Voiceprint authentication method and system based on StarGAN-VC

Abstract

The application provides a voiceprint authentication method and system based on StarGAN-VC, and relates to the field of voiceprint authentication. The method comprises the following steps: by taking the original registration audio as a training sample, iteratively training a generator network in a StarGAN-VC network architecture, and based on different iteration training times, obtaining a first generator network and a second generator network respectively. And by using the second generator network to transform the original registration voice signal into a first registration audio signal, the first registration audio signal and the original registration voice signal have certain differences, which guarantees the security of the voiceprint privacy of the original registration voice signal. Since the generator in the StarGAN-VC network architecture cannot know the specific transformation means in the transformation process of the voice signal, the confidentiality of the voice signal transformation is improved, and the voiceprint privacy security of the original voiceprint feature information of the user is reliably guaranteed.

Description

Technical Field

[0001] This invention relates to the field of voiceprint authentication technology, specifically to a voiceprint authentication method and system based on StarGAN-VC. Background Technology

[0002] With the development of information technology, higher requirements have been placed on trusted identity authentication capabilities. In February 2020, the People's Bank of China issued the "Technical Specifications for the Protection of Personal Financial Information," which for the first time separated "dynamic voiceprint passwords" from "personal biometric information" and listed them alongside dynamic passwords. This recognition by the industry as a type of personal information with lower privacy sensitivity demonstrates the unique advantages and broad prospects of voiceprints in personal privacy protection.

[0003] Existing voiceprint authentication methods emphasize the accuracy of recognition results and the efficiency of the authentication process. However, with the widespread application of voiceprint authentication, users' voice data has become a target for criminals. Specifically, users' voice data is often stored directly in the cloud without processing or with simple encryption. If attackers gain access to voice information collection devices such as mobile phones and self-service terminals, they can obtain or simply decrypt users' voice data and then directly replay the user's voiceprint features at the message level.

[0004] In summary, existing voiceprint authentication methods either do not encrypt the user's original voice or use relatively simple encryption techniques, resulting in low privacy and security of the user's original voiceprint feature information. Therefore, there is an urgent need for a voiceprint authentication method to solve the above problems. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a voiceprint authentication method and system based on StarGAN-VC, which solves the problem that the privacy and security of the user's original voiceprint feature information are low because existing voiceprint authentication methods do not encrypt the user's original voice or the encryption technology is too simple.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect of the present invention, a voiceprint authentication method based on StarGAN-VC is provided, the method comprising:

[0010] Obtain a first authentication voice signal; wherein the first authentication voice signal is obtained based on preprocessing of the original authentication voice signal;

[0011] Based on the pre-acquired first generator network, the first authentication speech signal is processed to obtain the second authentication speech signal; wherein, the acquisition process of the first generator network includes: using the original registration speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample; when the number of iterations reaches a first preset value, the corresponding generator network is acquired as the first generator network.

[0012] The similarity between the second certified speech signal and the first registered speech signal is calculated. If the similarity is greater than a first preset threshold, the original certified speech signal is successfully certified. The first registered speech signal is obtained based on the original registered speech signal and a pre-acquired second generator network. The process of acquiring the second generator network includes: using the original registered speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample. When the number of iterations reaches a second preset value, the corresponding generator network is acquired as the second generator network. The second preset value is less than the first preset value.

[0013] Optionally, the method further includes:

[0014] Obtain the original authentication voice signal;

[0015] The Mel-frequency cepstral coefficients of the original authentication voice signal are obtained and used as the first authentication voice signal.

[0016] Optionally, the processing of the first authentication voice signal based on the pre-acquired first generator network to obtain the second authentication voice signal includes:

[0017] The first authentication voice signal and the target voice tag are input together into a pre-acquired first generator network to obtain the signal output by the first generator network, which serves as the third authentication voice signal; wherein, the target voice tag is the voice tag of the pre-selected target voice signal;

[0018] The third authentication voice signal and the original authentication voice tag are input together into the first generator network to obtain the signal output by the first generator network, which serves as the fourth authentication voice signal; wherein, the original authentication voice tag is the voice tag of the original authentication voice signal.

[0019] Based on the third and fourth authentication voice signals, a first spectral gain function is obtained;

[0020] The spectral envelope of the original authenticated speech is multiplied by the first spectral gain function, and the multiplication result is processed based on the vocoder to obtain the second authenticated speech signal.

[0021] Optionally, the process of iteratively training the generator network in the StarGAN-VC network architecture using the original registered speech signal as training samples includes:

[0022] The Mel frequency cepstral coefficients of the original registered speech signal are obtained as the speech signal to be processed;

[0023] The speech signal to be processed and the target audio tag are input together into the generator network in the StarGAN-VC network architecture to be trained to obtain the first synthesized speech signal;

[0024] The first synthesized speech signal is input into a pre-trained discriminator network to obtain the probability that the first synthesized speech signal is the target speech signal; and the first synthesized speech signal and the original audio tag are input together into the generator network in the StarGAN-VC network architecture to be trained to obtain the reconstructed speech signal.

[0025] The loss value of the generator network is calculated based on the probability that the first synthesized speech signal is the target speech signal.

[0026] The network parameters of the generator network are adjusted based on the loss value of the generator network; wherein, after the network parameters of the generator network are adjusted, it is recorded as the completion of one iteration of training;

[0027] The reconstructed speech signal replaces the speech signal to be processed, and the replaced speech signal to be processed and the target audio tag are re-inputted into the generator network after adjusting the network parameters to obtain the first synthesized speech signal, and iterative training continues.

[0028] Optionally, before calculating the similarity between the second authenticated voice signal and the first registered voice signal, the method further includes:

[0029] Obtain a second registered voice signal; wherein the second registered voice signal is obtained by encrypting the first registered voice signal;

[0030] The second registered voice signal is decrypted to obtain the first registered voice signal.

