KR20190131806A - Voice recognition system for detecting spoofing in speaker voice authentication service - Google Patents

Abstract

A voice authentication system for detecting spoofing in a speaker voice authentication service may comprise: a voice synthesizing system including a voice synthesizing unit synthesizing a normal voice signal and a random noise signal of a first speaker and generating a plurality of synthesized voice signals for the first speaker and a voice determination unit receiving a voice signal for learning of the first speaker, comparing the voice signal for learning of the first speaker and the normal voice signal of the first speaker, and determining whether the voice signal for learning of the first speaker is synthesized; and a spoofing detecting system providing a speaker voice authentication service.

Description

VOICE RECOGNITION SYSTEM FOR DETECTING SPOOFING IN SPEAKER VOICE AUTHENTICATION SERVICE}

The present invention relates to a voice authentication system for detecting spoofing in a speaker voice authentication service.

In general, speaker authentication among authentication methods used for system security means a method of recognizing a voice spoken by a user and performing user authentication. In the conventional speaker authentication method, when an authentication requestor requests authentication using his / her own voice, it is determined whether the authentication requestor is authenticated after determining whether or not the voice of the authentication requester matches the voice of the registered user.

On the other hand, as the applicable range of the speaker authentication technology becomes wider and the market size increases, studies to increase the accuracy and reliability of the speaker authentication technology are also being actively conducted. Here, the reliability of the speaker authentication technique refers to the accuracy of spoofing detection.

Spoofing in speaker authentication refers to the act of impersonating a user to forge a voice and then use it to fool the system. The lower the reliability of the system, the more vulnerable to crime. Therefore, high reliability is essential to apply speaker authentication technology to areas where damages from impersonation are high.

Types of spoofing attacks include replay (recording the registered user's voice in advance and playing it upon authentication request), voice modulation (turn any voice into the registered user's voice), and speech synthesis (user's voice in the specific registered text). Generation).

Meanwhile, according to the 'BTAS 2016 speaker anti-spoofing competition', one of the open competitions in the field of spoofing detection, 'd-vector based spoof detection system' has the highest detection accuracy for replay attacks. . However, the devector spoofing detection system has a relatively low detection accuracy for synthesized speech attacks.

This is considered to be because the devector spoofing detection system uses only the prepared voice in detecting the spoofing attack.

Japanese Laid-Open Patent Publication No. 2001-318692 (published November 16, 2001)

The present invention is to solve the above-mentioned problems of the prior art, generating a synthesized speech signal in the speech synthesis system, to determine whether the learning speech signal synthesis, and using the learning speech signal and the evaluation speech signal in the spoofing detection system By learning the spoof detection, we want to provide security-enhanced speaker voice authentication service. However, the technical problem to be achieved by the present embodiment is not limited to the technical problems as described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, the voice authentication system for detecting spoofing in the speaker voice authentication service according to the first aspect of the present invention synthesizes the normal voice signal and the random noise signal of the first speaker to the first Receives a speech synthesizer for generating a plurality of synthesized speech signals for the speaker and the learning speech signal of the first speaker, by comparing the learning speech signal of the first speaker and the normal speech signal of the first speaker, A speech synthesis system including a speech discrimination unit configured to determine whether or not a speaker's learning speech signal is synthesized; And a spoofing detection system for providing the speaker voice authentication service.

A method for detecting spoofing in a speaker voice authentication service according to the second aspect of the present invention generates a plurality of synthesized voice signals for the first speaker by synthesizing a normal voice signal and a random noise signal of the first speaker in a voice synthesis system. Doing; Receiving a learning speech signal of the first speaker in the speech synthesis system; Comparing the speech signal for learning of the first speaker with the normal speech signal of the first speaker in the speech synthesis system to determine whether the speech signal for learning of the first speaker is synthesized and the speaker speech authentication in the spoofing detection system Providing a service may include.

The above-mentioned means for solving the problems are merely exemplary, and should not be construed to limit the present invention. In addition to the exemplary embodiments described above, there may be additional embodiments described in the drawings and detailed description of the invention.

