not the person is the person himself / herself as having no fluctuations that human voice should originally have. It is possible to prevent an unrelated third party from impersonating himself / herself using a speech synthesizer or the like.
[0011]
In the speaker verification device according to the present invention, it is preferable that the similarity is determined based on whether the verification process for the registered speaker model is the same. In human speech, since the length and spectrum of utterances differ for each utterance, the matching process cannot match, so by comparing the matching processes, a third party who has nothing to do with the person It is possible to prevent an act of impersonating the person using a speech synthesizer or the like.
[0012]
In the speaker verification device according to the present invention, when at least two or more voices are input by the voice input unit and conversion processing is performed on at least one voice, the voice input unit Of the input at least two or more sounds, it is preferable to perform signal processing on the sound that has not been subjected to conversion processing, or to perform normalization processing on at least two or more input sounds. About the act of artificially generating natural fluctuations in speech input and recognizing that it is not synthetic speech by applying some filter to the input speech for the second and subsequent input when inputting speech multiple times This is to prevent it.
[0013]
Moreover, it is preferable that the speaker collation apparatus according to the present invention rejects the input when it is determined that the similarities coincide with each other and it is determined that the input is not the user's voice input. This is because when the voice data that has been repeatedly spoken almost completely matches, it is not judged that the person is the person himself as a suspicion of the recorded material.
[0014]
Further, the present invention is characterized by software that executes the function of the speaker verification device as described above as a processing step of the computer. Specifically, the voice of the speaker to be input is registered in advance. A speaker verification method for determining whether or not the voice of a registered speaker matches, a step of giving an instruction to the speaker about the input utterance content, and a step of inputting one or more of the speaker's voice A step of analyzing the input voice and a step of calculating a similarity between two or more voices having the same utterance content, and the calculated similarity is a predetermined close to perfect match If the similarity is equal to or higher than the level, the speaker verification method for verifying the speaker using the information that the similarities match, and a computer-readable recording medium recording such a process as a program And features.
[0015]
With such a configuration, it is possible to determine whether the user is the person by loading and executing the program on the computer, and the artificially generated synthesized speech does not have any fluctuation that human speech should originally have. Therefore, it is possible to realize a speaker verification device that can prevent a third party who is completely unrelated to the person from impersonating the person using a speech synthesizer or the like. It becomes possible.
[0016]
DETAILED DESCRIPTION OF THE INVENTION
(Embodiment 1)
Hereinafter, a speaker verification device according to a first exemplary embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram of a speaker verification device according to a first exemplary embodiment of the present invention.
[0017]
In FIG. 1, reference numeral 1 denotes a personal ID input unit for inputting a personal ID at the time of speaker verification. Reference numeral 2 denotes an individual voice information registration unit, which stores voice information in a database for each individual ID in advance. Here, not only the speech waveform data but also the feature value obtained by analyzing the speech data is registered in advance.
[0018]
Next, reference numeral 3 denotes a voice input instructing unit for giving an instruction regarding voice to be input by the user at the time of speaker verification. Reference numeral 4 denotes a voice input unit, which is used by a user to actually speak and input voice data through an input medium such as a microphone.
[0019]
The voice input unit 4 inputs voice in accordance with instructions from the voice input instruction unit 3. In this case, a relatively long voice including the same utterance content may be input only once, or the same utterance content may be repeated twice or more. Further, different utterance contents including the same utterance contents may be input. For example, by saying “Mr. Maekawa” and “Mr. Hayakawa”, the portion of “Mr. Kawa” has the same utterance content, so that comparison can be performed. With such an input method, it is difficult to notice when the user is collating the same voice, and it can be expected that collation is performed with relatively high accuracy.
[0020]
Therefore, for example, by providing the input voice storage unit 21 as shown in FIG. 2, it is conceivable to collate the utterance contents based on the voice data input last time. If you are a human being, it is natural for voice fluctuations to occur when you change time and place, so it is possible to determine that an input that hardly sees fluctuations is synthetic voice or recorded voice, etc. It is.
[0021]
Reference numeral 5 denotes a voice analysis unit, which analyzes input voice data and obtains physical feature quantities of the voice waveform data. Based on the obtained feature amount, the registered speech similarity calculation unit 6 obtains the first similarity between the feature amount of the speech data registered in the individual speech information registration unit 2 and the feature amount of the speech data of the input speech. The input voice similarity calculation unit 7 calculates the second similarity of the feature amount of the voice data for the input voice having the same content.