[0031] Optionally, before acquiring the second registered voice signal, the method further includes:

[0032] Obtain the original registered voice signal;

[0033] The Mel-frequency cepstral coefficients of the original registered speech signal are obtained as the registered speech signal to be processed;

[0034] The registered speech signal to be processed and the target speech label are input into the pre-trained second generator network to obtain the signal output by the second generator network, which is used as the third registered speech signal;

[0035] The third registered speech signal and the original registered speech tag are input into the second generator network to obtain the signal output by the second generator network, which is used as the fourth registered speech signal; wherein, the original registered speech tag is the speech tag of the original registered speech signal;

[0036] Based on the third registered speech signal and the fourth registered speech signal, obtain the second spectral gain function;

[0037] The spectral envelope of the original registered speech is multiplied by the second spectral gain function, and the multiplication result is processed based on the vocoder to obtain the first registered speech signal;

[0038] The first registered voice signal is encrypted to obtain the second registered voice signal.

[0039] Optionally, the first registered voice signal is encrypted to obtain a second registered voice signal, including:

[0040] The original registered voice signal was deleted;

[0041] The first registered voice signal is encrypted using the AES encryption library to obtain the encrypted first registered voice signal, which is then used as the second registered voice signal.

[0042] Optionally, after encrypting the first registered voice signal using an AES encryption library to obtain an encrypted first registered voice signal, which is then used as the second registered voice signal, the method further includes:

[0043] The second registered voice signal is stored in the database, and the user password is deleted; wherein the user password is the same as the key used for encryption based on the AES encryption library.

[0044] Optionally, before deleting the user's password, the method further includes:

[0045] Hash the user password to obtain the corresponding hash value, which is used as the user hash value;

[0046] Before acquiring the original authentication voice signal, the method further includes:

[0047] Get the authentication password;

[0048] The authentication password is hashed to obtain the corresponding hash value, which is used as the authentication hash value;

[0049] Determine whether the authentication hash value is the same as the user hash value; if so, it indicates that the authentication password is successful.

[0050] The step of decrypting the second registered voice signal to obtain the first registered voice signal includes:

[0051] Obtain the authentication password;

[0052] The second registered voice signal is decrypted based on the authentication password to obtain the first registered voice signal.

[0053] In a second aspect of the invention, a voiceprint authentication system based on StarGAN-VC is provided, the system comprising:

[0054] The signal acquisition module is used to acquire a first authentication voice signal; wherein the first authentication voice signal is obtained based on preprocessing of the original authentication voice signal;

[0055] The signal processing module is used to process the first authentication speech signal based on the pre-acquired first generator network to obtain the second authentication speech signal; wherein, the acquisition process of the first generator network is as follows: the original registration speech signal is used as a training sample, and the generator network in the StarGAN-VC network architecture is iteratively trained based on the training sample. When the number of iterations reaches a first preset value, the corresponding generator network is acquired as the first generator network.

[0056] A similarity calculation module is used to calculate the similarity between the second certified speech signal and the first registered speech signal. If the similarity is greater than a first preset threshold, it indicates that the original certified speech signal has been successfully certified. The first registered speech signal is obtained based on the original registered speech signal and a pre-acquired second generator network. The generation process of the second generator network is as follows: the original registered speech signal is used as a training sample, and the generator network in the StarGAN-VC network architecture is iteratively trained based on the training sample. When the number of iterations reaches a second preset value, the corresponding generator network is obtained as the second generator network. The second preset value is less than the first preset value.

[0057] (III) Beneficial Effects

[0058] This invention provides a voiceprint authentication method and system based on StarGAN-VC. Compared with existing technologies, it has the following advantages:

[0059] This invention provides a voiceprint authentication method based on StarGAN-VC. The method includes: acquiring a first authentication voice signal; wherein the first authentication voice signal is obtained by preprocessing an original authentication voice signal; processing the first authentication voice signal based on a pre-acquired first generator network to obtain a second authentication voice signal; wherein the acquisition process of the first generator network includes: using the original registration voice signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample, and when the number of iterations reaches a first preset value, acquiring the corresponding generator network as the first generator network; calculating the similarity between the second authentication voice signal and the first registration voice signal, and if the similarity is greater than a first preset threshold, it indicates that the original authentication voice signal has been successfully authenticated; wherein the first registration voice signal is obtained based on the original registration voice signal and the pre-acquired second generator network; the generation process of the second generator network is: using the original registration voice signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample, and when the number of iterations reaches a second preset value, acquiring the corresponding generator network as the second generator network; the second preset value is less than the first preset value.

[0060] Based on the above processing, this invention uses the original registration audio as training samples to iteratively train the generator network in the StarGAN-VC network architecture. Based on different iterations, a first generator network and a second generator network are obtained. The second generator network then transforms the original registration speech signal into a first registration audio signal, ensuring a difference between the first and original registration speech signals and protecting the voiceprint privacy of the original registration speech signal. Furthermore, the first generator network transforms the original authentication speech signal into a second authentication speech signal, effectively protecting the voiceprint privacy of the original authentication speech signal as well. Because the specific transformation methods used by the generator in the StarGAN-VC network architecture to transform the speech signal are unknown in the prior art, the confidentiality of the speech signal transformation is greatly improved. Therefore, even if the first registration speech signal is obtained, the prior art cannot reverse-engineer it back into the original registration speech signal, providing reliable protection for the voiceprint privacy of the user's original voiceprint feature information.

[0061] Furthermore, since the first generator network and the second generator network are trained based on the same training samples and the same generator in the StarGAN-VC network architecture, with only the number of training iterations differing, the conversion process of the original registered speech signal is basically the same as that of the original certified speech signal. Therefore, by calculating the similarity between the second certified speech signal and the first registered speech signal, the voiceprint authentication of the original certified speech signal can be completed, thereby further ensuring the accuracy of voiceprint authentication while protecting the privacy of the voiceprint.

