A method and system for preventing speech synthesis attacks based on self-supervision
The self-supervised training method enhances voiceprint recognition systems' ability to detect speech synthesis attacks by extracting hidden features and context features, improving detection accuracy through unlabeled data utilization and model fusion techniques.
Patent Information
- Application Number
- CN202210675092.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-15
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2042-06-15
AI Technical Summary
Existing voiceprint recognition systems lack effective detection mechanisms for speech synthesis attacks, particularly due to limitations in audio feature extraction, environmental noise interference, and the challenge of distinguishing between real and fake voices with the advancement of voice cloning technology.
A self-supervised training method that extracts hidden features and context features from unlabeled data, fuses pre-trained models with synthetic detection models using a gated recurrent unit and attention module, and updates models based on contribution values to enhance detection accuracy.
Improves the generalization ability of synthetic speech detection by effectively utilizing unlabeled data and addressing gradient issues, resulting in enhanced model feature extraction and detection accuracy.
Smart Images
Figure CN115171725B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of voice attack prevention, and in particular to a method and system for preventing voice synthesis attacks based on self-supervision. Background Art
[0002] With the increasing popularity of voiceprint recognition technology, anti-attack has become more and more important. The main research purpose of voiceprint recognition anti-speech synthesis attack is to detect whether a recording is real or artificially synthesized audio, so as to prevent others from stealing voiceprint information. Anti-attack is essential for any voiceprint recognition system and is an important part of its system. With the advancement of technology in recent years (such as voice cloning technology), speech synthesis attacks have become easier to obtain, and the types of attacks have become more diverse, which has become one of the main means of attack. The voiceprint recognition system needs to distinguish the user identity based on the input audio and complete the user authorization. The existing voiceprint recognition system lacks an effective detection mechanism for speech synthesis attacks. The main detection method is to use artificially designed features (such as MFCC) and classify according to the extracted features.
[0003] However, the current methods for preventing speech synthesis attacks have certain limitations:
[0004] First, audio feature extraction is mainly manually designed, such as Mel-frequency cepstral coefficients (MFCC). These features are mostly designed for speech recognition systems and conform to the characteristics of the human auditory system. However, sometimes it is difficult for the human ear to distinguish whether it is synthesized speech, so it may not be suitable for speech synthesis detection.
[0005] Secondly, environmental noise, recording equipment and attack methods have a greater impact on the detection results of anti-speech synthesis attacks, and the coverage of scenarios and attack types is relatively small. These factors make it difficult to collect training data, and without sufficient data support, it is difficult for deep learning models to achieve the preset results.
[0006] Finally, with the advancement of technology, voice cloning technology has been able to easily imitate a person's voice characteristics, making it difficult to distinguish between the real and the fake, and the technology has become diversified and difficult to predict effectively. Summary of the invention
[0007] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art and propose a method and system for preventing speech synthesis attacks based on self-supervision, which can effectively utilize a large amount of unlabeled data and improve the generalization ability of synthesis detection.
[0008] The present invention adopts the following technical solution:
[0009] A method for preventing speech synthesis attacks based on self-supervision, comprising:
[0010] Obtain speech data training samples;
[0011] Input the speech data training sample into the pre-training model, extract the hidden features, and further obtain the context features based on the extracted hidden features. Then, reversely predict the hidden features through the context features to complete the self-supervised training of the pre-training model and obtain the pre-training model after initial training.
[0012] Use a variety of synthetic speech algorithms to construct a data set to prevent speech synthesis attacks;
[0013] The pre-trained model after initial training is fused with the synthesis detection model through the gated recurrent unit and the attention module to obtain a fusion model. The anti-speech synthesis attack dataset is input into the fusion model to complete the training of the fusion model and obtain the contribution value of each speech feature.