[0022]
Generally, in conventional speaker verification, similarity determination is performed only by calculating the first similarity between the input voice and the registered voice based on the feature amount of the voice. However, when a human utters, the voice fluctuates according to the situation and environment at that time, and it is impossible to utter as the same voice. If the similarity of 1 is within the range, a method of determining that the same person is the same is adopted.
[0023]
Therefore, by using a speech synthesizer or the like to generate synthesized speech adjusted so that the first similarity is within the allowable range, a third party can easily impersonate the person.
[0024]
On the other hand, the synthesized speech artificially generated by a speech synthesizer or the like does not fluctuate, and the same speech can be input no matter how many times it is input. Therefore, in addition to the conventional speaker verification, even when the same speech is input a plurality of times, the second similarity that is the similarity between the input speech is calculated as the same value every time. Therefore, it can be determined that the voice is unnatural such as a synthesized voice.
[0025]
Specifically, a case where the similarity determination criterion is a collation distance between audio data will be described with reference to FIG. FIG. 3 shows the frequency distribution of the collation distance between the audio data, and it is judged that the similarity is higher as the collation distance is shorter.
[0026]
In FIG. 3, an area A is an area indicating a distribution of distances between the feature amount of the voice data registered in the individual-specific voice information registration unit 2 and the feature amount of the input person's voice data. A region B shows a distribution of collation distances related to the feature amount of the voice data in the voice part having the same content, for example, a collation distance distribution between the first and second input voices. Region C shows the distribution of the collation distance between the feature amount of the speech data registered in the individual speech information registration unit 2 and the feature amount of the speech data of the spoiler input speech.
[0027]
That is, region A and region C are distributions of collation distances with audio data registered in the individual-specific audio information registration unit 2, whereas region B is a distribution of collation distances between input audio data. It is very different in a certain point.
[0028]
First, in the conventional method, it is determined whether or not the voice input between the area A and the area C is the person's own voice. That is, when the collation distance of the input voice data is smaller than the threshold value I which is a predetermined threshold value, it is determined that the similarity of the input voice is high, and the input voice is the voice of the user. Judge that there is.
[0029]
On the other hand, between region A and region B, it is determined whether the input sound is natural sound or synthesized sound. That is, when the collation distance between the input voice data is smaller than the threshold value II which is a predetermined threshold value, it is determined that the input voice does not have a natural fluctuation inherent in humans. It is determined that the voice is an unnatural voice such as a synthesized voice or a recorded voice.
[0030]
Next, the similarity calculation method in the registered voice similarity calculation unit 6 and the input voice similarity calculation unit 7 will be described. First, a fixed value that has been conventionally used is used as a threshold value for determining that the feature amount of the voice data corresponding to the specific personal ID is similar to the feature amount of the input voice data. There are many cases. For example, as shown in FIG. 4, the collation distance is calculated between the input voice and the pre-registered voice, and compared with a preset threshold value. Are the same or short ("-" in FIG. 4), it is determined that the person is the person, and if it is long ("+" in FIG. 4), the person is determined to be the other person.
[0031]
For setting such a threshold value, the following method is often used. FIG. 5 shows the identity rejection rate FRR (False Rejection error), which is the probability that the decision to reject a person who is not the person is wrong, with the comparison distance being used as an index for similarity judgment. Rate) is taken on the vertical axis. On the other hand, with the collation distance as the horizontal axis, the false acceptance rate FAR (False Acceptance error Rate), which is the probability when the determination that the person is an impersonator is wrong, is also taken on the vertical axis.
[0032]
If the threshold value is set to a small value, the rate FAR for erroneously accepting a spoofer decreases, but the rate FRR for erroneously rejecting the person increases. Conversely, if the threshold value is set to a large value, the rate FRR that erroneously rejects the person becomes small, but the rate FAR that erroneously accepts the false person becomes large. Therefore, it is desirable to set the threshold value to an appropriate value according to the importance of these two error rates.
[0033]
It is common to experimentally evaluate the value at which the two error rates are equal afterwards as a threshold value. In the first embodiment, the threshold value I in FIG. 3 is an experimental value based on human speech, and the threshold value II is an experimental value based on a synthesized speech generated by a speech synthesizer. The value that matches is the threshold value. That is, when a predetermined threshold is set, the area of the portion of the frequency distribution curve (FIG. 3) of the distance between the voices of the persons and the distance between the voices of the person and the other person that protrudes from the determined threshold Indicates FAR, FAR.