[0062] Furthermore, during the iterative training of the generator in the StarGAN-VC network architecture to obtain the generator network, since the number of training iterations for the second generator network is less than that for the first generator network, the difference between the original registered voice signal and the first registered voice signal is lower than the difference between the original certified voice signal and the second certified voice signal. This ensures the privacy and security of the original registered audio signal while allowing the first registered audio signal to retain more of the voice features of the original registered audio signal, further improving the accuracy of the subsequent voiceprint authentication process. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 A flowchart illustrating a voiceprint authentication method based on StarGAN-VC provided in an embodiment of the present invention;

[0065] Figure 2 A schematic diagram of a process for extracting Mel frequency cepstral coefficients provided in an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram of the structure for iterative training of a generator network provided in an embodiment of the present invention;

[0067] Figure 4 A flowchart of another voiceprint authentication method based on StarGAN-VC provided for an embodiment of the present invention;

[0068] Figure 5 The image shows the test results of the similarity between the synthesized audio and the original audio provided in the embodiments of the present invention.

[0069] Figure 6A structural diagram of a voiceprint authentication system based on StarGAN-VC provided in an embodiment of the present invention;

[0070] Figure 7 This is a diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] This application provides a voiceprint authentication method and system based on StarGAN-VC, addressing the issue of low privacy and security of users' original voiceprint feature information in existing voiceprint authentication methods. It achieves this by using the original registration audio as training samples to iteratively train the generator network in the StarGAN-VC network architecture. Based on different iteration counts, a first generator network and a second generator network are obtained. The second generator network then transforms the original registration speech signal into a first registration audio signal, ensuring a difference between the first and original registration speech signals and protecting the voiceprint privacy of the original registration speech signal. Furthermore, the first generator network transforms the original authentication speech signal into a second authentication speech signal, effectively protecting the voiceprint privacy of the original authentication speech signal as well.

[0073] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0074] This invention uses the original registration audio as training samples to iteratively train the generator network in the StarGAN-VC network architecture. Based on different numbers of iterations, a first generator network and a second generator network are obtained. The second generator network then transforms the original registration speech signal into a first registration audio signal, ensuring a difference between the first and original registration audio signals and protecting the voiceprint privacy of the original registration speech signal. Furthermore, the first generator network transforms the original authentication speech signal into a second authentication speech signal, effectively protecting the voiceprint privacy of the original authentication speech signal as well. Because the specific transformation methods used by the generator in the StarGAN-VC network architecture are unknown, the confidentiality of the speech signal transformation is greatly enhanced. Therefore, even with the first registration speech signal, existing technologies cannot reverse-engineer it back into the original registration speech signal, providing reliable protection for the voiceprint privacy of the user's original voiceprint features.

[0075] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0076] First, the basic knowledge involved in the technical solution of this invention will be introduced.

[0077] Voiceprints can be represented as the speech features inherent in speech that characterize and identify the speaker, or a speech model built based on these speech features (parameters). Theoretically, like fingerprints, very few two people have the same voiceprint features; therefore, voiceprints can be used to verify a user's identity. Correspondingly, voiceprint recognition is the process of identifying the speaker corresponding to a given segment of speech based on its voiceprint features.

[0078] Depending on the application scenario, voiceprint recognition can be divided into the following categories: voiceprint verification, voiceprint identification, voiceprint detection, and voiceprint tracking. Voiceprint verification can be understood as determining whether a given speaker's voiceprint model and a segment of speech containing only that speaker's voiceprint belong to that speaker. It is worth noting that the voiceprint authentication method of this invention is applied to the voiceprint verification scenario.

[0079] Because existing voiceprint authentication methods either do not encrypt the user's original voice or use relatively simple encryption techniques, the privacy and security of the user's original voiceprint feature information are low. Therefore, this invention provides a voiceprint authentication method based on StarGAN-VC to solve the above problems. See also Figure 1 , Figure 1A flowchart of a voiceprint authentication method based on StarGAN-VC provided for embodiments of the present invention is shown below. Figure 1 As shown, the method includes the following steps:

[0080] S1. Obtain the first authentication voice signal.

[0081] The first authentication voice signal is obtained by preprocessing the original authentication voice signal.

[0082] S2. Based on the pre-acquired first generator network, process the first authentication voice signal to obtain the second authentication voice signal.

[0083] The process of obtaining the first generator network includes: using the original registered speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample. When the number of iterations reaches a first preset value, the corresponding generator network is obtained as the first generator network.

[0084] S3. Calculate the similarity between the second authentication voice signal and the first registration voice signal. If the similarity is greater than the first preset threshold, the original authentication voice signal is successfully authenticated.

[0085] The first registered speech signal is obtained based on the original registered speech signal and the pre-acquired second generator network. The process of acquiring the second generator network includes: using the original registered speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample. When the number of iterations reaches a second preset value, the corresponding generator network is acquired as the second generator network. The second preset value is less than the first preset value.

[0086] Based on the above processing, this invention uses the original registration audio as training samples to iteratively train the generator network in the StarGAN-VC network architecture. Based on different iterations, a first generator network and a second generator network are obtained. The second generator network then transforms the original registration speech signal into a first registration audio signal, ensuring a difference between the first and original registration speech signals and protecting the voiceprint privacy of the original registration speech signal. Furthermore, the first generator network transforms the original authentication speech signal into a second authentication speech signal, effectively protecting the voiceprint privacy of the original authentication speech signal as well. Because the specific transformation methods used by the generator in the StarGAN-VC network architecture to transform the speech signal are unknown in the prior art, the confidentiality of the speech signal transformation is greatly improved. Therefore, even if the first registration speech signal is obtained, the prior art cannot reverse-engineer it back into the original registration speech signal, providing reliable protection for the voiceprint privacy of the user's original voiceprint feature information.

[0087] Furthermore, since the first generator network and the second generator network are trained based on the same training samples and the same generator in the StarGAN-VC network architecture, with only the number of training iterations differing, the conversion process of the original registered speech signal is basically the same as that of the original certified speech signal. Therefore, by calculating the similarity between the second certified speech signal and the first registered speech signal, the voiceprint authentication of the original certified speech signal can be completed, thereby further ensuring the accuracy of voiceprint authentication while protecting the privacy of the voiceprint.

[0088] Furthermore, during the iterative training of the generator in the StarGAN-VC network architecture to obtain the generator network, since the number of training iterations for the second generator network is less than that for the first generator network, the difference between the original registered voice signal and the first registered voice signal is lower than the difference between the original certified voice signal and the second certified voice signal. This ensures the privacy and security of the original registered audio signal while allowing the first registered audio signal to retain more of the voice features of the original registered audio signal, further improving the accuracy of the subsequent voiceprint authentication process.

