Synthetic speech detection method and system, computer equipment, and storage medium

Abstract

The embodiment of the present application provides a method and system for synthetic speech detection, a computer device, and a storage medium, which belongs to the field of financial technology. The method includes: inputting the acquired sample speech and sample audio labels into a preset initial network model, which includes a convolution layer, a residual network layer, and a classification layer; extracting features of the sample speech according to the convolution layer to obtain sample audio features; processing the sample audio features according to the residual network layer to obtain sample transformation features; predicting and classifying the sample transformation features according to the classification layer to obtain predicted classification data, which includes predicted audio labels; adjusting the parameters of the initial network model according to the sample audio labels and the predicted audio labels to obtain a speech detection model; inputting the target audio to be detected into the speech detection model for speech detection to obtain a speech category. The embodiment of the present application can avoid the influence of manual features and improve the detection accuracy of synthetic speech.

Description

Technical Field

[0001] The present application relates to the field of financial technology, and in particular to a synthetic speech detection method and system, computer equipment, and storage medium. Background Art

[0002] With the advancement of science and technology, and the gradual maturity of pattern recognition technology in the information security context of FinTech, biometric identification of individuals based on biological characteristics has begun to be applied and promoted in the field of identity recognition. This means that the application of biometric technology in daily life and security is becoming increasingly widespread. Voiceprint recognition, as an emerging biometric technology, is increasingly being used in the financial and security fields. However, with the development of speech synthesis technology, synthesizing audio with a specific timbre is becoming increasingly easier, making existing speech recognition solutions difficult to defend against synthetic voice attacks. In related technologies, voiceprint verification is achieved by synthesizing speech that simulates the timbre of a specific target subject. Therefore, research on synthetic voice detection methods is currently a key focus. Synthetic voice refers to speech generated through mechanical or electronic artificial speech technology with acoustic characteristics similar to real speech. Related synthetic voice detection systems generally consist of a front-end feature extractor and a back-end binary classifier. The discriminative features used in the front-end are often artificially generated features, such as spectral features and phase features. However, the artificial feature generation process can lose some information, seriously affecting the detection accuracy of synthetic speech. Therefore, how to provide a method that can avoid the influence of manual features and improve the detection accuracy of synthesized speech has become a technical problem that needs to be solved urgently. Summary of the Invention

[0003] The main purpose of the embodiments of the present application is to propose a synthetic speech detection method and system, computer equipment, and storage medium, which can avoid the influence of manual features and improve the detection accuracy of synthetic speech.

[0004] To achieve the above objectives, a first aspect of an embodiment of the present application provides a method for detecting synthesized speech, the method comprising:

[0005] In some embodiments, obtaining a sample speech and a sample audio tag of the sample speech;

[0006] Inputting the sample speech and the sample audio tag into a preset initial network model; wherein the initial network model includes a convolutional layer, a residual network layer and a classification layer;

[0007] Extracting features of the sample speech according to the convolutional layer to obtain sample audio features;

[0008] Performing feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features;

[0009] Performing prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels;

[0010] Adjusting parameters of the initial network model according to the sample audio label and the predicted audio label to obtain a speech detection model;

[0011] The target audio to be detected is input into the speech detection model for speech detection to obtain a speech category, where the speech category is used to characterize whether the target audio is synthesized speech or non-synthesized speech.

[0012] In some embodiments, the residual network layer includes a Ghost module and a first fully connected module, and performing feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features includes:

[0013] Performing linear transformation processing on the sample audio feature according to the Ghost module to obtain a first processed feature;

[0014] performing convolution transformation processing on the sample audio feature according to the first fully connected module to obtain a second processed feature;

[0015] Residual calculation is performed on the first processing feature and the second processing feature to obtain the sample transformation feature.

[0016] In some embodiments, the Ghost module includes a standard convolution submodule, an expanded group convolution submodule, and a correction submodule, and performing linear transformation processing on the sample audio feature according to the Ghost module to obtain the first processed feature includes:

[0017] Performing standard convolution processing on the sample audio feature according to the standard convolution submodule to obtain a first convolution feature;

[0018] Performing dilated group convolution processing on the first convolution feature according to the dilated group convolution submodule to obtain a second convolution feature;

[0019] Performing feature concatenation on the first convolution feature and the second convolution feature to obtain a third convolution feature;

[0020] The third convolution feature is corrected according to the correction submodule to obtain the first processed feature.

[0021] In some embodiments, performing correction processing on the third convolution feature according to the correction submodule to obtain the first processed feature includes:

[0022] Performing batch normalization on the third convolution feature according to the correction submodule to obtain a fourth convolution feature;

[0023] Feature calculation is performed on the fourth convolution feature according to a preset activation function to obtain a first processed feature.

[0024] In some embodiments, performing residual calculation on the first processing feature and the second processing feature to obtain a sample transformation feature includes:

[0025] Performing residual calculation on the first processing feature and the second processing feature to obtain a sample residual feature;

[0026] Pooling is performed on the sample residual features to obtain the sample transformation features.

[0027] In some embodiments, the classification layer includes a second fully connected module, and performing prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data includes:

[0028] Performing feature conversion processing on the sample transformation feature according to the second fully connected module to obtain a sample transformation feature;

[0029] Normalization processing is performed on the sample conversion features to obtain predicted classification data, and the predicted classification data is used to count the predicted classification values ​​of different sample audio labels to which the sample speech belongs.

[0030] In some embodiments, extracting features from the sample speech according to the convolutional layer to obtain sample audio features includes:

[0031] Extracting features of the sample speech according to the convolutional layer to obtain initial features of the sample;

[0032] Pooling is performed on the initial features of the sample to obtain the sample audio features.