[0034]
Similarly, various methods can be considered as a method for calculating the collation distance for determining whether or not the input speech is synthetic speech or the like. In the first embodiment, the feature amount of the voice data is set as an n-dimensional feature parameter, and the matching distance is obtained as an in-space distance in the n-th space. However, the method is not particularly limited to this method, and DP matching may be used as a method for calculating the matching distance. Here, DP means dynamic programming.
[0035]
For example, FIG. 6 shows a frequency distribution of distances within the same speaker with respect to word utterances included in the content uttered at the same time using DP matching. It is possible to obtain a distance distribution to be determined by such a method.
[0036]
FIG. 7 is a view showing an example of a DP path when DP matching is used. Here, the DP path means selecting a minimum value when time association is performed. Note that the horizontal axis of FIG. 7 is a feature parameter sequence of speech data based on the first speech input related to the same speech, and the vertical axis is a feature parameter sequence of speech data based on the second speech input related to the same speech. I and j respectively indicate the number of frames.
[0037]
For the first and second utterances related to the same utterance part, time matching (time normalization) is performed using DP matching or the like, and determination is performed using the distance after time normalization. When the distance is extremely small or extremely large, it is rejected as an unnatural utterance. Such a determination can be made more easily by using the result of the DP path.
[0038]
That is, in FIG. 7, in the case of a natural human utterance, as shown in 71, it is possible that local DP path fluctuations occur in the first and second utterances, and the feature parameters completely match. Absent. However, in the case of synthetic speech or the like artificially generated by a speech synthesizer or the like, the feature parameters are the same no matter how many times they are input. Is an exact match. By detecting such an unnatural utterance, it is possible to prevent impersonation.
[0039]
Therefore, as a method for checking whether or not the voice is a synthesized speech or the like, it is possible to check whether or not the checking process for the registered speaker model is the same. FIG. 8 is a block diagram of a speaker verification device according to an embodiment of the present invention to which the matching process identity determination is applied. FIG. 8 is characterized in that a similarity calculation process comparison unit 81 is provided instead of the input speech similarity calculation unit 7.
[0040]
In the similarity calculation process comparison unit 81, for example, if DP matching is used, the DP path is backtraced. If the HMM (Hidden Markov Model) using the Viterbi algorithm is used, the state transition is backtraced. The results are investigated and compared for each input speech. In general, in the case of human utterances, the length and spectrum of utterances are different, so the matching process cannot occur. In the case of synthesized speech or recorded speech, the speech information of the registered speaker Since the matching process with respect to the number of times matches, it becomes possible to detect such illegal input.
[0041]
Then, in the overall judgment unit 8, the similarity between the feature amount of the voice data registered in the individual voice information registration unit 2 calculated by the registered voice similarity calculation unit 6 and the feature amount of the voice data of the input voice And whether or not the speech input comprehensively belongs to the user based on the similarity of the feature amount of the speech data related to the input speech of the same content calculated by the input speech similarity calculation unit 7 To do.
[0042]
First, the similarity between the feature quantity of the voice data registered in the individual voice information registration section 2 calculated by the registered voice similarity calculation section 6 and the feature quantity of the voice data of the input voice is determined from a predetermined threshold value. Is smaller, that is, larger than the threshold I of the collation distance obtained by the method as described above, it is assumed that the range of fluctuation at the time of voice utterance originally possessed by humans is exceeded, and the input voice is It is judged that it is not. When the similarity is equal to or greater than a predetermined threshold, that is, equal to or smaller than the threshold I of the collation distance obtained by the above-described method, it is determined that the human is originally within the range of fluctuation at the time of voice utterance. Move on to the following judgment.
[0043]
Next, the similarity of the feature amount of the voice data related to the input voice having the same content calculated by the input voice similarity calculation unit 7 is equal to or greater than a predetermined threshold, that is, the collation distance threshold obtained by the method as described above. When the value is less than or equal to the value II, it is determined that the input voice is not that of the person himself / herself, as it is an unnatural voice that does not even cause fluctuations at the time of voice utterance inherent in humans. When the similarity is smaller than a predetermined threshold value, that is, larger than the threshold value II of the collation distance obtained by the method as described above, it is determined that the fluctuation at the time of voice utterance inherent in humans has occurred, It is determined that the input voice belongs to the user.
[0044]
Finally, the determination result output unit 9 outputs a determination result as to whether or not the input voice belongs to the user. As an output method, it may be displayed on a display device or the like, or may be passed as a file to an application or the like that operates according to the determination result.
[0045]
Next, artificial speech is subjected to different signal processing in the first and second times to artificially add fluctuations to the synthesized speech, thereby avoiding the above-described synthetic speech rejection conditions. It is also possible to do. In order to prevent such a circumvention, a signal processing unit is provided as a post-process of the voice input unit 4 as shown in FIG.