[0089] The voiceprint authentication method of this invention can be applied to voiceprint verification scenarios such as access control and user account login, and is executed by electronic devices such as computers, mobile phones, and voice processing devices. In the following description of this invention, "computer" refers to the aforementioned electronic device. The original registration voice signal refers to the audio signal input by the user to the computer during the registration process before voiceprint authentication; the original authentication voice signal refers to the audio signal to be authenticated input by the user to the computer during the voiceprint authentication process.

[0090] For step S1, the first authentication voice signal is represented by a voice signal obtained by preprocessing the original authentication voice signal.

[0091] Since the original voice signal entered by the user during registration or authentication (i.e., the original registration voice signal and the original authentication voice signal in this invention) is a time-series signal, which is a one-dimensional array of data, while the generator in the StarGAN-VC model framework can process two-dimensional data (i.e., matrix data, or it can also be understood as a spectrogram), the original voice signal needs to be preprocessed before being input into the generator. This preprocessing method involves converting the original voice signal from a one-dimensional time-series array into a two-dimensional matrix of data.

[0092] Since MFCC (Mel-scale Frequency Cepstral Coefficients) contains the speech features of the speech signal and is stored in matrix form, the preferred preprocessing method of this invention is to extract the Mel-scale frequency cepstral coefficients of the original speech signal.

[0093] In some embodiments, prior to step S1, the method may include the following steps:

[0094] S4. Obtain the original authentication voice signal.

[0095] S5. Obtain the Mel frequency cepstral coefficients of the original authentication voice signal as the first authentication voice signal.

[0096] In one implementation method, the process of obtaining the Mel-frequency cepstral coefficients of the original authenticated speech signal can be divided into the following steps:

[0097] Step 1: Pre-emphasize the original authentication voice signal, that is, increase the high-frequency part of the voice to flatten the original authentication voice signal and filter out noise.

[0098] Step 2: Perform frame segmentation on the pre-emphasis processed original authentication speech signal to discretize the continuous speech signal. Specifically, sample the pre-emphasis processed original authentication speech signal 22.5k times per second, and use the number of sampling points in each sample as a frame, adding a Hamming window.

[0099] Step 3: Perform FFT (Fast Fourier Transform) on each frame of the original authenticated speech signal after frame segmentation to convert the time-domain signal into a frequency-domain signal. The frequency-domain signal after FFT transformation is then concatenated over time to obtain the spectrogram of the speech signal. Since the obtained spectrogram is quite large, it is further processed to obtain appropriately sized sound features.

[0100] Step 4: Input the original certified speech signal after FFT transformation into the Mel-scale filter bank to obtain the Mel spectrum;

[0101] Step 5: Perform cepstral analysis on the Mel spectrum to obtain the Mel frequency cepstral coefficients. The cepstral analysis includes taking the logarithm and performing a Discrete Cosine Transform (DCT).

[0102] See Figure 2 , Figure 2 This is a schematic diagram of a process for extracting Mel frequency cepstral coefficients according to an embodiment of the present invention.

[0103] It is understandable that the process of obtaining the Mel frequency cepstral coefficients of the original registered speech signal is the same as the process of obtaining the Mel frequency cepstral coefficients of the original certified speech signal described above.

[0104] In some embodiments, step S2 may include the following steps:

[0105] S201. The first authentication voice signal and the target voice tag are input together into the pre-acquired first generator network to obtain the signal output by the first generator network as the third authentication voice signal; wherein, the target voice tag is the voice tag of the pre-selected target voice signal.

[0106] S202. The third authentication voice signal and the original authentication voice tag are input together into the first generator network to obtain the signal output by the first generator network, which serves as the fourth authentication voice signal. The original authentication voice tag is the voice tag of the original authentication voice signal.

[0107] S203. Based on the third and fourth certified voice signals, obtain the first spectral gain function.

[0108] S204. Multiply the spectral envelope of the original authentication speech with the first spectral gain function, and process the multiplication result based on the vocoder to obtain the second authentication speech signal.

[0109] It can be understood that the first authentication voice signal is two-dimensional data obtained by preprocessing the original authentication voice signal. The second authentication voice signal is a time-domain signal processed by the first generator network's first spectral gain function, vocoder, etc.

[0110] Based on the above processing, the first authentication voice signal is converted into the second authentication voice signal, effectively protecting the voiceprint privacy of the original authentication voice signal.

[0111] In some embodiments, the process of obtaining the first generator network includes: using the original registered speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample; when the number of iterations reaches a first preset value, obtaining the corresponding generator network as the first generator network.

[0112] In one implementation, the process of iteratively training the generator network in the StarGAN-VC network architecture using the original registered speech signal as training samples includes the following steps:

[0113] S301. Obtain the Mel frequency cepstral coefficients of the original registered speech signal as the speech signal to be processed.

[0114] S302. Input the speech signal to be processed and the target audio label together into the generator network in the StarGAN-VC network architecture to be trained to obtain the first synthesized speech signal.

[0115] S303. Input the first synthesized speech signal into the pre-trained discriminator network to obtain the probability that the first synthesized speech signal is the target speech signal; and input the first synthesized speech signal and the original audio tag into the generator network in the StarGAN-VC network architecture to be trained to obtain the reconstructed speech signal.

[0116] S304. Calculate the loss value of the generator network based on the probability that the first synthesized speech signal is the target speech signal.

[0117] S305. Adjust the network parameters of the generator network based on the loss value of the generator network; wherein, after adjusting the network parameters of the generator network, it is recorded as completing one iteration of training.

[0118] S306. The reconstructed speech signal replaces the speech signal to be processed, and the replaced speech signal to be processed and the target audio tag are re-inputted into the generator network after adjusting the network parameters to obtain the first synthesized speech signal, and iterative training continues.

[0119] In this context, the audio tag represents the speech characteristics of the speech signal. The pre-selected target speech signal is a speech signal pre-selected by the user, and the audio tag of this speech signal serves as the target audio tag.