[0033] A second aspect of the embodiments of the present application provides a synthetic speech detection system, the system comprising:

[0034] A sample audio acquisition module is used to acquire sample speech and sample audio tags of the sample speech;

[0035] A model acquisition module, configured to input the sample speech and the sample audio label into a preset initial network model; wherein the initial network model includes a convolutional layer, a residual network layer, and a classification layer;

[0036] A feature extraction module, configured to extract features from the sample speech according to the convolutional layer to obtain sample audio features;

[0037] A feature transformation module, configured to perform feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features;

[0038] A prediction and classification module, configured to perform prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels;

[0039] A parameter adjustment module, configured to adjust parameters of the initial network model according to the sample audio label and the predicted audio label to obtain a speech detection model;

[0040] The speech detection module is used to input the target audio to be detected into the speech detection model for speech detection to obtain a speech category, where the speech category is used to characterize whether the target audio is synthesized speech or non-synthesized speech.

[0041] A third aspect of an embodiment of the present application provides a computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor is used to execute a method as described in any one of the embodiments of the first aspect of the present application.

[0042] The fourth aspect of the embodiments of the present application proposes a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by a computer, the computer is used to execute the method described in any one of the embodiments of the first aspect of the present application.

[0043] The synthetic speech detection method and system, computer device, and storage medium proposed in the embodiments of the present application first obtain sample speech and sample audio labels of the sample speech, and input the sample speech and sample audio labels into a preset initial network model, wherein the initial network model includes a convolutional layer, a residual network layer, and a classification layer. Feature extraction is performed on the sample speech according to the convolutional layer to obtain sample audio features; feature transformation processing is performed on the sample audio features according to the residual network layer to obtain sample transformation features. Prediction and classification processing is performed on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels. Afterwards, the parameters of the initial network model are adjusted according to the sample audio labels and the predicted audio labels to obtain a speech detection model, which can improve the detection accuracy of synthetic speech. Finally, the target audio to be detected is input into the speech detection model for speech detection to obtain a speech category, which is used to characterize whether the target audio is synthetic speech or non-synthetic speech. The embodiment of the present application constructs an initial network model including a convolutional layer, a residual network layer, and a classification layer, so that when the sample speech is predicted and classified according to the initial network model, the influence of manual features on the prediction and classification processing can be avoided. The parameters of the initial network model are adjusted by sample audio labels and preset audio labels to obtain a speech detection model with the same structure as the initial network model but higher detection accuracy. It can be seen that when speech detection is performed on the target audio according to the speech detection model provided in the embodiment of the present application, speech detection can be performed on the target audio that does not contain manual features, thereby improving the detection accuracy of whether the target audio is a synthetic speech. For example, in the information security scenario of financial technology, the speech of the target object can be detected by using the speech detection model of the present application to avoid the influence of synthetic speech on the security of the target object's account information, thereby improving the security of voiceprint recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a first flow chart of the synthetic speech detection method provided in an embodiment of the present application;

[0045] Figure 2 yes Figure 1 Flowchart of step S130 in FIG.

[0046] Figure 3 yes Figure 1 Flowchart of step S140 in FIG.

[0047] Figure 4 Schematic diagram of the module structure of the Ghost-ResNet module provided in an embodiment of the present application;

[0048] Figure 5 yes Figure 3 Flowchart of step S310 in FIG.

[0049] Figure 6 yes Figure 5 Flowchart of step S540 in FIG.

[0050] Figure 7 yes Figure 3 Flowchart of step S330 in FIG.

[0051] Figure 8 yes Figure 1 Flowchart of step S150 in FIG.

[0052] Figure 9 This is a block diagram of the module structure of the synthetic speech detection system provided in an embodiment of the present application;

[0053] Figure 10 This is a schematic diagram of the hardware structure of the computer device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0055] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0057] In addition, the described features, structures or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, well-known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present disclosure.

[0058] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0059] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0060] First, let’s analyze some of the terms used in this application:

[0061] Artificial Intelligence (AI) is a new technical discipline that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. A branch of computer science, AI seeks to understand the essence of intelligence and create new intelligent machines that can respond in a manner similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thinking. It also encompasses theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results.

[0062] Gradient disappearance: In a neural network, when the learning rate of the current hidden layer is lower than the learning rate of the subsequent hidden layer, that is, as the number of hidden layers increases, the classification accuracy decreases instead. This phenomenon is called gradient disappearance.

[0063] Residual Connections: Superposition of input and nonlinear changes of input.

[0064] Pooling: It is a form of downsampling that compresses the input feature image. On the one hand, it reduces the image features, resulting in a reduction in parameters, thereby simplifying the complexity of convolutional network calculations; on the other hand, it maintains a certain invariance of the features (rotation, translation, scaling, etc.).

[0065] Batch Normalization (BN): Also known as batch normalization, it is a technique used to improve the performance and stability of artificial neural networks.