According to any one of the above-described problem solving means of the present invention, the present invention generates a synthesized speech signal in the speech synthesis system, determines whether the learning speech signal is synthesized, and the learning speech signal and the evaluation speech signal in the spoofing detection system By learning the spoofing detection by using can provide a security-enhanced speaker voice authentication service.

In addition, the present invention induces competitive learning between the speech synthesis system and the spoofing detection system so that the speech synthesis system can generate a synthesized speech signal that is similar to the normal speech signal of a more natural speaker, and the spoofing detection system can Even if very similar synthesized speech signals can be detected accurately, the detection accuracy of synthesized speech signals can be improved.

1 is a block diagram of a voice authentication system according to an embodiment of the present invention.
2 is a block diagram of the speech synthesis system shown in FIG. 1, in accordance with an embodiment of the present invention.
3A to 3C are diagrams for describing a voice authentication learning method in a voice synthesis system and a spoofing detection system according to an embodiment of the present invention.
4A to 4C are diagrams for describing a method of improving performance of a speech synthesizer and a spoofing detection system according to whether or not a spoofing is detected in a spoofing detection system according to an embodiment of the present invention.
5 is a diagram for describing a hostile generating network according to an embodiment of the present invention.
6 is a diagram for describing a conditional hostile generating network according to an embodiment of the present invention.
7 is a flowchart illustrating a method for detecting spoofing in a speaker voice authentication service according to an embodiment of the present invention.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element in between. . In addition, when a part is said to "include" a certain component, which means that it may further include other components, except to exclude other components unless otherwise stated.

In the present specification, the term 'unit' includes a unit realized by hardware, a unit realized by software, and a unit realized by both. In addition, one unit may be realized using two or more pieces of hardware, and two or more units may be realized by one piece of hardware.

Some of the operations or functions described as being performed by a terminal or a device in the present specification may instead be performed in a server connected to the terminal or device. Similarly, some of the operations or functions described as being performed by the server may be performed by a terminal or a device connected to the server.

Hereinafter, with reference to the accompanying configuration diagram or processing flow chart, it will be described in detail for the practice of the present invention.

1 is a block diagram of a voice authentication system according to an embodiment of the present invention.

Referring to FIG. 1, a voice authentication system may include a voice synthesis system 100, a spoofing detection system 110, and a voice database 120. However, since the voice authentication system of FIG. 1 is only one embodiment of the present invention, the present invention is not limitedly interpreted through FIG. 1 and may be configured differently from FIG. 1 according to various embodiments of the present invention.

Hereinafter, the speech synthesis system 100 will be described with reference to FIGS. 1 and 2.

The speech synthesis system 100 may include a speech synthesizer 200 and a speech discriminator 210.

The speech synthesizer 200 may generate a plurality of synthesized speech signals for the first speaker by synthesizing the normal speech signal and the random noise signal of the first speaker.

In this case, the random noise signal may correspond to a pseudo random number generated by a random number generation algorithm. At this time, the random number generation algorithm is an algorithm for generating a pseudo random number using the input initial value. The initial value input to the random number generation algorithm is called a seed value. If the seed values are the same, pseudo random numbers generated by the random number generation algorithm may be the same.

Here, the plurality of synthesized speech signals for the first speaker may include the synthesized speech signal for learning of the first speaker and the synthesized speech signal for evaluation.

The voice discriminating unit 210 may receive the first voice's learning voice signal and compare the first voice's learning voice signal with the first speaker's normal voice signal to determine whether the first speaker's learning voice signal is synthesized. have.

Here, the learning voice signal of the first speaker may be a learning normal voice signal or a learning synthesized voice signal of the first speaker.

For example, the voice discriminator 210 may compare the normal voice signal of the first speaker and the normal voice signal of the first speaker to determine whether the first normal voice signal for learning of the first speaker is a normal voice signal.