[0014] The speech features corresponding to the anti-speech synthesis attack dataset and the set contribution value are input into the pre-trained model after initial training for re-training to obtain a trained pre-trained model;
[0015] Update the fusion model according to the trained pre-trained model to obtain a speech synthesis attack protection model;
[0016] Input speech into the anti-speech synthesis attack model to implement speech synthesis detection.
[0017] Specifically, the pre-training model is constructed based on a convolutional neural network, a recurrent neural network and a fully connected layer, including a latent space feature module and a context feature module.
[0018] The backbone network of the latent space feature module and the context feature module includes a convolutional neural network unit, a long short-term memory unit and a fully connected prediction unit.
[0019] Specifically, the synthetic detection model is constructed based on a recurrent neural network and a fully connected layer.
[0020] Specifically, the training process of the model is:
[0021] Select a model to be trained, and select corresponding input data, output data and corresponding training loss function according to the model;
[0022] Set training hyperparameters and select training strategies; the hyperparameters include learning rate, batch size and maximum number of iterations, and the strategies include learning rate decay strategy, verification strategy and training termination strategy;
[0023] The model parameters are updated by gradient descent and back propagation algorithm until the error is stable.
[0024] Another aspect of the present invention provides a self-supervision-based anti-speech synthesis attack system, comprising:
[0025] Training sample acquisition unit: acquire speech data training samples;
[0026] Initial training unit: input new training samples of speech data into the pre-training model, extract hidden features, and further obtain context features based on the extracted hidden features. Then, reversely predict the hidden features through the context features to complete the self-supervised training of the pre-training model and obtain the pre-training model after initial training.
[0027] Speech synthesis unit: uses a variety of synthetic speech algorithms to construct a data set to prevent speech synthesis attacks;
[0028] Model fusion unit: The pre-trained model after initial training is fused with the synthesis detection model through the gated recurrent unit and the attention module to obtain a fusion model. The anti-speech synthesis attack dataset is input into the fusion model to complete the training of the fusion model and obtain the contribution value of each speech feature.
[0029] Retraining unit: input the speech features corresponding to the anti-speech synthesis attack data set and the set contribution value into the pre-trained model after initial training for retraining to obtain a trained pre-trained model;
[0030] Update unit: updates the fusion model according to the trained pre-trained model to obtain a speech synthesis attack protection model;
[0031] Detection unit: Input speech into the anti-speech synthesis attack model to implement speech synthesis detection.
[0032] Specifically, the pre-training model is constructed based on a convolutional neural network, a recurrent neural network and a fully connected layer, including a latent space feature module and a context feature module.
[0033] The backbone network of the latent space feature module and the context feature module includes a convolutional neural network unit, a long short-term memory unit and a fully connected prediction unit.
[0034] Specifically, the synthetic detection model is constructed based on a recurrent neural network and a fully connected layer.
[0035] Specifically, the training process of the model is:
[0036] Select a model to be trained, and select corresponding input data, output data and corresponding training loss function according to the model;
[0037] Set training hyperparameters and select training strategies; the hyperparameters include learning rate, batch size and maximum number of iterations, and the strategies include learning rate decay strategy, verification strategy and training termination strategy;
[0038] The model parameters are updated by gradient descent and back propagation algorithm until the error is stable.
[0039] On the other hand, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method steps for preventing speech synthesis attacks based on self-supervision when executing the computer program.
[0040] Yet another aspect of an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps for preventing speech synthesis attacks based on self-supervision are implemented.
[0041] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:
[0042] (1) The present invention proposes a new self-supervised training method, which first extracts hidden features, and then further obtains context features based on the extracted hidden features, and then reversely predicts the hidden features through the context features to complete the self-supervised training of the pre-trained model; and applies this self-supervised learning to synthetic speech detection, which can not only effectively utilize a large amount of unlabeled data, but also improve the generalization ability of synthetic detection, and can achieve good results when using a small amount of labeled data.