[0066] With the development of science and technology, and the gradual maturity of pattern recognition technology in the information security context of FinTech, biometric identification of individuals based on biological characteristics has begun to be applied and promoted in the field of identity recognition. This means that the application of biometric technology in daily life and security is becoming increasingly widespread. Voiceprint recognition, as an emerging biometric technology, is increasingly being used in the financial and security fields. However, with the advancement of speech synthesis technology, synthesizing audio with specific timbre is becoming increasingly easier, making existing speech recognition solutions difficult to defend against synthetic voice attacks. In related technologies, voiceprint verification is achieved by synthesizing speech that mimics the timbre of a specific target subject. Therefore, research on synthetic voice detection methods is currently a key focus. Synthetic voice refers to speech generated through mechanical or electronic artificial speech technology with acoustic characteristics similar to real speech. In recent years, with the rapid development of deep learning technology, the naturalness of synthesized speech has become increasingly higher. Related synthetic voice detection systems generally consist of a front-end feature extractor and a back-end binary classifier. The discriminative features in the front-end are often artificially generated features, such as spectral features and phase features. Designing network structures, loss functions, and training methods can improve the performance of synthesized speech detection models, but the model's potential ultimately depends on the information captured in the initial features. However, the process of manually crafting features can result in the loss of some information, severely impacting the accuracy of synthesized speech detection. Therefore, developing a method that can mitigate the impact of handcrafted features and improve synthesized speech detection accuracy has become a pressing technical challenge.

[0067] Based on this, the embodiments of the present application propose a synthetic speech detection method and system, a computer device, and a storage medium, which can improve the detection accuracy of synthetic speech.

[0068] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results.

[0069] Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0070] The synthetic speech detection method provided in the embodiment of the present application relates to the field of artificial intelligence. The synthetic speech detection method provided in the embodiment of the present application can be applied to a terminal, can be applied to a server side, or can be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer or a smart watch, etc.; the server side can be configured as an independent physical server, or can be configured as a server cluster or a distributed system composed of multiple physical servers, or can be configured as 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, content delivery networks (CDNs) and big data and artificial intelligence platforms; the software can be an application that implements the synthetic speech detection method, etc., but is not limited to the above forms.

[0071] Embodiments of the present application can be used in numerous general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0072] It should be noted that in each specific embodiment of the present application, when it comes to the need to perform relevant processing based on data related to the user's identity or characteristics, such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first, and the collection, use, and processing of such data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or consent, the necessary user-related data for the normal operation of the embodiment of the present application will be obtained.

[0073] Please refer to Figure 1 , Figure 1This is an optional flow chart of the synthetic speech detection method provided in the embodiment of the present application. In some embodiments, Figure 1 The method may specifically include but is not limited to steps S110 to S170. Figure 1 These seven steps are introduced in detail.

[0074] Step S110, obtaining a sample speech and a sample audio tag of the sample speech;

[0075] Step S120: inputting the sample speech and sample audio labels into a preset initial network model; wherein the initial network model includes a convolutional layer, a residual network layer, and a classification layer;

[0076] Step S130, extracting features from the sample speech using the convolutional layer to obtain sample audio features;

[0077] Step S140, performing feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features;

[0078] Step S150, performing prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels;

[0079] Step S160, adjusting parameters of the initial network model according to the sample audio label and the predicted audio label to obtain a speech detection model;

[0080] In step S170 , the target audio to be detected is input into a speech detection model for speech detection to obtain a speech category, where the speech category is used to characterize whether the target audio is synthesized speech or non-synthesized speech.

[0081] It should be noted that the synthetic speech detection method provided in the embodiment of the present application can be executed by the terminal or the server respectively, or by the terminal and the server in collaboration. Among them, the terminal communicates with the server through the network to realize end-to-end synthetic speech detection. For example, the terminal execution corresponding to the synthetic speech detection method is taken as an example for explanation: the terminal obtains the sample speech to be trained and the sample audio label of the sample speech from the local, and the sample speech and the sample audio label are input into the preset initial network model, which includes a convolution layer, a residual network layer and a classification layer. The sample speech is subjected to feature extraction according to the convolution layer to obtain sample audio features. Since there is a lot of redundant information in the features output by the convolution, the sample audio features are subjected to feature transformation processing according to the residual network layer to obtain sample transformation features. Afterwards, the sample transformation features are subjected to prediction classification processing according to the classification layer to obtain predicted classification data, which includes predicted audio labels. And the parameters of the initial network model are adjusted according to the sample audio labels and the predicted audio labels to obtain a speech detection model. After the speech detection model is trained, the target audio to be detected is received and input into the speech detection model for speech detection to obtain a speech category for characterizing the target audio as synthesized speech or non-synthesized speech. The trained speech detection model can also be deployed on the server so that the server can also implement the steps of the above-mentioned synthetic speech detection method.

[0082] Terminals can be smartphones, tablets, laptops, desktop computers, smart speakers, smart watches, etc.

[0083] In step S110 of some embodiments, to train the speech detection model, first, the terminal or server obtains a training sample set, which includes at least one training data, and the training data is used to train the model. The training data includes sample speech and sample audio labels of the sample speech. In order to enable the constructed speech detection model to accurately identify synthesized speech and non-synthesized speech, the sample audio labels include positive labels and negative labels. The positive label can be used to indicate that the sample speech belongs to non-synthesized speech, and the negative label can be used to indicate that the sample speech belongs to synthesized speech. The sample speech can be in the form of an audio spectrum such as a mel spectrum.

[0084] It should be noted that the synthetic speech detection method proposed in the embodiment of the present application can be applied in different fields, such as access control systems, telephone banking, etc.

[0085] It should be noted that in actual applications, for example, when the synthetic voice detection method of the present application is used in an access control system, the sample voice can be the preset access control voice of the target object, and the sample audio label corresponding to the preset access control voice is a positive label; the sample voice can also be the synthetic access control voice of the target object, and the synthetic access control voice is used to represent the voice generated by imitating the target object's speaking style, intonation and other information, and the sample audio label corresponding to the synthetic access control voice is a negative label. In addition, when the synthetic voice detection method of the present application is used in the technology of using telephone banking for information confirmation, the sample voice can be the confirmation voice of the target object, and the sample audio label corresponding to the confirmation voice is a positive label; the sample voice can also be the synthetic confirmation voice of the target object, and the synthetic confirmation voice is used to represent the confirmation voice generated by imitating the target object's speaking style, intonation and other information, and the sample audio label corresponding to the synthetic confirmation voice is a negative label. Therefore, the sample voice can be selected according to different actual applications and is not specifically limited here.