For example, the speech discriminating unit 210 may determine whether the first synthesized speech signal for learning by the first speaker is a synthesized signal by comparing the first synthesized speech signal for the first speaker with the normal speech signal of the first speaker.

The speech synthesizer 200 may perform a learning to generate a synthesized speech signal similar to the normal speech signal of the first speaker based on a hostile generation network. Here, the hostile generation network is a network to which competitive learning is applied to the deep neural network.

The voice discriminating unit 210 may perform learning to distinguish the first voice's learning voice signal from the first speaker's normal voice signal based on the hostile generating network.

In detail, the speech synthesizer 200 and the speech discriminator 210 may perform competitive learning based on a neural network model.

로부터 표본 샘플을 샘플링하여 얻은 변수로서 일종의 랜덤 시드(randon seed)의 역할을 수행할 수 있다. For example, when the speech synthesis unit 200 receives the random noise signal z, the speech synthesis unit 200 maps the random noise signal z to a data space to obtain a probability value G (z) corresponding to the synthesized speech signal. You can print Here, the random noise signal z may be of any prior distribution.

As a variable obtained by sampling a sample sample from, it may serve as a kind of random seed.

The speech discriminating unit 210 may output a probability value D (x) for synthesizing upon receiving a speech signal x for determining whether to synthesize. Here, the probability value D (x) as to whether or not synthesis is a probability value indicating that the speech signal x is a normal speech signal may be determined within a preset probability range (for example, [0, 1]).

For example, as the probability value D (x) for synthesis is closer to the first probability value (for example, 1), the speech signal x is closer to the normal speech signal, and the probability value D (x) for synthesis or not is determined. The closer to the second probability value (eg, 0), the more implied that the speech signal x is closer to the composite data.

The speech synthesizer 200 and the speech discriminator 210 may simultaneously perform competitive learning. In detail, the speech synthesizer 200 may learn to minimize the objective function for the hostile generating network, and the speech discriminator 210 may learn to maximize the objective function for the hostile generating network. Here, the objective function for the hostile generation network may be expressed as shown in [Equation 1].

[Equation 1]

The speech discriminating unit 210 increases the probability of distinguishing the learning normal speech signal from the learning synthesized speech signal as the learning speech signal of the first speaker so that the speech discriminating unit 210 maximizes the objective function for the hostile generating network. To learn.

)에 있어서의 합성 여부에 대한 확률값(D(x))이 제 1 확률값에 가까워지도록 매개 변수 x를 조정할 경우, 적대적 생성 네트워크에 대한 목적 함수를 최대화할 수 있다. 여기서, 합성 여부에 대한 확률값(D(x))이 제 1 확률값에 가까워지도록 매개 변수 x를 조정한다는 의미는 정상 음성 신호를 정상 음성 신호로서 판단하도록 학습한다는 것을 의미할 수 있다. For example, the normal voice signal for training (

When the parameter x is adjusted so that the probability value D (x) for the synthesis in the step 1) is close to the first probability value, the objective function for the hostile generation network can be maximized. Here, the adjustment of the parameter x such that the probability value D (x) for synthesis is closer to the first probability value may mean that the learner determines to determine the normal speech signal as a normal speech signal.

)로부터 생성된 합성 음성 신호(G(z))에 있어서의 합성 여부에 대한 확률값(D(G(z)))이 제 2 확률값에 가까워지도록 매개변수 G(z)를 조정할 경우, 적대적 생성 네트워크에 대한 목적 함수를 최대화할 수 있다. 여기서, 합성 여부에 대한 확률값 (D(G(z)))이 제 2 확률값에 가까워지도록 매개변수 G(z)를 조정한다는 의미는 합성 음성 신호를 합성 음성 신호로서 판단하도록 학습한다는 것을 의미할 수 있다. In addition, the random noise signal (

When the parameter G (z) is adjusted so that the probability value D (G (z)) for synthesis in the synthesized speech signal G (z) generated from the value is close to the second probability value, the hostile generation network You can maximize the objective function for. Here, the adjustment of the parameter G (z) such that the probability value D (G (z)) as to whether the synthesis is close to the second probability value may mean that the training is performed to determine the synthesized speech signal as the synthesized speech signal. have.