[0046]
FIG. 9 is a block diagram of a speaker verification apparatus according to an embodiment of the present invention that implements such a method. Compared to FIG. 1, a signal processing unit 91 is added as preprocessing of the voice input unit 4.
[0047]
The signal processing unit 91 does not process all the voices input from the voice input unit 4. At least one of the audio inputs to be compared is subjected to assumed signal processing, and the similarity is actually determined by the method described above for the input audio and the audio after signal processing, so that the characteristic parameters actually match. The input voice that pretends that the synthesized voice is a pseudo-different voice can be detected as the synthesized voice, which can contribute to further improvement in security performance.
[0048]
Also, the voice input environment varies depending on the surrounding conditions such as time and place, and even if the same person inputs the same voice, it is unlikely that the surrounding environment is the same. It can also be used to perform appropriate signal processing on the input voice in order to minimize misidentification of the speaker due to fluctuations.
[0049]
It is also conceivable that the signal processing unit 91 performs normalization processing instead of signal conversion processing. As the normalization processing, it is conceivable to use a CMN method (Cepstral Mean Normalization) performed by subtracting the cepstrum value averaged over the entire speech section from the cepstrum value in each frame. By performing the normalization process, it is possible to treat the input speech that is the target of similarity calculation as speech in the same environment, and it can be expected that the accuracy of the determination is improved. Note that the normalization processing method is not particularly limited to the CMN method.
[0050]
As described above, according to the present embodiment, artificially generated synthesized speech can be identified as having no fluctuation that human speech should originally have. It is possible to prevent an unauthorized third party from impersonating himself / herself using a speech synthesizer or the like.
[0051]
Next, a processing flow of a program that realizes the speaker verification apparatus according to the embodiment of the present invention will be described. FIG. 10 shows a flowchart of the processing of a program that realizes the speaker verification apparatus according to the embodiment of the present invention.
[0052]
In FIG. 10, first, in order to extract voice information registered in advance, the personal ID of the user to be collated is input, and the registered voice information is extracted (step S101).
[0053]
Next, in order to calculate the similarity of the voice, an instruction is given as to what kind of voice is to be input (step S102), and one or more voices that are included at least twice for the same content are input. (Step S103).
[0054]
First, a first similarity between the extracted registered voice and the input voice is calculated (step S104). If the calculated first similarity is smaller than the predetermined threshold (step S105: No), it is determined that the difference is more than the natural fluctuation of human beings, and the voice is by a spoofer. Judgment is made (step S109).
[0055]
Next, when the calculated first similarity is greater than or equal to a predetermined threshold (step S105: Yes), a second similarity between two or more input voices that are input is calculated. (Step S106). If the calculated second similarity is greater than or equal to a predetermined threshold (step S107: Yes), it is determined that the voice is an unnatural voice that does not have even natural fluctuations that humans have, and an impersonator (Step S109).
[0056]
If the calculated second similarity is smaller than a predetermined threshold (step S107: No), it is determined that the input is a natural voice input and is determined to be a voice input by the person ( Step S108).
[0057]
The recording medium storing the program for realizing the speaker verification device according to the embodiment of the present invention is a CD-ROM 112-1, floppy disk 112-2, etc. as shown in the example of the recording medium shown in FIG. In addition to the portable recording medium 112, any other storage device 111 provided at the end of the communication line, or a recording medium 114 such as a hard disk or a RAM of the computer 113 may be used. Runs on main memory.
[0058]
In addition, as shown in the example of the recording medium shown in FIG. 11, the recording medium on which the individual voice information generated by the speaker verification device according to the embodiment of the present invention is recorded is also a CD-ROM 112-1. Not only the portable recording medium 112 such as the disk 112-2 but also any other storage device 111 provided at the end of the communication line, or the recording medium 114 such as the hard disk or RAM of the computer 113, for example, the present invention. It is read by the computer 113 when using such a speaker verification device.
[0059]
【The invention's effect】
As described above, according to the speaker verification device according to the present invention, artificially generated synthesized speech can be identified as having no fluctuation that human speech should originally have, It becomes possible to prevent a third party who has nothing to do with the person from impersonating the person using a speech synthesizer or the like.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of a speaker verification device according to an embodiment of the present invention.
FIG. 2 is a block diagram of a speaker verification device according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating the principle of the speaker verification device according to the embodiment of the present invention.
FIG. 4 is an explanatory diagram of a threshold value concept in the speaker verification device according to the embodiment of the present invention.