[0086] It should be noted that the training sample set used in the embodiment of the present application can be a training set such as ASVspoof2015, ASVspoof2019, etc., or a training set constructed according to actual needs.

[0087] In step S120 of some embodiments, in order to optimize the network structure to solve the problem of gradient disappearance during sample training and extract richer feature information, the present application pre-constructs an initial network model, which includes a convolutional layer, a residual network layer and a classification layer, and inputs the sample speech and sample audio labels into the preset initial network model for processing.

[0088] In step S130 of some embodiments, to extract audio feature information from the input speech sample, the sample speech is first subjected to feature extraction using a convolutional layer to obtain sample audio features. The sample audio features may be an audio feature sequence or a mel-spectrogram feature in the frequency domain. By performing feature extraction on the sample speech using a convolutional layer, the present embodiment can avoid the influence of manual features on speech detection.

[0089] Please refer to Figure 2 , Figure 2 This is a flowchart of the specific method of step S130 provided by this application. In some embodiments, step S130 may specifically include but is not limited to step S210 and step S220. Figure 2 These two steps are introduced in detail.

[0090] Step S210, extracting features from the sample speech using the convolutional layer to obtain initial features of the sample;

[0091] Step S220: Pooling is performed on the initial features of the sample to obtain the sample audio features.

[0092] In steps S210 and S220 of some embodiments, since there is a large amount of redundant information in the features output by the convolution layer, in order to reduce the feature dimension, remove redundant information and reduce the amount of calculation, a maximum pooling layer can be added at the output end of the convolution layer to retain more effective feature information. Specifically, the audio of the input sample speech is subjected to feature extraction according to the convolution layer to obtain the initial features of the sample. Then, the initial features of the sample are pooled to obtain the sample audio features. The embodiment of the present application can improve the efficiency of model training and prevent the problem of model overfitting by adding a maximum pooling layer at the output end of the convolution layer.

[0093] In step S140 of some embodiments, since deep neural network models are prone to problems such as gradient disappearance during training, in order to improve the training efficiency of the network, the sample audio features are subjected to feature transformation processing according to the residual network layer to obtain sample transformation features. The embodiment of the present application can effectively solve the gradient disappearance problem in the propagation process by introducing residual connections, thereby optimizing the network structure of the model.

[0094] It is understandable that when the synthetic speech detection method of the present application is used in the technology of confirming information using telephone banking, the sample speech can be the confirmation speech of the target object, and the sample audio label corresponding to the confirmation speech is a positive label; the sample speech can also be the synthesized confirmation speech of the target object, and the sample audio label corresponding to the synthesized confirmation speech is a negative label. The confirmation speech of the target object is subjected to feature extraction according to the convolutional layer to obtain the sample initial features, and the sample initial features are pooled to obtain the sample audio features. The embodiment of the present application can avoid the influence of manual features.

[0095] Please refer to Figure 3 , Figure 3 This is a flowchart of the specific method of step S140 provided by this application. In some embodiments, the residual network layer includes a Ghost module and a first fully connected module. Step S140 may specifically include but is not limited to steps S310 to S330. Figure 3 These three steps are introduced in detail.

[0096] Step S310: performing linear transformation processing on the sample audio feature according to the Ghost module to obtain a first processed feature;

[0097] Step S320: performing convolution transformation processing on the sample audio features according to the first fully connected module to obtain second processed features;

[0098] Step S330 , performing residual calculation on the first processing feature and the second processing feature to obtain a sample transformation feature.

[0099] In step S310 of some embodiments, since the Ghost module can obtain richer feature information without changing the number of output channels, the residual network layer of the present application may include a Ghost module and a first fully connected module, and the sample audio features are linearly transformed according to the Ghost module to obtain the first processed features.

[0100] In step S320 of some embodiments, in order to change the number of convolution output channels to achieve a change in the output dimension and thus reduce the dimensionality of the first processing feature, the sample audio feature is subjected to convolution transformation processing according to the first fully connected module to obtain a second processing feature. Among them, the first fully connected module can use 1*1 convolution to adjust the number of feature sequence channels of different branches to be consistent and increase the nonlinearity of the network so that the model can express more complex local features. In addition, the first fully connected module can also use a global average pooling layer for operation.

[0101] In step S330 of some embodiments, a residual calculation is performed on the first processing feature and the second processing feature to obtain a sample transformation feature. Specifically, the first processing feature and the second processing feature are nonlinearly superimposed to obtain the sample transformation feature. In this embodiment, the introduction of residual connections can effectively solve the gradient vanishing problem during the propagation process and optimize the network structure of the model.

[0102] It should be noted that the residual network layer of the embodiment of the present application can be represented as a form of connection of M Ghost-ResNet modules, wherein each Ghost-ResNet module includes a Ghost module and a first fully connected module, that is, each Ghost-ResNet module can be sequentially connected to construct a residual network layer, and the sample audio features are subjected to feature transformation processing according to the residual network layer to obtain sample transformation features. The Ghost-ResNet module may include at least one Ghost module and at least one first fully connected module. The embodiment of the present application can achieve end-to-end synthetic speech detection by stacking Ghost-ResNet modules, thereby improving the detection accuracy of synthetic speech.