On the other hand, the speech discriminator 210 may learn to increase the probability of synthesizing the first speech learner speech signal so that the speech synthesizer 200 minimizes the objective function for the hostile generating network.

For example, the probability synthesizer 200 determines whether the speech synthesizer 200 synthesizes the synthesized speech signal G (z) so that the speech discriminator 210 incorrectly judges the synthesized speech signal as a normal speech signal. Adjusting the parameter G (z) to bring (z))) closer to the first probability value, the objective function for the hostile generation network can be minimized.

Referring to FIG. 5, a method of performing learning by the voice synthesizer 200 and the voice discriminator 210 using a hostile generation network will be described. 5 is a diagram illustrating a learning process of a hostile generating network for each step 500, 510, 520, and 530.

)의 분포(

)를 나타내고, 제 2 선(55, 초록색 실선)은 합성 음성 신호(G(z))의 분포(

)를 나타내고, 제 3 선(53, 파란색 점선)은 음성 판별부(210)에서 산출된 합성 여부에 대한 확률값(D(x))의 분포를 나타낸다. x에서 z로 향하는 화살표는 z를 x=G(z)로 매핑하는 것을 의미한다. 각 단계에서 x에서 z로 향하는 화살표의 끝이 합성 음성 신호(G(z))의 분포(

)의 값이 큰 쪽으로 몰려있는 것을 확인할 수 있다. Referring to FIG. 5, the first line 51 (solid black line) represents a normal voice signal (

) Distribution

), And the second line 55 (solid green line) shows the distribution (of the synthesized speech signal G (z) (

), And a third line 53 (blue dotted line) represents a distribution of probability values D (x) for synthesis or not calculated by the voice discriminator 210. An arrow from x to z means map z to x = G (z). At each step, the end of the arrow from x to z indicates the distribution of the synthesized speech signal G (z) (

You can see that the value of) is in the larger direction.

)와 합성 음성 신호의 분포(

)가 차이를 보이므로 음성 판별부(210)는 정상 음성 신호와 합성 음성 신호를 어렵지 않게 분류할 수 있다. In the first step 500, the distribution of the normal speech signal (

) And the distribution of synthesized speech signals (

), The voice discrimination unit 210 can easily classify the normal voice signal and the synthesized voice signal.

)와 합성 음성 신호의 분포(

)를 잘 구분하도록

에 수렵하도록 학습할 수 있다. 이 때,

는 전역 최적 해를 의미한다. 음성 판별부(210)가 정상 음성 신호 및 합성 음성 신호에 대한 분류 정확도가 높은 경우는

일 때 일 수 있다. In a second step 510, the voice discriminating unit 210 distributes a normal voice signal (

) And the distribution of synthesized speech signals (

)

You can learn to hunt. At this time,

Means the global optimal solution. When the speech discriminating unit 210 has a high classification accuracy for the normal speech signal and the synthesized speech signal,

Can be when

)가 정상 음성 신호의 분포(

)에 보다 가까워지도록 학습할 수 있다. In the third step 520, the speech synthesizer 200 distributes the synthesized speech signal so that it is difficult for the speech discriminator 210 to distinguish the normal speech signal from the synthesized speech signal.

) Is the distribution of normal speech signals (

You can learn to get closer to).

)와 합성 음성 신호의 분포(

)는 동일한 분포를 보인다. 이 경우, 음성 판별부(210)는 정상 음성 신호와 합성 음성 신호를 구분하기 어려우므로 어느 구간에서든 합성 여부에 대한 확률값(D(x))은 0.5로 출력될 수 있다. In a fourth step 530, the distribution of the normal voice signal

) And the distribution of synthesized speech signals (

) Shows the same distribution. In this case, since it is difficult to distinguish the normal speech signal from the synthesized speech signal, the speech discriminating unit 210 may output 0.5 as a probability value D (x) for synthesis in any section.