FIG. 5 is an explanatory diagram of a threshold value concept in the speaker verification device according to the embodiment of the present invention.
FIG. 6 is a view showing an example of a DP matching result in the speaker verification apparatus according to the embodiment of the present invention.
FIG. 7 is a conceptual explanatory diagram of DP matching in the speaker verification device according to the embodiment of the present invention.
FIG. 8 is a block diagram of a speaker verification device according to an embodiment of the present invention.
FIG. 9 is a block diagram of a speaker verification device according to an embodiment of the present invention.
FIG. 10 is a flowchart of processing in the speaker verification apparatus according to the embodiment of the present invention.
FIG. 11 is an exemplary diagram of a recording medium.
[Explanation of symbols]
1 Personal ID input section
2 Individual voice information registration department
3 Voice input instruction section
4 Voice input part
5 Voice analysis part
6 Input voice similarity calculator
7 Registered voice similarity calculator
8 Comprehensive judgment part
9 Judgment result output part
21 Input voice storage
81 Similarity calculation process comparison unit
91 Signal processor
111 Line destination storage device
112 Portable recording media such as CD-ROM and floppy disk
112-1 CD-ROM
112-2 Floppy disk
113 computer
114 Recording medium such as RAM / hard disk on computer

Claims (6)

A speaker verification device for determining whether or not a voice of a speaker for voice input matches a voice of a registered speaker registered in advance,
A speaker input instruction unit that gives instructions to the speaker about the utterance content to be input;
A voice input unit that includes one or more voices of the speaker, including a voice part of the same content;
A voice analysis unit for analyzing the voice input by the voice input unit;
The similarity between the voice input by the voice input unit and the registered voice of the speaker registered in advance is calculated as the first similarity, and the same one or more of the input voices an input speech similarity calculation unit for calculating a similarity between each other as the second degree of similarity in the two or more audio portion is utterance contents,
When the calculated first similarity is smaller than a predetermined threshold, it is determined that the input is not the voice input of the person, and even if the first similarity is equal to or higher than the predetermined threshold, the second A speaker verification device comprising a determination unit that determines that the input is not the voice input of the person himself / herself when the similarity is equal to or higher than a predetermined level close to perfect match.   The speaker verification apparatus according to claim 1, wherein the second similarity is determined based on whether the verification process for the registered speaker model is the same. When at least two or more voices are input in the voice input unit,
The speaker verification apparatus according to claim 1, wherein normalization processing is performed on at least two or more voices input by the voice input unit. When the calculated first similarity is smaller than a predetermined threshold, the determination unit determines that the voice is an impersonator, and the first similarity is equal to or higher than the predetermined threshold. 2. The speaker verification device according to claim 1 , wherein even if the second similarity is a similarity equal to or higher than a predetermined level close to a perfect match, it is determined that the voice is a spoofer and the input is rejected. A speaker verification method for determining whether or not a voice of a speaker for voice input matches a voice of a registered speaker registered in advance,
Giving instructions to the speaker about the utterance content to be entered;
Inputting one or more voices including the same voice part of the speaker;
Analyzing the input speech;
The similarity between the voice input by the voice input unit and the registered voice of the speaker registered in advance is calculated as the first similarity, and the same one or more of the input voices Calculating a similarity between two or more audio parts that are utterance contents as a second similarity;
When the calculated first similarity is smaller than a predetermined threshold, it is determined that the input is not the voice input of the person, and even if the first similarity is equal to or higher than the predetermined threshold, the second And a step of determining that the input is not the voice input of the person himself / herself when the similarity is equal to or higher than a predetermined level close to perfect match. A computer-readable recording medium that records a program for realizing a speaker verification method for determining whether or not a voice of a speaker to be input matches a voice of a registered speaker registered in advance,
Giving instructions to the speaker about the utterance content to be entered;
Inputting one or more voices including the same voice part of the speaker;
Analyzing the input speech;
The similarity between the voice input by the voice input unit and the registered voice of the speaker registered in advance is calculated as the first similarity, and the same one or more of the input voices Calculating a similarity between two or more audio parts that are utterance contents as a second similarity;
When the first similarity is smaller than a predetermined threshold range, it is determined that the voice input is not the person's own voice. Even when the first similarity is equal to or higher than the predetermined threshold, the second similarity is determined. A computer-readable recording medium on which a program is recorded, which causes a computer to execute a step of determining that the input is not the voice input of the user when the degree of similarity is equal to or higher than a predetermined level close to perfect coincidence.

Images