[0120] Based on the above, it can be seen that when the number of iterations trained on the generator network in the StarGAN-VC network architecture meets a first preset value, the generator network can be used as the first generator network; when the number of iterations trained meets a second preset value, the generator network can be used as the first generator network. The first preset value ranges from 75 to 85, preferably 80. The second preset value ranges from 15 to 25, preferably 20.

[0121] Based on existing technology, the purpose of converting the original speech signal through the generator network in the StarGAN-VC network architecture is to make the converted original speech signal as close as possible to the pre-defined target speech signal.

[0122] In the technical solution of this invention, when the generator network in the StarGAN-VC network architecture is used to convert the original signal (i.e., the original authentication voice signal or the original registration voice signal of this invention), the purpose is to make the converted voice signal retain the voiceprint features of the original signal as much as possible, while also having a certain degree of difference from the original signal.

[0123] Understandably, in order to ensure that the first registered audio retains more of the voiceprint features from the original registered speech signal, thereby improving the accuracy of voiceprint recognition, the range of the second preset value is set to 15-25, preferably 20 times.

[0124] In the technical solution of this invention, once a user completes account registration, that user is considered a legitimate user. All other users are considered illegitimate users. The voiceprint authentication method of this invention is used to determine whether the original authentication voice signal belongs to a legitimate user, that is, whether the user to be authenticated is a legitimate user.

[0125] If the first preset value is very large (e.g., exceeding 100 times), the second authentication voice signal, whether from an illegitimate or legitimate user, will contain many voiceprint features from the pre-target voice signal, resulting in a very high similarity to the first registered voice signal. Therefore, it will be impossible to determine whether the original registered voice signal has passed voiceprint authentication based on similarity, and voiceprint authentication will be impossible. Thus, the first preset value should not be set too large.

[0126] If the first preset value is small (e.g., between 20 and 50), the second authentication voice signal converted by the first generator network will contain fewer voiceprint features of the pre-selected target voice signal. Therefore, even if the second authentication voice signal belongs to a legitimate user, its similarity to the first registered voice signal will be low. To address this, to ensure successful voiceprint authentication, the value of the first preset threshold needs to be reduced. However, reducing the value of the first preset threshold will directly decrease the accuracy of the voiceprint authentication in this invention. Therefore, the first preset value should not be set too small. It is worth noting that the similarity calculation between the second authentication voice signal and the first registered voice signal can be simply understood as superimposing the voiceprint features of the legitimate user and the pre-selected target voice signal in the second authentication voice signal and the first registered voice signal.

[0127] Furthermore, the values ​​of the first and second preset values ​​can be adjusted according to the number of speech features contained in the original registered speech signal. If the number of speech features is very large, the first and second preset values ​​can be appropriately increased.

[0128] To verify the above, the inventors conducted simulation verification with different settings for the first and second preset values. SF1 represents a legitimate user, while SF2, SF3, SF4, and TM2 represent illegitimate users. The test text consisted of five different registered voice recordings input by legitimate users. The content of the middle table in Tables 1 and 2 represents the similarity scores. Table 1 shows the test results of this technical solution, with the first preset value being 80 and the second preset value being 20. The test results are as follows.

[0129] Table 1 Test results of this technical solution

[0130]

[0131] Table 2 shows the test results of the comparative technical solutions. The first preset value is 50 and the second preset value is 20. The test results are as follows.

[0132] Table 2 Test results of the comparative technical solutions

[0133]

[0134]

[0135] The first preset threshold is set to 0.7, and the voiceprint authentication results are judged based on the similarity in Tables 1 and 2. There are only two voiceprint authentication errors in Table 1, while there are nine authentication errors in Table 2. Therefore, the preferred first preset value of this invention is 80, and the preferred second preset value is 20, which has better results.

[0136] See Figure 3 , Figure 3 This is a schematic diagram illustrating the structure of an iterative training generator network provided in an embodiment of the present invention. Figure 3 As shown, the speech signal to be processed and the target audio tag are input together into the generator network G (i.e., the generator network to be trained in this invention) to obtain a first synthesized speech signal. Then, the first synthesized speech signal is input into a pre-trained discriminator D to obtain the probability that the first synthesized speech signal is the target speech signal. At the same time, the first synthesized speech signal and the original audio tag are input together into the generator network G (i.e., the generator network to be trained in this invention) to obtain a reconstructed speech signal.

[0137] Based on the probability that the first synthesized speech signal is the pre-selected target speech signal, the loss value of the generator network is obtained, and the network parameters of the generator network are adjusted based on the loss value. Then, the reconstructed speech signal is used as the speech signal to be processed, and the target audio label is re-inputted into the generator network after the network parameters are adjusted, and iterative training continues.

[0138] In some embodiments, the computer processes the original registration audio signal during the user registration process as follows:

[0139] a. Obtain the Mel frequency cepstral coefficients of the original registered audio signal as the speech signal to be processed.

[0140] b. Iteratively train the generator network in the StarGAN-VC network architecture based on the speech signal to be processed to obtain the first generator network and the second generator network.

[0141] c. Input the registered speech signal to be processed and the target speech label into the pre-trained second generator network to obtain the signal output by the second generator network as the third registered speech signal.

[0142] d. Input the third registered speech signal and the original registered speech tag into the second generator network to obtain the signal output by the second generator network, which serves as the fourth registered speech signal. The original registered speech tag is the speech tag of the original registered speech signal.

[0143] e. Based on the third registered speech signal and the fourth registered speech signal, obtain the second spectral gain function.

[0144] f. Multiply the spectral envelope of the original registered speech with the second spectral gain function, and process the multiplication result based on the vocoder to obtain the first registered speech signal.

[0145] g. Encrypt the first registered voice signal to obtain the second registered voice signal.

[0146] In some embodiments, step g, which encrypts the first registered voice signal to obtain a second registered voice signal, may include the following steps:

[0147] G01. Delete the original registered voice signal.

[0148] G02. Encrypt the first registered voice signal using the AES encryption library to obtain the encrypted first registered voice signal, which is then used as the second registered voice signal.