1 and 2, the speech synthesizer 200 and the speech discriminator 210 may perform learning based on a conditional hostile generating network. Here, the conditional hostile generation network is a concept in which the hostile generation network is extended to receive the conditions 301 and 309 for each of the voice synthesis unit 200 and the voice discriminator 210 as shown in FIG. 6.

The objective function for the conditional hostile generation network can be expressed as shown in [Equation 2]. That is, the condition y may be added to the objective function of the hostile generating network.

[Equation 2]

Referring to FIG. 3A for a while, the speech synthesizer 200 may receive a condition 301 for the normal speech signal of the first speaker. Here, the condition 301 is registered by the first speaker for use as information about the first speaker and text information about a normal voice signal received from the first speaker (e.g., voice authentication of the user for payment or security, etc.). Text information).

The speech synthesizer 200 uses the condition 301 for the normal speech signal of the first speaker, the normal speech signal 303 of the first speaker, and the random noise signal 305 to synthesize the speech signal for the first speaker ( 307 may be generated. In this case, the normal voice signal 303 of the first speaker may correspond to a voice signal (eg, “hello”) corresponding to the text information (eg, “hello”) included in the condition 301.

Referring to FIG. 6 for a while, the synthesized speech signal for the first speaker that can be generated by the speech synthesizer 200 with respect to the condition 301 for the normal speech signal of the first speaker input to the speech synthesizer 200 is random. It may be equal to the number of times the noise signal 600 is sampled. That is, even if only one condition is input to the speech synthesis unit 200, a plurality of synthesized speech signals having the same condition may be generated.

Referring again to FIG. 3A, the speech synthesis unit 200 may generate the normal speech signal 303 and the random noise signal 305 of the first speaker based on the condition 310 for the normal speech signal of the first speaker. The synthesized speech signal corresponding to the speech signal or the text information similar to that spoken by the first speaker can be generated.

을 입력으로 받고, 해당 구간 [t, t+u-1]의 바로 다음 시간인 t+u에서 표본 값을 출력한다.Meanwhile, the speech synthesizer 200 may generate a wavenet for generating a speech waveform as a synthesized speech signal. Here, the wave net corresponds to a sample sequence of length u corresponding to an arbitrary interval [t, t + u-1].

Is inputted, and the sample value is output at t + u immediately after the corresponding interval [t, t + u-1].

라고 가정하면, 웨이브 넷은 길이 u의 학습용 음성 신호의 표본열을 입력으로 받는다. At this time, the sample string may be obtained from a specific voice or set to an arbitrary voice. The wave net may perform learning using quantized sample sequences of speech. For example, quantum samples of a learning speech signal

In this example, the wave net receives a sample sequence of a learning speech signal of length u as an input.

을 입력으로 받고, 웨이브 넷은 시간 u+1에서의 예상 표본 값(

)을 생성할 수 있다. 그 다음에는 [2, u+1] 구간에 해당하는 표본 열

로부터

을 생성하고,

와

의 차이가 줄어들도록 학습할 수 있다. 이 과정을 순차적으로 반복하면, 1개의 음성 신호에 대한 학습이 완료된다. 이러한 학습 과정을 목적 함수로 나타내면 [수학식 3]과 같이 나타낼 수 있다. For example, the sample column initially corresponds to the interval [1, u].

As the input, and the wavenet returns the expected sample value at time u + 1 (

) Can be created. Next, the sample column corresponding to the interval [2, u + 1]

from

Creates a,

Wow

You can learn to reduce the difference. If this process is repeated sequentially, the learning of one voice signal is completed. If the learning process is expressed as an objective function, it can be expressed as shown in [Equation 3].

[Equation 3]

는

와

를 나타내는 함수이며, 교차 엔트로피(cross-entropy)를 주로 사용한다. here,

Is

Wow

This function indicates that cross-entropy is used.

3B and 6, the voice discriminating unit 210 may receive a condition 309 for the learning voice signal of the first speaker. Here, the condition may include information on the first speaker and text information on the voice signal to be determined.