[0103] Please note that, please refer to Figure 4, the Ghost-ResNet module 400 constructed in the embodiment of the present application can be a connection form of two Ghost modules and two fully connected modules, wherein the fully connected module can be operated using a 1*1 convolution layer. Specifically, after the sample speech is feature extracted according to the convolution layer 410 to obtain the sample audio feature, the sample audio feature is linearly transformed according to the first Ghost module 420 to obtain the first processing feature. The sample audio feature is convolutionally transformed according to the first fully connected layer 430 to obtain the second processing feature, and the first processing feature is convolutionally transformed according to the second fully connected layer 440 to obtain the third processing feature. Afterwards, the second Ghost module 450 performs linear transformation on the third processing feature to obtain the fourth processing feature. Finally, the residual calculation is performed on the second processing feature and the fourth processing feature to obtain the sample transformation feature, and the obtained sample transformation feature is input into the classification layer 460 for predictive classification processing. The embodiment of the present application connects the Ghost module and the fully connected module to form a residual network by residual connection, which can help the model effectively capture the discriminative features for identifying the authenticity of speech, thereby improving the detection ability and generalization ability of the model.

[0104] Please refer to Figure 5 , Figure 5 This is a flowchart of the specific method of step S310 provided by this application. In some embodiments, the Ghost module includes a standard convolution submodule, an expansion group convolution submodule and a correction submodule. Step S310 may specifically include but is not limited to steps S510 to S540. Figure 5 These four steps are introduced in detail.

[0105] In step S510, a standard convolution process is performed on the sample audio feature according to the standard convolution submodule to obtain a first convolution feature;

[0106] In step S520, the first convolution feature is subjected to dilated group convolution processing according to the dilated group convolution submodule to obtain a second convolution feature;

[0107] In step S530, the first convolution feature and the second convolution feature are concatenated to obtain a third convolution feature;

[0108] In step S540, the third convolution feature is corrected according to the correction submodule to obtain the first processed feature.

[0109] In step S510 of some embodiments, since the Ghost module can obtain richer feature information without changing the number of output channels, the embodiments of the present application connect the Ghost module and the fully connected module to a residual network by residual connection, wherein the Ghost module can specifically include a standard convolution submodule, an expanded group convolution submodule and a correction submodule. Specifically, after obtaining the sample audio features output by the convolution layer, first, the sample audio features are subjected to standard convolution processing according to the standard convolution submodule to obtain a first convolution feature. That is, the standard convolution processing can be represented as sliding a convolution kernel on the input feature sequence at a specified step size, traversing each pixel in the input feature sequence. For each step size, the convolution kernel will have an overlapping area with the input feature sequence, and the corresponding elements of the overlapping area are multiplied, summed and then added with the bias term to obtain a pixel of the output feature. For example, the sequence of the input sample audio features is denoted as X, and the size of the sequence X is expressed as l×c, where l is used to represent the length of the input sequence X in the time dimension, c represents the number of input channels of the input sequence X, and n represents the number of output channels. The sequence X is input into the standard convolution submodule with the convolution kernel k for standard convolution processing to obtain the first convolution feature with reduced channel number, which is recorded as the eigensequence Y'. The size of the eigensequence Y' is expressed as l'×m, where m represents the number of channels of the eigensequence.

[0110] In step S520 of some embodiments, in order to obtain the output of n channels corresponding to the first processing feature, let m = n / s, s is an integer, and m is usually much smaller than n. The first convolution feature is subjected to dilated grouped convolution processing according to the dilated grouped convolution submodule to obtain a second convolution feature. Specifically, the eigensequence Y' is linearly transformed according to the dilated grouped convolution submodule, and the dilation coefficient of the dilated grouped convolution submodule increases layer by layer, as shown in formula (1). The first convolution feature, i.e., each channel of the eigensequence Y', is linearly transformed to generate s-1 feature channels, where s represents the total number of mappings generated by each channel in Y'.

[0111]

[0112] Among them, y i ' represents the characteristics of the i-th channel in the eigensequence Y', Φ i,j Indicates the effect on y i 'th linear operation, we get a sequence of second convolution features with m×(s-1) feature channels, which can be expressed as [y 11 ,y 12 ,…,y n(s-1) ]. Each eigensequence Y' is linearly transformed through s-1 channels with a convolution kernel of d.

[0113] In step S530 of some embodiments, since there is a large amount of redundant information in the features output by the convolution layer, certain channel features have a high degree of similarity. The Ghost module adopted in the embodiment of the present application divides the mapping calculation process into two parts, including standard convolution processing, i.e., identity transformation, and dilated grouped convolution processing, i.e., multiple linear transformations. Specifically, the identity transformation result of the standard convolution submodule and the result of multiple linear transformations by the dilated grouped convolution submodule are spliced, that is, the eigenvalue sequence Y' and the second convolution feature are feature spliced ​​to obtain a third convolution feature, which can be recorded as Y, the output channel number n of Y. Compared with traditional standard convolution, the Ghost module constructed in the embodiment of the present application can obtain richer feature information without changing the number of output channels, and reduce the number of parameters and calculations of the model.

[0114] In step S540 of some embodiments, in order to make the model training more stable and reduce the number of model training failures, the third convolution feature is corrected according to the correction submodule to obtain the first processed feature.

[0115] Please refer to Figure 6 , Figure 6 This is a flowchart of the specific method of step S540 provided by this application. In some embodiments, step S540 may specifically include but is not limited to step S610 and step S620. Figure 6 These two steps are introduced in detail.

[0116] Step S610, performing batch normalization processing on the third convolution feature according to the correction submodule to obtain a fourth convolution feature;

[0117] Step S620: Perform feature calculation on the fourth convolution feature according to a preset activation function to obtain a first processed feature.