[0043] (2) The present invention fuses the pre-trained model after initial training with the synthetic detection model through a gated recurrent unit and an attention module to obtain a fusion model. On the one hand, it solves the gradient vanishing / exploding problem and retains the long-term information of the sequence at the same time. On the other hand, it can derive the contribution value in feature synthesis detection and retrain the pre-trained model according to the contribution value, further improving the accuracy of model feature extraction and the accuracy of synthetic detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is a diagram of an application environment of a self-supervised method for preventing speech synthesis attacks in an embodiment of the present invention;
[0045] Figure 2 A flow chart of a self-supervised method for preventing speech synthesis attacks in an embodiment of the present invention;
[0046] Figure 3 Schematic diagram of initial training of the pre-trained model provided in an embodiment of the present invention
[0047] Figure 4 A self-supervised anti-speech synthesis attack system architecture diagram provided by an embodiment of the present invention;
[0048] Figure 5A schematic diagram of an electronic device provided by an embodiment of the present invention;
[0049] Figure 6 A schematic diagram of an embodiment of a computer-readable storage medium provided in an embodiment of the present invention.
[0050] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION
[0051] The solution provided in the embodiments of the present application involves technologies such as artificial intelligence synthetic speech detection, which is specifically described by the following embodiments:
[0052] The present application provides a method for preventing speech synthesis attacks based on self-supervision, which can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The terminal 102 can control the server 104 to train the accent classification model. When the terminal 102 sends a training instruction to the server, the server 104 obtains training samples from the database, and the training data includes training audio; the server 104 inputs the speech data training samples into the pre-trained model, extracts hidden features, and further obtains context features based on the extracted hidden features, and then reversely predicts the hidden features through the context features to complete the self-supervised training of the pre-trained model and obtain the pre-trained model after initial training; a variety of synthetic speech algorithms are used to construct a data set to prevent speech synthesis attacks; the pre-trained model after initial training is fused with the synthetic detection model through the gated recurrent unit and the attention module to obtain a fusion model, and the anti-speech synthesis attack data set is input into the fusion model to complete the training of the fusion model, and obtain each speech The contribution value of the feature; the speech feature corresponding to the anti-speech synthesis attack data set and the set contribution value is input into the pre-trained model after the initial training for re-training to obtain a trained pre-trained model; the server 104 updates the fusion model according to the trained pre-trained model to obtain an anti-speech synthesis attack model; wherein the terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices, and the server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected by wired or wireless communication, and this application is not limited here.
[0053] like Figure 2, is a flow chart of a method for preventing speech synthesis attacks based on self-supervision provided by an embodiment of the present invention, specifically comprising:
[0054] S201: Acquire speech data training samples;
[0055] The training sample refers to the sample for training the pre-trained model. The speech data training sample in the embodiment of the present invention has no corresponding label. Specifically, the server can directly obtain the training sample from the database. The server can also collect the training sample from the Internet, for example, the server collects the training audio from the Internet, and the server can also obtain the training sample from the service provider providing data services.
[0056] S202: Input the speech data training sample into the pre-training model, extract hidden features, further obtain context features based on the extracted hidden features, and then reversely predict the hidden features through the context features to complete the self-supervised training of the pre-training model and obtain the pre-training model after initial training;
[0057] like Figure 3 The schematic diagram of the initial training of the pre-training model provided in the embodiment of the present invention is as follows: first, the speech data training sample X is input into the pre-training model, the hidden feature Z is extracted according to the convolutional neural network, and then the context feature C is extracted through another convolutional neural network. The innovative operation of the embodiment of the present invention is that the hidden feature Z is reversely predicted by the context feature C until the error is minimized, thereby completing the initial training of the pre-training model.
[0058] The pre-training model is constructed based on convolutional neural networks, recurrent neural networks and fully connected layers, including latent space feature modules and context feature modules.
[0059] The backbone network of the latent space feature module and the context feature module includes a convolutional neural network unit, a long short-term memory unit and a fully connected prediction unit.