The voice discriminating unit 210 uses the condition 309 for the learning voice signal of the first speaker, the learning voice signal 311 of the first speaker, and the normal voice signal 303 of the first speaker for learning of the first speaker. Whether the voice signal 311 is synthesized or not may be determined.

For example, the voice discrimination unit 210 has the same voice as the learning voice signal 311 of the first speaker and the normal voice signal 303 of the first speaker as the condition 309 for the learning voice signal of the first speaker. Learning to distinguish whether it is a signal can be performed.

For example, the voice discriminator 210 determines that the voice of the first speaker's normal voice signal and the first speaker's normal voice signal 303 are equal to each other based on the condition 309 of the first speaker's learning voice signal. If it is determined as a signal, a first probability value (eg, 1) corresponding to a normal voice signal may be output.

As another example, when the voice signal of the same speaker as the condition 309 for the learning voice signal of the first speaker is input, the voice determination unit 210 may output a first probability value. If it is determined that the first speaker's learning normal voice signal and the first speaker's normal voice signal 303 are different voice signals, the voice discriminator 210 determines the first speaker's learning normal voice signal and the first speaker. It is possible to learn that the speaker's normal voice signal 303 is recognized as the same voice signal.

 For example, the voice discriminator 210 determines that the voice of the first speaker's learning synthesized voice signal and the first speaker's normal voice signal 303 are different from each other based on a condition 309 for the first speaker's learning voice signal. If it is determined as a signal, a second probability value (eg, 0) corresponding to the synthesized speech signal may be output.

As another example, when the voice signal of the speaker different from the condition 309 for the learning voice signal of the first speaker is input, the voice determination unit 210 may output a second probability value (eg, 0). If it is determined that the first synthesizer speech signal for learning of the first speaker and the normal speech signal 303 of the first speaker are the same voice signals, the voice discriminator 210 determines the first normal learner speech signal for the first speaker and the first speaker. The speaker's normal speech signal 303 can be learned to be distinguished.

The present invention trains the speech synthesizer 200 so that the output value of the speech discriminator 210 approaches the first probability value with respect to the synthesized speech signal to the first speaker. That is, the speech synthesizer 200 learns to generate a synthesized speech signal similar to the normal speech signal of the first speaker, and the speech discriminator 210 accurately distinguishes the normal speech signal of the first speaker from the synthesized speech signal. When learning is performed, even if a synthesized speech signal similar to the normal speech signal of the first speaker is input, the synthesized speech signal can be accurately distinguished.

Referring to FIG. 3C for a while, the spoofing detection system 110 receives a training voice signal 315 of a first speaker and detects whether to spoof the training voice signal 315 of the first speaker. ) Can be performed.

For example, the spoofing detection system 110 may output a first probability value corresponding to the normal voice signal when the input first learning normal voice signal is detected as an unsynthesized normal voice signal. If a second probability value corresponding to the synthesized speech signal is output with respect to the received first speaker's first normal speech signal, the spoofing detection system 110 determines that the first speaker's first normal speech signal for the first speaker is the first speech signal. Learn to recognize the speaker's normal voice signal.

In addition, when the received first learning synthesis speech signal is detected as a synthesized speech signal, the spoofing detection system 110 may output a second probability value corresponding to the synthesized speech signal. If the first probability value corresponding to the normal speech signal is output with respect to the received first speech synthesis speech signal of the first speaker, the spoofing detection system 110 may generate the first speech synthesis speech signal and the first speech signal of the first speaker. Learning to distinguish the speaker's normal speech signal can be performed.

On the other hand, the spoofing detection system 110 may evaluate the detection accuracy of the synthesis of the evaluation speech signal using the evaluation speech signal. Here, the evaluation voice signal may include a plurality of normal voice signals for evaluation and a plurality of synthesized voice signals for evaluation.

For example, the spoofing detection system 110 verifies whether the first detection value corresponding to the normal voice signal is output to the at least one evaluation normal voice signal, and synthesizes the synthesized voice signal with respect to the at least one evaluation synthesized voice signal. It may be verified whether a second detection value corresponding to the output is output.

