Universal audio tampering detection method and system, storage medium and electronic equipment

Abstract

The invention provides a general audio tampering detection method and system, a storage medium and electronic equipment, and relates to the technical field of audio recognition. According to the general audio tampering detection method, a pre-constructed integrated detection framework based on a deep neural network is adopted, the integrated detection framework comprises a detection network, and the detection network is used for processing a Mel spectrogram of an audio to be detected and outputting a positioning mask consistent with the Mel spectrogram in size; the classification network outputs a tampering category label for tampering the audio mel spectrogram; wherein the detection network comprises an encoder and a decoder. The dense convolution block attention module and the self-attention module are introduced into the encoder of the detection network, gradient disappearance and information redundancy are effectively relieved, local and global features are effectively fused, the perception ability of the detection network to tiny tampering traces is enhanced, and the sensitivity and generalization ability of the detection network to tiny tampering are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of audio recognition technology, and more specifically to a general audio tampering detection method, system, storage medium, and electronic device. Background Technology

[0002] In the field of voice tampering detection, existing methods can be mainly divided into detection methods based on traditional signal processing and audio tampering detection methods based on deep learning.

[0003] Traditional signal processing-based detection methods typically rely on acoustic features (such as fundamental frequency, formants, and ENF) for tampering trace analysis, combined with Support Vector Machines (SVM) or Gaussian Mixture Models (GMM) for classification. Deep learning-based audio tampering detection methods mainly include using basic Convolutional Neural Networks (CNNs) to automatically learn forgery patterns, introducing deep residual networks into audio deception detection, or combining the self-supervised learning model Wav2Vec 2.0 with data augmentation detection methods.

[0004] However, traditional signal processing-based tampering detection methods are limited by the strong binding between acoustic features and specific tampering methods, failing to cover diverse and novel tampering scenarios. Furthermore, SVM and GMM further amplify these limitations, relying on manually designed acoustic features for training, resulting in inherently weak generalization capabilities and an inability to identify new acoustic change patterns when faced with unseen tampering types. Most deep learning-based audio tampering detection models tend to rely on acoustic traces left by specific forgery algorithms during training, leading to significant performance degradation when encountering novel or complex tampering methods. In short, existing methods, whether based on traditional signal processing or deep learning, can only detect specific types of tampering, exhibiting poor generalization capabilities. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a universal audio tampering detection method, system, storage medium, and electronic device, solving the technical problem of poor generalization ability of current methods.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a general audio tampering detection method, which utilizes a pre-constructed integrated detection framework based on a deep neural network. The integrated detection framework includes a detection network that processes the Mel spectrogram of the audio to be detected and outputs a localization mask with the same size as the Mel spectrogram. A classification network outputs a tampering category label for the tampered audio Mel spectrogram. The detection network includes an encoder and a decoder. The encoder includes a cascaded channel-level high-pass filter module, a dense convolutional block attention module, and a self-attention module; The channel-level high-pass filtering module splits the Mel spectrum into three channels: H, W, and C. It applies three learnable high-pass convolution kernels to each channel independently, and after highlighting edges and gray-level abrupt changes, it splices the results to obtain the channel-level high-pass filtered features. The dense convolutional block attention module consists of a first dense block, a dual-branch attention unit, and a second dense block. The first dense block performs multi-scale feature extraction and reuse on the channel-level high-pass filtered features, outputting fine-grained detail features. The dual-branch attention unit performs channel attention and spatial attention processing on the fine-grained detail features, enhancing useful features and suppressing irrelevant noise, outputting filtered features enhanced with key features. The second dense block integrates the filtered features, outputting refined features that combine fine-grained and high-level semantics. The self-attention module includes a spatial attention submodule and a channel attention submodule. The spatial attention submodule generates position-sensitive weights by calculating the similarity between positions in the refined features, thereby amplifying the features of the tampered region and generating spatial attention features. The channel attention submodule strengthens the task-related semantic information in the refined features and generates channel attention features. The spatial attention features and channel attention features are fused and processed through a convolutional layer to output deep features containing global semantics. The decoder gradually restores the feature resolution by upsampling and feature concatenation, concatenating deep features with refined features and then with fine-grained features, and finally outputting a positioning mask.

[0007] Preferably, the integrated detection framework further includes a classification network, which is used to process the Mel spectrogram of the audio to be detected and output a classification result.