[0060] S203: Using multiple synthetic speech algorithms to construct a speech synthesis attack protection dataset;
[0061] In this embodiment, the forged speech is synthesized using a Wavenet model or a Tacotron model, and a speech synthesis attack prevention data set is constructed. The data set is as diverse as possible. There is no need for a lot of each type of synthesized speech, but the synthesized speech algorithm is as diverse as possible so that the next step of model training can learn more effective features.
[0062] S204: fusing the pre-trained model after initial training with the synthesis detection model through the gated recurrent unit and the attention module to obtain a fusion model, inputting the anti-speech synthesis attack dataset into the fusion model, completing the training of the fusion model, and obtaining the contribution value of each speech feature;
[0063] The present invention fuses the pre-trained model after initial training with the synthetic detection model through a gated recurrent unit and an attention module to obtain a fusion model. On the one hand, it solves the gradient vanishing / exploding problem and retains the long-term information of the sequence at the same time. On the other hand, it can derive the contribution value in feature synthesis detection and retrain the pre-trained model according to the contribution value, further improving the accuracy of model feature extraction and the accuracy of synthetic detection.
[0064] Specifically, the synthetic detection model is constructed based on a recurrent neural network and a fully connected layer.
[0065] S205: inputting the speech features corresponding to the anti-speech synthesis attack data set and the set contribution value into the pre-trained model after initial training for re-training to obtain a trained pre-trained model;
[0066] Specifically, the training process of the model is:
[0067] Select a model to be trained, and select corresponding input data, output data and corresponding training loss function according to the model;
[0068] Set training hyperparameters and select training strategies; the hyperparameters include learning rate, batch size and maximum number of iterations, and the strategies include learning rate decay strategy, verification strategy and training termination strategy;
[0069] The model parameters are updated by gradient descent and back propagation algorithm until the error is stable.
[0070] S206: updating the fusion model according to the trained pre-trained model to obtain a speech synthesis attack prevention model;
[0071] S207: Input speech to the anti-speech synthesis attack model to implement speech synthesis detection.
[0072] like Figure 4 Another aspect of the present invention provides a self-supervision-based anti-speech synthesis attack system, including:
[0073] Training sample acquisition unit 401: acquires speech data training samples;
[0074] The training sample refers to the sample for training the pre-trained model. The speech data training sample in the embodiment of the present invention has no corresponding label. Specifically, the server can directly obtain the training sample from the database. The server can also collect the training sample from the Internet, for example, the server collects the training audio from the Internet, and the server can also obtain the training sample from the service provider providing data services.
[0075] Initial training unit 402: input new training samples of speech data into the pre-training model, extract hidden features, further obtain context features based on the extracted hidden features, and then reversely predict the hidden features through the context features to complete the self-supervised training of the pre-training model and obtain the pre-training model after initial training;
[0076] like Figure 3 The schematic diagram of the initial training of the pre-training model provided in the embodiment of the present invention is as follows: first, the speech data training sample X is input into the pre-training model, the hidden feature Z is extracted according to the convolutional neural network, and then the context feature C is extracted through another convolutional neural network. The innovative operation of the embodiment of the present invention is that the hidden feature Z is reversely predicted by the context feature C until the error is minimized, thereby completing the initial training of the pre-training model.
[0077] The pre-training model is constructed based on convolutional neural networks, recurrent neural networks and fully connected layers, including latent space feature modules and context feature modules.
[0078] The backbone network of the latent space feature module and the context feature module includes a convolutional neural network unit, a long short-term memory unit and a fully connected prediction unit.
[0079] Speech synthesis unit 403: uses a variety of synthetic speech algorithms to construct a speech synthesis attack protection data set;
[0080] In this embodiment, the forged speech is synthesized using a Wavenet model or a Tacotron model, and a speech synthesis attack prevention data set is constructed. The data set is as diverse as possible. There is no need for a lot of each type of synthesized speech, but the synthesized speech algorithm is as diverse as possible so that the next step of model training can learn more effective features.