4A to 4C, when the spoofing detection system 110 outputs the first detection value 403 corresponding to the normal speech signal of the first speaker to the at least one evaluation synthesized speech signal 401. The spoofing detection system 110 may lower the detection accuracy of the corresponding synthesis speech signal 401.

When the spoofing detection system 110 determines the at least one evaluation synthesized speech signal 401 as the normal voice signal of the first speaker, the spoofing detection system 110 determines that at least one of the spoofing detection system 110 is determined as the normal voice signal of the first speaker. Re-learning may be performed to distinguish the synthesized speech signal 401 for evaluation from the normal speech signal of the first speaker.

When the spoofing detection system 110 outputs a second detection value 407 corresponding to the synthesized speech signal with respect to the at least one evaluation normal speech signal 405, the spoofing detection system 110 determines the corresponding normal speech for evaluation. The detection accuracy for the signal 405 can be evaluated low.

When the spoofing detection system 110 determines that the at least one evaluation normal speech signal 405 is a synthesized speech signal, the spoofing detection system 110 determines at least one evaluation normal speech signal ( Re-learning may be performed to recognize 405 as a normal voice signal of the first speaker.

When the spoofing detection system 110 outputs the second detection value 411 corresponding to the synthesized speech signal with respect to the at least one evaluation synthesized speech signal 409, the spoofing detection system 110 determines the corresponding synthesized speech for evaluation. The detection accuracy for the signal 409 can be highly evaluated.

When the spoofing detection system 110 determines that the at least one evaluation synthesized speech signal 409 is a synthesized speech signal, the speech synthesizer 200 learns to reproduce a synthesized speech signal similar to the normal speech signal of the first speaker. can do.

As such, the speech synthesis system 100 learns to generate a synthesized speech signal similar to the normal speech signal of the first speaker to the extent that it is difficult for the spoofing detection system 110 to accurately detect whether the speech is synthesized, and the spoofing detection system 110 Learning to accurately distinguish the synthesized speech signal from the normal speech signal of the first speaker, the speech synthesis system 100 can generate a synthesized speech signal that is substantially similar to the normal speech signal of the first speaker, and the spoofing detection system 110 Can provide a speaker voice authentication service that accurately detects whether or not a synthesized voice signal is substantially similar to the normal voice signal of the first speaker.

Referring to FIG. 1, the voice database 120 includes a first voice database 122 for storing a plurality of learning voice signals for each of the plurality of speakers and a plurality of evaluation voice signals for each of the plurality of speakers. Two voice databases 124. For example, the first voice database 122 may store a plurality of learning voice signals including a plurality of learning normal voice signals of the first speaker and a plurality of learning synthesis voice signals. The second voice database 124 may store a plurality of evaluation voice signals including the plurality of evaluation normal voice signals of the first speaker and the plurality of evaluation synthesized voice signals.

7 is a flowchart illustrating a method for detecting spoofing in a speaker voice authentication service according to an embodiment of the present invention.

The spoofing detection method according to the embodiment shown in FIG. 7 includes steps that are processed in time series in the speech synthesis system 100 and the spoofing detection system 110 according to the embodiment shown in FIGS. 1 to 6. Therefore, although omitted below, the contents described with respect to the speech synthesis system 100 and the spoofing detection system 110 of FIGS. 1 to 6 may also be applied to the spoofing detection method according to the embodiment shown in FIG. 7.

Referring to FIG. 7, in operation S701, the speech synthesis system 100 may generate a plurality of synthesized speech signals for the first speaker by synthesizing the normal speech signal and the random noise signal of the first speaker.

In operation S703, the speech synthesis system 100 may receive a learning speech signal of the first speaker.

In operation S705, the speech synthesis system 100 may compare the speech signal for learning of the first speaker with the normal speech signal of the first speaker to determine whether the speech signal for learning of the first speaker is synthesized.