[0008] Preferably, the dual-branch attention includes channel branching and spatial branching; In the channel branch, fine-grained detail features are processed through parallel max pooling and average pooling layers. The processed features are then passed through a shared MLP layer. The output of the shared MLP layer is assigned different weight vectors. The features with different weights are fused and then passed through an activation layer to output channel features. After multiplying the channel features with the fine-grained detail features, the data is input into the spatial branch. The spatial branch performs max pooling, average pooling, convolution, and activation on the input data. The output features are then multiplied with the channel features to obtain the filtered features.

[0009] Preferably, the decoder includes a first module and a second module; The first module is used to perform upsampling, convolution, and VGG processing on the first spliced ​​feature and output the first feature. The first spliced ​​feature and the fine-grained detail feature are spliced ​​together and then input into the second module for sampling, convolution, and activation processing, and output a localization mask. The first spliced ​​feature refers to the feature after splicing the deep feature and the refined feature.

[0010] Preferably, the loss function during the training process of the detection network includes: in, The total number of samples, Indicates the first The true label of the frame, To detect the probability that the frame is tampered with.

[0011] Preferably, the classification network includes a feature extraction module and a feature learning and output module, wherein, The feature extraction module is used to extract spectral features from the Mel spectrogram step by step through a multi-level convolution-max pooling alternating structure; The feature learning and output module is used to process spectral features and output a classification of tampering types.

[0012] Preferably, the loss function during the training process of the classification network includes: in, Indicates that the sample belongs to the first The true label of the class, For the network to the first Predicted probability of class This represents the total number of categories.

[0013] Secondly, the present invention provides a universal audio tampering detection system, wherein the audio deepfake identification system is used to perform the universal audio tampering detection method as described above.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for general audio tampering detection, wherein the computer program causes a computer to perform the general audio tampering detection method as described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: One or more processors; Memory; and One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the general audio tampering detection method as described above.

[0016] (III) Beneficial Effects This invention provides a universal method, system, storage medium, and electronic device for detecting audio tampering. Compared with existing technologies, it has the following advantages: This invention introduces a dense convolutional block attention module and a self-attention module into the encoder of the detection network, which effectively alleviates gradient vanishing and information redundancy, and effectively integrates local and global features, enhances the detection network's ability to perceive subtle tampering traces, and significantly improves the detection network's sensitivity to and generalization ability to minor tampering. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the detection network structure in the integrated detection framework of this invention. Figure 2 This is a schematic diagram of the structure of the dual-branch attention unit in the detection network in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the self-attention module in the detection network in an embodiment of the present invention; Figure 4 This is a schematic diagram of the classification network structure in the integrated detection framework of this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides a general audio tampering detection method, system, storage medium, and electronic device, which solves the technical problem of poor generalization ability of current methods, improves the ability to identify multiple types of tampering, and enhances cross-type generalization performance.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: This invention addresses the field of voice content security, specifically the risks of tampering arising from the large-scale collection of personal voice data, including copying-moving, deletion, same-source splicing, and heterogeneous-source splicing. It provides an integrated detection framework based on deep neural networks. This framework integrates "blind detection, precise localization, and detailed classification," and can be directly deployed in applications such as forensics and media authentication. It can quickly identify the authenticity of unknown recordings, pinpoint tampered segments, and characterize the tampering methods.

[0022] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0023] This invention provides a universal audio tampering detection method. This method utilizes a pre-constructed ensemble detection framework based on a deep neural network. The ensemble framework includes a detection network that processes the Mel spectrogram of the audio to be detected and outputs a localization mask with the same size as the Mel spectrogram. Figure 1 As shown. The detection network includes an encoder and a decoder; The encoder includes a cascaded channel-level high-pass filter module, a dense convolutional block attention module, and a self-attention module; The channel-level high-pass filtering module splits the Mel spectrum into three channels: H, W, and C. It applies three learnable high-pass convolution kernels to each channel independently, and after highlighting edges and gray-level abrupt changes, it splices the results to obtain the channel-level high-pass filtered features. The dense convolutional block attention module consists of a first dense block, a dual-branch attention unit, and a second dense block. The first dense block performs multi-scale feature extraction and reuse on the channel-level high-pass filtered features, outputting fine-grained detail features. The dual-branch attention unit performs channel attention and spatial attention processing on the fine-grained detail features, enhancing useful features and suppressing irrelevant noise, outputting filtered features enhanced with key features. The second dense block integrates the filtered features, outputting refined features that combine fine-grained and high-level semantics. The self-attention module includes a spatial attention submodule and a channel attention submodule. The spatial attention submodule generates position-sensitive weights by calculating the similarity between positions in the refined features, thereby amplifying the features of the tampered region and generating spatial attention features. The channel attention submodule strengthens the task-related semantic information in the refined features and generates channel attention features. The spatial attention features and channel attention features are fused and processed through a convolutional layer to output deep features containing global semantics. The decoder gradually restores the feature resolution by upsampling and feature concatenation, concatenating deep features with refined features and then with fine-grained features, and finally outputting a positioning mask.