[0081] Model fusion unit 404: fuses the pre-trained model after initial training with the synthesis detection model through the gated recurrent unit and the attention module to obtain a fusion model, inputs the anti-speech synthesis attack dataset into the fusion model, completes the training of the fusion model, and obtains the contribution value of each speech feature;
[0082] The present invention fuses the pre-trained model after initial training with the synthetic detection model through a gated recurrent unit and an attention module to obtain a fusion model. On the one hand, it solves the gradient vanishing / exploding problem and retains the long-term information of the sequence at the same time. On the other hand, it can derive the contribution value in feature synthesis detection and retrain the pre-trained model according to the contribution value, further improving the accuracy of model feature extraction and the accuracy of synthetic detection.
[0083] Specifically, the synthetic detection model is constructed based on a recurrent neural network and a fully connected layer.
[0084] Retraining unit 405: inputting the speech feature corresponding to the anti-speech synthesis attack data set and the set contribution value into the pre-trained model after initial training for retraining to obtain a trained pre-trained model;
[0085] Specifically, the training process of the model is:
[0086] Select a model to be trained, and select corresponding input data, output data and corresponding training loss function according to the model;
[0087] Set training hyperparameters and select training strategies; the hyperparameters include learning rate, batch size and maximum number of iterations, and the strategies include learning rate decay strategy, verification strategy and training termination strategy;
[0088] Update the model parameters by gradient descent and back propagation algorithm until the error is stable
[0089] Updating unit 406: updating the fusion model according to the trained pre-trained model to obtain a speech synthesis attack prevention model;
[0090] Detection unit 407: inputs speech to the anti-speech synthesis attack model to implement speech synthesis detection.
[0091] like Figure 5 As shown, an embodiment of the present invention provides an electronic device 500, including a memory 510, a processor 520, and a computer program 511 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 511, a method for preventing speech synthesis attacks based on self-supervision provided in an embodiment of the present invention is implemented.
[0092] Since the electronic device introduced in this embodiment is the device used to implement the embodiment of the present invention, based on the method introduced in the embodiment of the present invention, the technical personnel in this field can understand the specific implementation mode of the electronic device of this embodiment and its various variations. Therefore, how the electronic device implements the method in the embodiment of the present invention is not introduced in detail here. As long as the equipment used by the technical personnel in this field to implement the method in the embodiment of the present invention falls within the scope of protection of the present invention.
[0093] See also Figure 6 , Figure 6 A schematic diagram of an embodiment of a computer-readable storage medium provided in an embodiment of the present invention.
[0094] like Figure 6As shown, this embodiment provides a computer-readable storage medium 600, on which a computer program 611 is stored. When the computer program 611 is executed by a processor, a method for preventing speech synthesis attacks based on self-supervision provided by an embodiment of the present invention is implemented;
[0095] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0096] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0097] The present invention proposes a new self-supervised training method, which first extracts hidden features, and further obtains context features based on the extracted hidden features, and then reversely predicts the hidden features through the context features to complete the self-supervised training of the pre-trained model; and applies this self-supervised learning to synthetic speech detection, which can not only effectively utilize a large amount of unlabeled data, but also improve the generalization ability of synthetic detection, and can achieve good results when using a small amount of labeled data.
[0098] The present invention fuses the pre-trained model after initial training with the synthetic detection model through a gated recurrent unit and an attention module to obtain a fusion model. On the one hand, it solves the gradient vanishing / exploding problem and retains the long-term information of the sequence at the same time. On the other hand, it can derive the contribution value in feature synthesis detection and retrain the pre-trained model according to the contribution value, further improving the accuracy of model feature extraction and the accuracy of synthetic detection.
[0099] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The above is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. The various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features applied for herein.
[0100] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.