[0118] In steps S610 and S620 of some embodiments, the correction submodule includes batch normalization processing and activation function processing. Specifically, the third convolution feature is batch normalized according to the correction submodule to obtain a fourth convolution feature, and the fourth convolution feature is feature calculated according to a preset activation function to obtain a first processing feature, wherein the preset activation function can be a rectified linear unit (Rectified Linear Unit, ReLU) function or a Sigmoid function, etc., which are not specifically limited here. The embodiment of the present application constructs a Ghost-ResNet module, that is, connects the Ghost module and the fully connected module in a residual connection manner, which can help the model capture the discriminative features for identifying the authenticity of speech, improve the detection ability and generalization ability of the model, and thus improve the detection accuracy of synthetic speech.

[0119] Please refer to Figure 7 , Figure 7 This is a flowchart of the specific method of step S330 provided by this application. In some embodiments, step S330 may specifically include but is not limited to step S710 and step S720. Figure 7 These two steps are introduced in detail.

[0120] Step S710, performing residual calculation on the first processing feature and the second processing feature to obtain a sample residual feature;

[0121] Step S720: performing pooling processing on the sample residual features to obtain the sample transformation features.

[0122] In some embodiments, in steps S710 and S720, to compress features, simplify the complexity of the model network, and improve model training efficiency, embodiments of the present application connect a pooling layer to the output of each Ghost-ResNet module. Specifically, residuals are calculated on the first processed features and the second processed features to obtain sample residual features, and the sample residual features are pooled to obtain the sample transformed features.

[0123] It should be noted that the residual network layer of the embodiment of the present application can be expressed as a connection of M Ghost-ResNet modules, and the output ends of the first M-1 Ghost-ResNet modules can be connected to a maximum pooling layer respectively, which can effectively downsample the features and reduce the amount of model calculation. A global maximum pooling layer can be connected to the output end of the Mth Ghost-ResNet module, which can retain more texture information of the features and improve the detection accuracy of the synthesized speech.

[0124] In step S150 of some embodiments, in order to accurately train a model capable of detecting synthesized speech, after obtaining the sample transformation features, the sample transformation features are subjected to predictive classification processing according to the classification layer to obtain predicted classification data, wherein the predicted classification data includes a predicted audio label, which is used to characterize the audio label obtained after the model performs synthesized speech detection on the input sample speech. The predicted audio label can be any of a preset positive label or negative label. Specifically, the sample transformation features are subjected to predictive classification processing according to the classification layer to obtain a predicted classification value for each preset label. By numerically comparing all the predicted classification values, the label corresponding to the largest predicted classification value is determined as the predicted audio label.

[0125] Please refer to Figure 8 , Figure 8This is a flowchart of the specific method of step S150 provided by this application. In some embodiments, the classification layer includes a second fully connected module, and step S150 may specifically include but is not limited to step S810 and step S820. Figure 8 These two steps are introduced in detail.

[0126] Step S810: performing feature conversion processing on the sample transformation feature according to the second fully connected module to obtain the sample transformation feature;

[0127] Step S820: normalize the sample conversion features to obtain predicted classification data, which is used to calculate the predicted classification values ​​of different sample audio labels to which the sample speech belongs.

[0128] In steps S810 and S820 of some embodiments, the second fully connected module performs feature conversion on the sample transformation features to obtain sample transformation features, thereby reducing the influence of feature position on the classification results and improving the robustness of the entire model network. The sample transformation features are then normalized to obtain predicted classification data, which is used to calculate the predicted classification values ​​of different sample audio labels to which the sample speech belongs.

[0129] It should be noted that the second fully connected module may include at least two fully connected layers to more deeply map the obtained sample transformation features to the sample label space, thereby reducing the influence of feature position on the classification results and improving the robustness of the entire network of the model.

[0130] It should be noted that the sample conversion features can be normalized using a Softmax function or a normalized exponential function.

[0131] In step S160 of some embodiments, a model loss value is determined based on the sample audio label and the predicted audio label, and the model parameters of the entire initial network model are adjusted based on the model loss value. When the initial network model meets the preset end condition, a speech detection model is obtained.

[0132] It should be noted that the preset training end condition can be when the model loss value of the speech detection model is less than or equal to the preset loss value threshold, or when the initial recognition accuracy of the speech detection model is greater than or equal to the preset accuracy threshold.

[0133] In step S170 of some embodiments, the terminal obtains the target audio to be detected locally, or obtains the target audio to be detected from the server, performs speech detection on the target audio through the trained speech detection model, and obtains a speech category, which is used to characterize the target audio as synthesized speech or non-synthesized speech.

[0134] It should be noted that when the synthetic speech detection method of the present application is executed on the terminal, after the target audio to be detected is input into the speech detection model, the terminal sends a detection signal to the speech detection model, and the speech detection model responds to the detection signal to start synthetic speech detection on the input target audio, and returns the detected speech category to the terminal.

[0135] Exemplarily, for example, in the telephone banking scenario of financial technology, when information confirmation is required, the sample voice can be the confirmation voice of the target object and the synthesized confirmation voice of the target object. The sample audio label corresponding to the confirmation voice is a positive label, and the sample audio label corresponding to the synthesized confirmation voice is a negative label. For example, the confirmation voice of the target object is "confirmed as myself", and the synthesized confirmation voice is a voice that simulates the pronunciation of the target object through speech synthesis technology, that is, it is also "confirmed as myself". The speech detection model trained in the embodiment of the present application includes a convolutional layer, a residual network layer and a classification layer, wherein the residual network layer is in the form of M Ghost-ResNet modules connected, and the output ends of the first M-1 Ghost-ResNet modules are connected to a maximum pooling layer, which can effectively downsample the features and reduce the amount of model calculation. A global maximum pooling layer can be connected to the output end of the Mth Ghost-ResNet module, which can retain more texture information of the features. Each Ghost-ResNet module can include at least one Ghost module and at least one first fully connected module, and each Ghost module includes a standard convolution submodule, an expanded group convolution submodule and a correction submodule. The classification layer of the embodiment of the present application includes at least two fully connected layers and a normalization process. The speech detection model obtained through training can perform speech synthesis detection on the input target speech to determine whether the target speech is the confirmation speech of the target object or a synthesized confirmation speech generated by imitating the sound characteristics of the target object. When the target speech is a synthesized confirmation speech, the server is notified to suspend subsequent information confirmation to ensure the security of the target object's account information; when the target speech is the confirmation speech of the target object, the subsequent business process continues to be executed.