[0149] In one implementation, after step G02, the method further includes the following steps:

[0150] G03. Store the second registered voice signal in the database and delete the user password. This user password refers to the account password set by the user during account registration. Furthermore, this account password is the same as the encryption key.

[0151] In some embodiments, after a user's account registration is completed, the user's password is hashed to obtain the corresponding hash value, which is then used as the user's hash value.

[0152] Once the user's hash value is obtained, the plaintext password in the user's account will be deleted.

[0153] In one implementation, during voiceprint authentication, the user to be authenticated first needs to input a password into the voiceprint recognition system. This password, obtained by the voiceprint recognition system, serves as the authentication password. Then, the authentication password is hashed to obtain the corresponding hash value, which is used as the authentication hash value.

[0154] The system checks if the authentication hash value is the same as the user hash value. If they are the same, the authentication password is the same as the user's password during registration, indicating that the authentication password is successful. If they are different, it means that the password entered by the user to be authenticated is different from the user's password, and the voiceprint authentication operation is stopped.

[0155] Since the account password and encryption key are the same, once the user's authentication password is entered and passed, the second registered voice signal is decrypted based on the authentication password to obtain the first registered voice signal.

[0156] Since the original password cannot be derived from the hash value, once the plaintext password of an account is deleted, even if the user's hash value on the device containing the account is leaked, users other than the account holder will not be able to obtain the account's password. Furthermore, because the password key is the same as the account password, this process effectively prevents attackers from using secure audio for voiceprint authentication after the database is stolen, thus preventing the leakage of user information.

[0157] For step S3, the similarity between the second authentication voice signal and the first registration voice signal is calculated. If the similarity is greater than a first preset threshold, the original authentication voice signal is successfully authenticated. The first preset threshold is set by the user and ranges from 0 to 1.

[0158] In one implementation, the second authentication voice signal and the first registration voice signal are input to a preset discriminator. The discriminator calculates the similarity between the second authentication voice signal and the first registration voice signal and determines whether the similarity is greater than 0.7 (i.e., the first preset threshold in this invention). If yes, the voiceprint authentication of the original authentication voice signal passes. If no, the voiceprint authentication of the original authentication voice signal fails.

[0159] See Figure 4 , Figure 4 A flowchart illustrating another voiceprint authentication method based on StarGAN-VC provided in this embodiment of the invention. Figure 4 As shown, the method includes:

[0160] During user registration, the computer first obtains the user's account password, hashes the password to convert it into a user hash value, and deletes the plaintext password (e.g., the account password) from the account. The computer then processes the obtained user registration voice (i.e., the original registration voice signal of this invention) through the second generator network in the preprocessing module and the voiceprint privacy module to obtain a first registration voice signal. The first registration voice signal is then encrypted using AES and stored in the user database.

[0161] When performing voiceprint authentication, the computer first obtains the user's password and hashes it to obtain an authentication hash value. The system then compares the authentication hash value with the user's hash value to determine whether to continue the voiceprint authentication process.

[0162] When performing voiceprint authentication, the computer obtains the original authentication voice signal of the user to be authenticated. This original authentication voice signal is processed by a preprocessing module and a first generation network to obtain the protected authentication voice signal (i.e., the second authentication voice signal of this invention). At the same time, the computer obtains the encrypted first registration voice signal from the user database and decrypts it using AES based on the authentication password to obtain the first registration voice signal.

[0163] Then, the protected authentication voice signal and the first registered voice signal are input into a preset discriminator to obtain the comparison result, that is, to calculate the similarity between the protected authentication voice signal and the first registered voice signal, and then to determine whether the voiceprint authentication is successful.

[0164] The inventors conducted performance tests on the voiceprint authentication method based on StarGAN-VC provided by this invention.

[0165] During performance testing, the voices of four source speakers and four target speakers (comprising female and male speakers) from the VCC2016 fixed corpus were selected as training data. Each speaker's voice dataset contains 150 different voice recordings, with a one-to-one correspondence between the voice datasets of different speakers. The data in this dataset was divided into two subsets: the VCC2016 training set and the VCC2016 evaluation set. The VCC2016 evaluation set was used as the test set.

[0166] In this test, the speech data of four speakers SF1, SF2, TM1, and TM2 from the VCC2016 training set were extracted and placed in the training set, while the data from their corresponding evaluation sets were placed in the test set. The text information entered by the speakers is divided into Text 1 and Text 2.

[0167] After a legitimate user TM1 registers and enters text 1 (or text 2), its original speech (i.e., the original registration speech signal of this invention) is preprocessed and then processed by the second generator network to obtain the synthesized audio SF1-TM1+1 (or SF1-TM1+2); after a legitimate user TM2 registers and enters text 1 (or text 2), its original speech is preprocessed and then processed by the second generator network to obtain the synthesized audio SF2-TM2+1 (or SF2-TM2+2).

[0168] First, the privacy protection performance of the second generator network was tested. It is understandable that if the synthesized audio differs significantly from the original audio in terms of its corresponding voiceprint spectrum, even if the encrypted synthesized audio stored in the database is stolen and deciphered, attackers will not be able to obtain the user's original voiceprint information, thus achieving the goal of protecting the user's original voiceprint characteristics.

[0169] Therefore, the synthesized speech can be compared with the original audio for similarity testing. The lower the similarity, the higher the privacy protection of the technical solution of this invention. For the similarity comparison method, this invention selects three comparison methods: SSIM (Structure Similarity Index Measure), MSE (Mean Square Error), and histogram similarity calculation.

[0170] Histogram similarity calculation, based on simple vector similarity, measures image similarity and provides good normalization. SSIM is a metric for measuring the similarity between two images, primarily considering three key features: brightness, contrast, and structure. SSIM returns a structural similarity index as a floating-point number between 0 and 1 (the closer to 1, the higher the similarity). MSE is a metric used to calculate the similarity between two images; a smaller MSE indicates greater similarity.