In operation S707, the spoofing detection system 110 may provide a speaker voice authentication service.

In the above description, steps S701 to S707 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, and the order between the steps may be changed.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by the computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The foregoing description of the present invention is intended for illustration, and it will be understood by those skilled in the art that the present invention may be easily modified in other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is shown by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention. .

100: speech synthesis system
110: spoofing detection system
120: voice database
200: speech synthesis unit
210: voice discriminating unit

Claims (14)

In a voice authentication system for detecting spoofing in a speaker voice authentication service,
Receives a speech synthesizer for synthesizing a normal speech signal and a random noise signal of the first speaker to generate a plurality of synthesized speech signals for the first speaker and a learning speech signal of the first speaker, and for learning the first speaker A speech synthesizing system including a speech discriminating unit configured to compare a speech signal with a normal speech signal of the first speaker and determine whether to synthesize the speech signal for learning of the first speaker; And
Spoofing detection system for providing the speaker voice authentication service
Voice authentication system comprising a.
The method of claim 1,
The speech synthesis unit performs learning to generate a synthesized speech signal similar to the normal speech signal of the first speaker based on the hostile generating network,
And the voice discriminating unit performs learning for distinguishing the learning voice signal of the first speaker from the normal voice signal of the first speaker based on the hostile generating network.
The method of claim 2,
The speech synthesizer learns to minimize the objective function for the hostile generating network,
And the speech discriminating unit learns to maximize the objective function.
The method of claim 3, wherein
And the speech synthesizer learns to increase the probability of synthesizing the learning speech signal of the first speaker so that the speech synthesizer minimizes the objective function.
The method of claim 4, wherein
And the speech discriminating unit learns to increase a probability of distinguishing a learning normal speech signal from a learning synthesized speech signal as the learning speech signal of the first speaker so that the speech discriminating unit maximizes the objective function.
The method of claim 1,
The synthesized speech signal includes a synthesized speech signal for learning and a synthesized speech signal for evaluation,
The learning speech signal of the first speaker is a learning normal speech signal or the learning synthesized speech signal of the first speaker.
The method of claim 1,
And the speech synthesizing unit and the speech discriminating unit perform learning based on a conditional hostile generating network.
The method of claim 7, wherein
The speech synthesizer receives a condition for the normal speech signal of the first speaker,
And the voice discrimination unit receives a condition for the learning voice signal of the first speaker.
The method of claim 8,
Wherein the condition includes information about the speaker and textual information about the received voice signal.
The method of claim 1,
The spoofing detection system receives the learning voice signal of the first speaker,
And learning to detect whether to spoof the received speech signal for learning of the first speaker.
The method of claim 6,
And when the spoofing detection system determines the evaluation synthesized speech signal as a synthesized speech signal, the speech synthesizer learns to reproduce a synthesized speech signal similar to the normal speech signal of the first speaker.
The method of claim 11,
When the spoofing detection system determines that at least one evaluation normal voice signal of the evaluation normal voice signal is a synthesized voice signal, the spoofing detection system determines at least one evaluation normal voice signal determined as the synthesized voice signal. Re-learning to recognize as a normal voice signal of the first speaker.
The method of claim 11,
The spoofing detection system determines that at least one evaluation synthesized speech signal of the evaluation synthesized speech signal is a normal speech signal of the first speaker;
And the spoofing detection system performs relearning to distinguish at least one evaluation synthesized speech signal determined as the normal speech signal of the first speaker from the normal speech signal of the first speaker.
A voice synthesis system for synthesizing a speaker's voice used in a spoofing detection system that provides a speaker voice authentication service,
A speech synthesizer configured to synthesize a normal speech signal and a random noise signal of a first speaker to generate a plurality of synthesized speech signals for the first speaker; And
A voice discriminating unit configured to receive a learning voice signal of the first speaker, compare the learning voice signal of the first speaker with a normal voice signal of the first speaker, and determine whether to synthesize the learning voice signal of the first speaker;
Speech synthesis system comprising a.

Images