[0136] Please refer to Figure 9 , Figure 9 It is a structural diagram of the synthetic speech detection system provided in an embodiment of the present application. The synthetic speech detection system can implement the above-mentioned synthetic speech detection method. The system includes: a sample audio acquisition module 910, a model acquisition module 920, a feature extraction module 930, a feature transformation module 940, a prediction and classification module 950, a parameter adjustment module 960 and a speech detection module 970.

[0137] A sample audio acquisition module 910 is used to acquire sample speech and sample audio tags of the sample speech;

[0138] The model acquisition module 920 is used to input the sample speech and sample audio labels into a preset initial network model; wherein the initial network model includes a convolutional layer, a residual network layer and a classification layer;

[0139] A feature extraction module 930 is used to extract features from the sample speech according to the convolution layer to obtain sample audio features;

[0140] A feature transformation module 940 is configured to perform feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features;

[0141] Prediction and classification module 950, configured to perform prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels;

[0142] A parameter adjustment module 960 is used to adjust the parameters of the initial network model according to the sample audio label and the predicted audio label to obtain a speech detection model;

[0143] The speech detection module 970 is used to input the target audio to be detected into the speech detection model for speech detection to obtain a speech category. The speech category is used to characterize whether the target audio is synthesized speech or non-synthesized speech.

[0144] The synthetic speech detection system of the embodiment of the present application is used to execute the synthetic speech detection method in the above embodiment. Its specific processing process is the same as the synthetic speech detection method in the above embodiment, and will not be repeated here.

[0145] An embodiment of the present application further provides a computer device, which includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor is used to execute the synthetic speech detection method in the above embodiment.

[0146] The following combination Figure 10 The hardware structure of the computer device is described in detail. The computer device includes: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040 and a bus 1050.

[0147] The processor 1010 can be implemented as a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0148] The memory 1020 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1020 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1020 and is called by the processor 1010 to execute the synthetic speech detection method of the embodiments of this application.

[0149] Input / output interface 1030, used to implement information input and output;

[0150] Communication interface 1040, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0151] bus 1050 , which transmits information between various components of the device (e.g., processor 1010 , memory 1020 , input / output interface 1030 , and communication interface 1040 );

[0152] The processor 1010 , the memory 1020 , the input / output interface 1030 , and the communication interface 1040 are connected to each other in communication within the device via a bus 1050 .

[0153] An embodiment of the present application further provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by a computer, the computer is used to execute the synthetic speech detection method of the embodiment of the present application.

[0154] The embodiments of the present application provide a method and system for detecting synthetic speech, a computer device, and a storage medium. In order to train a speech detection model, first, a sample speech and a sample audio label of the sample speech are obtained, and the sample speech and the sample audio label are input into a preset initial network model, wherein the initial network model includes a convolution layer, a residual network layer, and a classification layer. The initial network model can optimize the network structure to solve the problem of gradient disappearance during sample training and extract richer feature information. Feature extraction is performed on the sample speech according to the convolution layer to obtain the sample initial feature, and the sample initial feature is pooled to obtain the sample audio feature. Since the Ghost module can obtain richer feature information without changing the number of output channels, the sample audio feature is linearly transformed according to the Ghost module to obtain the first processing feature, and the sample audio feature is convolutionally transformed according to the first fully connected module to obtain the second processing feature, and the first processing feature and the second processing feature are residually calculated to obtain the sample transformation feature. To compress features to simplify the complexity of the model network and improve model training efficiency, residual calculation is performed on the first and second processed features to obtain sample residual features, which are then pooled to obtain the sample transformation features. The sample transformation features are then predicted and classified using a classification layer to obtain predicted classification data, wherein the predicted classification data includes predicted audio labels. Subsequently, the parameters of the initial network model are adjusted based on the sample audio labels and the predicted audio labels to obtain a speech detection model that can improve the detection accuracy of synthesized speech. In an application, the target audio to be detected is input into the speech detection model for speech detection to obtain a speech category, which is used to characterize whether the target audio is synthesized speech or non-synthesized speech. This embodiment of the present application constructs an initial network model comprising a convolutional layer, a residual network layer, and a classification layer. This allows the initial network model to avoid the influence of manual features on the prediction and classification process when predicting and classifying sample speech. The initial network model parameters are adjusted based on the sample audio labels and preset audio labels to obtain a speech detection model with the same structure as the initial network model but with higher detection accuracy. Among them, in the residual network layer, the embodiment of the present application realizes end-to-end synthetic speech detection by stacking Ghost-ResNet modules. The residual connection of ResNet can introduce forward feature information to solve the gradient disappearance problem in the propagation process. The Ghost module can obtain richer feature information without changing the number of output channels, and reduce the number of parameters and calculations of the model. In addition, the embodiment of the present application connects the Ghost module and the fully connected module by residual connection, which can help the model capture the discriminative features for identifying the authenticity of speech, thereby improving the detection ability and generalization ability of the model to accurately detect synthetic speech.It can be seen from this that when speech detection is performed on the target audio according to the speech detection model provided in the embodiment of the present application, speech detection can be performed on the target audio that does not contain manual features, thereby improving the detection accuracy of whether the target audio is synthetic speech.