Claims
1. A method for preventing speech synthesis attacks based on self-supervision, characterized in that: include: Obtain speech data training samples; Input the speech data training sample into the pre-training model, extract the hidden features, and further obtain the context features according to the extracted hidden features, and then reversely predict the hidden features through the context features to complete the self-supervised training of the pre-training model, and obtain the pre-training model after initial training. The pre-training model is constructed based on a convolutional neural network, a recurrent neural network and a fully connected layer, including a latent space feature module and a context feature module. The backbone network of the latent space feature module and the context feature module includes a convolutional neural network unit, a long short-term memory unit and a fully connected prediction unit; Use a variety of synthetic speech algorithms to construct a data set to prevent speech synthesis attacks; The pre-trained model after initial training is fused with the synthesis detection model through the gated recurrent unit and the attention module to obtain a fusion model. The anti-speech synthesis attack dataset is input into the fusion model to complete the training of the fusion model and obtain the contribution value of each speech feature. The speech features corresponding to the anti-speech synthesis attack dataset and the set contribution value are input into the pre-trained model after initial training for re-training to obtain a trained pre-trained model; Update the fusion model according to the trained pre-trained model to obtain a speech synthesis attack protection model; Input speech into the anti-speech synthesis attack model to implement speech synthesis detection.
2. The method for preventing speech synthesis attacks based on self-supervision according to claim 1, characterized in that: The synthetic detection model is built based on a recurrent neural network and a fully connected layer.
3. The method for preventing speech synthesis attacks based on self-supervision according to claim 1, characterized in that: The training process of the model is: Select a model to be trained, and select corresponding input data, output data and corresponding training loss function according to the model; Set training hyperparameters and select training strategies; the hyperparameters include learning rate, batch size and maximum number of iterations, and the strategies include learning rate decay strategy, verification strategy and training termination strategy; The model parameters are updated by gradient descent and back propagation algorithm until the error is stable.
4. A self-supervised anti-speech synthesis attack system, characterized in that: include: Training sample acquisition unit: acquire speech data training samples; Initial training unit: input new training samples of speech data into the pre-training model, extract hidden features, and further obtain context features based on the extracted hidden features, and then reversely predict hidden features through context features to complete the self-supervised training of the pre-training model to obtain the pre-training model after initial training. The pre-training model is constructed based on convolutional neural networks, recurrent neural networks and fully connected layers, including latent space feature modules and context feature modules. The backbone network of the latent space feature module and the context feature module includes a convolutional neural network unit, a long short-term memory unit and a fully connected prediction unit; Speech synthesis unit: uses a variety of synthetic speech algorithms to construct a data set to prevent speech synthesis attacks; Model fusion unit: The pre-trained model after initial training is fused with the synthesis detection model through the gated recurrent unit and the attention module to obtain a fusion model. The anti-speech synthesis attack dataset is input into the fusion model to complete the training of the fusion model and obtain the contribution value of each speech feature. Retraining unit: input the speech features corresponding to the anti-speech synthesis attack data set and the set contribution value into the pre-trained model after initial training for retraining to obtain a trained pre-trained model; Update unit: updates the fusion model according to the trained pre-trained model to obtain a speech synthesis attack protection model; Detection unit: Input speech into the anti-speech synthesis attack model to implement speech synthesis detection.
5. The self-supervised anti-speech synthesis attack system according to claim 4, characterized in that: The synthetic detection model is built based on a recurrent neural network and a fully connected layer.
6. The self-supervised anti-speech synthesis attack system according to claim 4, characterized in that: The training process of the model is: Select a model to be trained, and select corresponding input data, output data and corresponding training loss function according to the model; Set training hyperparameters and select training strategies; the hyperparameters include learning rate, batch size and maximum number of iterations, and the strategies include learning rate decay strategy, verification strategy and training termination strategy; The model parameters are updated by gradient descent and back propagation algorithm until the error is stable.
7. An electronic device, characterized in that: include: A memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method steps described in any one of claims 1 to 3 when executing the computer program.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of claims 1 to 3 are implemented.