[0171] See Figure 5 , Figure 5 This is a graph showing the test results of the similarity between the synthesized audio and the original audio provided in an embodiment of the present invention. Figure 5 As shown in the test results, the SSIM index is generally below 65%, the histogram similarity is also below 80%, and the MSE error is generally high. Therefore, it can be determined that there are significant differences between the synthesized audio and the original audio in auditory analysis and spectrogram analysis. The test meets the requirements, thus proving that the technical solution of this invention has good privacy protection performance for voice technology and meets the actual privacy protection needs.

[0172] The above tests quantitatively demonstrated a significant difference between the synthesized audio and the original audio, thus achieving the goal of protecting user voiceprint privacy. Accordingly, the voiceprint authentication performance of the technical solution of this invention was tested. In this test, the first preset threshold was set to 0.7. The test results are shown in Tables 3 and 4.

[0173] Table 3. Comparison results of synthesized audio between different users and SF1

[0174]

[0175]

[0176] Table 4. Comparison results of synthesized audio between different users and SF2

[0177]

[0178] As can be seen from the test results in Tables 1 and 2, the voiceprint authentication method based on StarGAN-VC provided by this invention can ensure the accuracy of voiceprint authentication while protecting the voiceprint privacy and security of the original voice signal.

[0179] See Figure 6 , Figure 6 This is a structural diagram of a voiceprint authentication system based on StarGAN-VC provided by the present invention. Figure 6 As shown, the system includes:

[0180] The signal acquisition module 601 is used to acquire the first authentication voice signal. The first authentication voice signal is obtained by preprocessing the original authentication voice signal.

[0181] The signal processing module 602 is used to process the first authentication speech signal based on a pre-acquired first generator network to obtain a second authentication speech signal. The acquisition process of the first generator network is as follows: the original registration speech signal is used as a training sample, and the generator network in the StarGAN-VC network architecture is iteratively trained based on the training sample. When the number of iterations reaches a first preset value, the corresponding generator network is acquired and used as the first generator network.

[0182] The similarity calculation module 603 is used to calculate the similarity between the second authenticated speech signal and the first registered speech signal. If the similarity is greater than a first preset threshold, it indicates that the original authenticated speech signal has been successfully authenticated. The first registered speech signal is obtained based on the original registered speech signal and a pre-acquired second generator network. The generation process of the second generator network is as follows: the original registered speech signal is used as a training sample, and the generator network in the StarGAN-VC network architecture is iteratively trained based on the training sample. When the number of iterations reaches a second preset value, the corresponding generator network is obtained and used as the second generator network. The second preset value is less than the first preset value.

[0183] It is understood that the voiceprint authentication system based on StarGAN-VC provided in this embodiment of the invention corresponds to the voiceprint authentication method based on StarGAN-VC described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the voiceprint authentication method based on StarGAN-VC, and will not be repeated here.

[0184] This invention also provides an electronic device, such as... Figure 7 As shown, it includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704.

[0185] Memory 703 is used to store computer programs.

[0186] The processor 701, when executing the program stored in the memory 703, implements any of the aforementioned voiceprint authentication methods based on StarGAN-VC.

[0187] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0188] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0189] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0190] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0191] In summary, compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0192] This invention uses the original registration audio as training samples to iteratively train the generator network in the StarGAN-VC network architecture. Based on different numbers of iterations, a first generator network and a second generator network are obtained. The second generator network then transforms the original registration speech signal into a first registration audio signal, ensuring a difference between the first and original registration audio signals and protecting the voiceprint privacy of the original registration speech signal. Furthermore, the first generator network transforms the original authentication speech signal into a second authentication speech signal, effectively protecting the voiceprint privacy of the original authentication speech signal as well. Because the specific transformation methods used by the generator in the StarGAN-VC network architecture are unknown, the confidentiality of the speech signal transformation is greatly enhanced. Therefore, even with the first registration speech signal, existing technologies cannot reverse-engineer it back into the original registration speech signal, providing reliable protection for the voiceprint privacy of the user's original voiceprint features.

[0193] Furthermore, since the first generator network and the second generator network are trained based on the same training samples and the same generator in the StarGAN-VC network architecture, with only the number of training iterations differing, the conversion process of the original registered speech signal is basically the same as that of the original certified speech signal. Therefore, by calculating the similarity between the second certified speech signal and the first registered speech signal, the voiceprint authentication of the original certified speech signal can be completed, thereby further ensuring the accuracy of voiceprint authentication while protecting the privacy of the voiceprint.

[0194] Furthermore, during the iterative training of the generator in the StarGAN-VC network architecture to obtain the generator network, since the number of training iterations for the second generator network is less than that for the first generator network, the difference between the original registered voice signal and the first registered voice signal is lower than the difference between the original certified voice signal and the second certified voice signal. This ensures the privacy and security of the original registered audio signal while allowing the first registered audio signal to retain more of the voice features of the original registered audio signal, further improving the accuracy of the subsequent voiceprint authentication process.

[0195] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0196] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A voiceprint authentication method based on StarGAN-VC, characterized in that, The method includes: Obtain a first authentication voice signal; wherein the first authentication voice signal is obtained based on preprocessing of the original authentication voice signal; Based on the pre-acquired first generator network, the first authentication speech signal is processed to obtain the second authentication speech signal; wherein, the acquisition process of the first generator network includes: using the original registration speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample; when the number of iterations reaches a first preset value, the corresponding generator network is acquired as the first generator network. The similarity between the second certified speech signal and the first registered speech signal is calculated. If the similarity is greater than a first preset threshold, the original certified speech signal is successfully certified. The first registered speech signal is obtained based on the original registered speech signal and a pre-acquired second generator network. The process of acquiring the second generator network includes: using the original registered speech signal as a training sample, and iteratively training the generator network in the StarGAN-VC network architecture based on the training sample. When the number of iterations reaches a second preset value, the corresponding generator network is acquired as the second generator network. The second preset value is less than the first preset value. The first authentication voice signal is processed based on a pre-acquired first generator network to obtain a second authentication voice signal, including: The first authentication voice signal and the target voice tag are input together into a pre-acquired first generator network to obtain the signal output by the first generator network, which serves as the third authentication voice signal; wherein, the target voice tag is the voice tag of the pre-selected target voice signal; The third authentication voice signal and the original authentication voice tag are input together into the first generator network to obtain the signal output by the first generator network, which serves as the fourth authentication voice signal; wherein, the original authentication voice tag is the voice tag of the original authentication voice signal. Based on the third and fourth authentication voice signals, a first spectral gain function is obtained; The spectral envelope of the original authenticated voice signal is multiplied by the first spectral gain function, and the multiplication result is processed based on the vocoder to obtain the second authenticated voice signal.