[0155] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0156] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0157] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0159] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0160] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0161] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0162] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0163] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0164] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0165] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes multiple instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0166] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A method for detecting synthetic speech, characterized in that: The method comprises: Obtaining a sample speech and a sample audio tag of the sample speech; Inputting the sample speech and the sample audio tag into a preset initial network model; wherein the initial network model includes a convolutional layer, a residual network layer and a classification layer; Extracting features of the sample speech according to the convolutional layer to obtain sample audio features; According to the residual network layer, the sample audio features are subjected to feature transformation processing to obtain sample transformation features; the residual network layer includes a Ghost module and a first fully connected module, and the feature transformation processing of the sample audio features according to the residual network layer to obtain sample transformation features includes: performing linear transformation processing on the sample audio features according to the Ghost module to obtain a first processing feature; performing convolution transformation processing on the sample audio features according to the first fully connected module to obtain a second processing feature; performing residual calculation on the first processing feature and the second processing feature to obtain the sample transformation feature; wherein the Gho The st module includes a standard convolution submodule, an expanded group convolution submodule and a correction submodule. The linear transformation processing is performed on the sample audio feature according to the Ghost module to obtain the first processing feature, including: performing standard convolution processing on the sample audio feature according to the standard convolution submodule to obtain the first convolution feature; performing expanded group convolution processing on the first convolution feature according to the expanded group convolution submodule to obtain the second convolution feature; performing feature splicing on the first convolution feature and the second convolution feature to obtain the third convolution feature; performing correction processing on the third convolution feature according to the correction submodule to obtain the first processing feature; Performing prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels; Adjusting parameters of the initial network model according to the sample audio label and the predicted audio label to obtain a speech detection model; The target audio to be detected is input into the speech detection model for speech detection to obtain a speech category, where the speech category is used to characterize whether the target audio is synthesized speech or non-synthesized speech.

2. The method according to claim 1, characterized in that The correcting the third convolution feature according to the correction submodule to obtain the first processed feature includes: Performing batch normalization on the third convolution feature according to the correction submodule to obtain a fourth convolution feature; Feature calculation is performed on the fourth convolution feature according to a preset activation function to obtain a first processed feature.

3. The method according to claim 1, characterized in that The performing residual calculation on the first processing feature and the second processing feature to obtain a sample transformation feature includes: Performing residual calculation on the first processing feature and the second processing feature to obtain a sample residual feature; Pooling is performed on the sample residual features to obtain the sample transformation features.

4. The method according to claim 1, wherein The classification layer includes a second fully connected module, and the prediction classification processing of the sample transformation features according to the classification layer to obtain predicted classification data includes: Performing feature conversion processing on the sample transformation feature according to the second fully connected module to obtain a sample transformation feature; Normalization processing is performed on the sample conversion features to obtain predicted classification data, and the predicted classification data is used to count the predicted classification values ​​of different sample audio labels to which the sample speech belongs.

5. The method according to any one of claims 1 to 4, characterized in that The extracting features of the sample speech according to the convolutional layer to obtain sample audio features includes: Extracting features of the sample speech according to the convolutional layer to obtain initial features of the sample; Pooling is performed on the initial features of the sample to obtain the sample audio features.

6. A synthetic speech detection system, characterized in that: The system comprises: A sample audio acquisition module is used to acquire sample speech and sample audio tags of the sample speech; A model acquisition module, configured to input the sample speech and the sample audio label into a preset initial network model; wherein the initial network model includes a convolutional layer, a residual network layer, and a classification layer; A feature extraction module, configured to extract features from the sample speech according to the convolutional layer to obtain sample audio features; A feature transformation module is used to perform feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features; the residual network layer includes a Ghost module and a first fully connected module, and the feature transformation processing on the sample audio features according to the residual network layer to obtain sample transformation features includes: performing linear transformation processing on the sample audio features according to the Ghost module to obtain a first processing feature; performing convolution transformation processing on the sample audio features according to the first fully connected module to obtain a second processing feature; performing residual calculation on the first processing feature and the second processing feature to obtain the sample transformation feature; wherein, The Ghost module includes a standard convolution submodule, an expanded group convolution submodule and a correction submodule. The linear transformation processing of the sample audio feature according to the Ghost module to obtain the first processing feature includes: performing standard convolution processing on the sample audio feature according to the standard convolution submodule to obtain the first convolution feature; performing expanded group convolution processing on the first convolution feature according to the expanded group convolution submodule to obtain the second convolution feature; performing feature splicing on the first convolution feature and the second convolution feature to obtain a third convolution feature; performing correction processing on the third convolution feature according to the correction submodule to obtain the first processing feature; A prediction and classification module, configured to perform prediction and classification processing on the sample transformation features according to the classification layer to obtain predicted classification data; wherein the predicted classification data includes predicted audio labels; A parameter adjustment module, configured to adjust parameters of the initial network model according to the sample audio label and the predicted audio label to obtain a speech detection model; The speech detection module is used to input the target audio to be detected into the speech detection model for speech detection to obtain a speech category, where the speech category is used to characterize whether the target audio is synthesized speech or non-synthesized speech.

7. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is configured to perform: The method according to any one of claims 1 to 5.

8. A storage medium, wherein the storage medium is a computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program. When the computer program is executed by a computer, the computer is configured to: The method according to any one of claims 1 to 5.

Images