2. The method according to claim 1, characterized in that, Before acquiring the first authentication voice signal, the method further includes: Obtain the original authentication voice signal; The Mel-frequency cepstral coefficients of the original authentication voice signal are obtained and used as the first authentication voice signal.

3. The method according to claim 1, characterized in that, The process of iteratively training the generator network in the StarGAN-VC network architecture using the original registered speech signals as training samples includes: The Mel frequency cepstral coefficients of the original registered speech signal are obtained as the speech signal to be processed; The speech signal to be processed and the target audio tag are input together into the generator network in the StarGAN-VC network architecture to be trained to obtain the first synthesized speech signal; The first synthesized speech signal is input into a pre-trained discriminator network to obtain the probability that the first synthesized speech signal is the target speech signal; and the first synthesized speech signal and the original audio tag are input together into the generator network in the StarGAN-VC network architecture to be trained to obtain the reconstructed speech signal. Based on the probability that the first synthesized speech signal is the target speech signal, the loss value of the generator network is calculated. The network parameters of the generator network are adjusted based on the loss value of the generator network; wherein, after the network parameters of the generator network are adjusted, it is recorded as the completion of one iteration of training; The reconstructed speech signal replaces the speech signal to be processed, and the replaced speech signal to be processed and the target audio tag are re-inputted into the generator network after adjusting the network parameters to obtain the first synthesized speech signal, and iterative training continues.

4. The method according to claim 1, characterized in that, Before calculating the similarity between the second authenticated voice signal and the first registered voice signal, the method further includes: Obtain a second registered voice signal; wherein the second registered voice signal is obtained by encrypting the first registered voice signal; The second registered voice signal is decrypted to obtain the first registered voice signal.

5. The method according to claim 4, characterized in that, Before acquiring the second registered voice signal, the method further includes: Obtain the original registered voice signal; The Mel-frequency cepstral coefficients of the original registered speech signal are obtained as the registered speech signal to be processed; The registered speech signal to be processed and the target speech label are input into the pre-trained second generator network to obtain the signal output by the second generator network, which is used as the third registered speech signal; The third registered speech signal and the original registered speech tag are input into the second generator network to obtain the signal output by the second generator network, which is used as the fourth registered speech signal; wherein, the original registered speech tag is the speech tag of the original registered speech signal; Based on the third registered speech signal and the fourth registered speech signal, obtain the second spectral gain function; The spectral envelope of the original registered speech signal is multiplied by the second spectral gain function, and the multiplication result is processed based on the vocoder to obtain the first registered speech signal. The first registered voice signal is encrypted to obtain the second registered voice signal.

6. The method according to claim 5, characterized in that, The first registered voice signal is encrypted to obtain the second registered voice signal, which includes: The original registered voice signal was deleted; The first registered voice signal is encrypted using the AES encryption library to obtain the encrypted first registered voice signal, which is then used as the second registered voice signal.

7. The method according to claim 6, characterized in that, After encrypting the first registered voice signal using an AES encryption library to obtain the encrypted first registered voice signal, which is then used as the second registered voice signal, the method further includes: The second registered voice signal is stored in the database, and the user password is deleted; wherein the user password is the same as the key used for encryption based on the AES encryption library.

8. The method according to claim 7, characterized in that, Before deleting the user's password, the method further includes: Hash the user password to obtain the corresponding hash value, which is used as the user hash value; Before acquiring the original authentication voice signal, the method further includes: Get the authentication password; The authentication password is hashed to obtain the corresponding hash value, which is used as the authentication hash value; Determine whether the authentication hash value is the same as the user hash value; if so, it indicates that the authentication password is successful. The step of decrypting the second registered voice signal to obtain the first registered voice signal includes: Obtain the authentication password; The second registered voice signal is decrypted based on the authentication password to obtain the first registered voice signal.

9. A voiceprint authentication system based on StarGAN-VC, characterized in that, The system includes: The signal acquisition module is used to acquire a first authentication voice signal; wherein the first authentication voice signal is obtained based on preprocessing of the original authentication voice signal; The signal processing module is used to process the first authentication speech signal based on the pre-acquired first generator network to obtain the second authentication speech signal; wherein, the acquisition process of the first generator network is as follows: the original registration speech signal is used as a training sample, and the generator network in the StarGAN-VC network architecture is iteratively trained based on the training sample. When the number of iterations reaches a first preset value, the corresponding generator network is acquired as the first generator network. A similarity calculation module is used to calculate the similarity between the second authenticated speech signal and the first registered speech signal. If the similarity is greater than a first preset threshold, the original authenticated speech signal is successfully authenticated. The first registered speech signal is obtained based on the original registered speech signal and a pre-acquired second generator network. The generation process of the second generator network is as follows: the original registered speech signal is used as a training sample, and the generator network in the StarGAN-VC network architecture is iteratively trained based on the training sample. When the number of iterations reaches a second preset value, the corresponding generator network is obtained as the second generator network. The second preset value is less than the first preset value. The first authentication voice signal is processed based on a pre-acquired first generator network to obtain a second authentication voice signal, including: The first authentication voice signal and the target voice tag are input together into a pre-acquired first generator network to obtain the signal output by the first generator network, which serves as the third authentication voice signal; wherein, the target voice tag is the voice tag of the pre-selected target voice signal; The third authentication voice signal and the original authentication voice tag are input together into the first generator network to obtain the signal output by the first generator network, which serves as the fourth authentication voice signal; wherein, the original authentication voice tag is the voice tag of the original authentication voice signal. Based on the third and fourth authentication voice signals, a first spectral gain function is obtained; The spectral envelope of the original authenticated voice signal is multiplied by the first spectral gain function, and the multiplication result is processed based on the vocoder to obtain the second authenticated voice signal.

Images