[0024] It should be noted that, in the specific implementation process, the integrated detection framework also includes a classification network, which is used to process the Mel spectrogram of the audio to be detected and output the classification result.

[0025] The following is a detailed description of each component in the integrated detection framework: The encoder comprises a cascaded channel-level high-pass filter (CW-HPF) module, a dense convolutional block attention module (Dense-CBAM), and a self-attention module (SAM). CW-HPF is responsible for extracting high-frequency residuals and accurately capturing local edges and details; Dense-CBAM preserves multi-layer semantics while eliminating redundancy through dual channel-space attention, outputting robust coarse features; SAM further amplifies the distributional differences between the tampered region and the real region, maximizing the expression of task-related information.

[0026] In a Mel spectrum, pixel brightness directly reflects energy intensity. Alteration of boundaries is often accompanied by weak abrupt changes in high-frequency components, manifesting as localized grayscale jumps. Therefore, as... Figure 1 As shown, CW-HPF splits the input Mel spectrum (H×W×C) by channel, applies three learnable high-pass convolution kernels independently to each channel, highlights edges and gray-level abrupt changes, and then splices the results to obtain H×W×3C features with enhanced details.

[0027] Subsequently, the enhanced H×W×3C features are fed into the Dense-CBAM module, which includes Block-1, CBAM, and Block-2.

[0028] This feature first fully reuses low-level details through Block-1 (4 Dense layers + 0.5 compression transition layer); then through... Figure 2The CBAM described uses a dual-branch attention mechanism (channel-space branch: parallel average / max pooling + shared MLP; spatial branch: channel-level pooling + 7×7 convolution) to suppress irrelevant responses. Finally, a Block-2 layer (two dense layers + a transition layer) integrates the context, forming refined features with both fine-grained and high-level semantics. The entire encoding pathway deepens the network's receptive field while effectively mitigating gradient vanishing and information redundancy through dense connections, attention gating, and residual fusion, significantly improving the network's sensitivity to minor tampering and its generalization ability. The shared MLP output is assigned different weight vectors, which are not fixed values ​​but dynamic parameters adaptively learned from the training data by the shared MLP layer. The specific selection rules are automatically optimized by the model through backpropagation to capture the differences in importance between different channels in the input features.

[0029] To more effectively highlight the differences between global and local features and to make the network focus more on the tampered region, the encoder introduces a self-attention module (SAM). The SAM includes a pre-feature processing sub-stage and a self-control mechanism. The pre-feature processing sub-stage pre-processes the features input to the SAM using VGG blocks and multiple dilated convolutional layers. For example... Figure 3 As shown, its spatial attention submodule generates location-sensitive weights by calculating the similarity between locations, thereby amplifying the features of the tampered region: in, For input features; , , This is the reshaped feature map; The learnable coefficient; This refers to the enhanced spatial attention features.

[0030] Meanwhile, the channel attention submodule enhances task-related semantic information and adaptively learns the most suitable channel combination for tamper detection in the following ways: in, To reshape features, It is also a learnable coefficient. This is the obtained channel attention feature. The two attention mechanisms work together to enable the inspection network to accurately focus on key clues in the tampered area while maintaining the global context.

[0031] The encoder network decoder employs a "deep and shallow dual-path fusion" strategy to progressively refine features, restoring the abstract semantics of the encoder output to a pixel-level tampering mask. First, the refined features (Block-2 output) are concatenated across layers with the output of the self-attention module (containing deep features with global semantics), and then fused with fine-grained detail features (Block-1 output). Next, the resolution is progressively amplified through two "3×3 convolution + bilinear upsampling" units, with a lightweight VGG module embedded after each upsampling to suppress checkerboard noise and restore fine structures. Subsequently, the local details from Block-1 are fused again with the first upsampling result to further supplement spatial information. Finally, a 1×1 convolution compresses the channels and activates with Sigmoid, outputting a localization mask (PLM) with the same size as the original image. This complementary fusion of deep and shallow features enhances the fine-grained characterization of tampering boundaries while ensuring semantic consistency, thus significantly improving the robustness and detection accuracy of the detection network.

[0032] The classification network draws on established image classification concepts, primarily employing a pure convolutional structure. It utilizes convolutional, pooling, and fully connected layers to collaboratively achieve end-to-end acoustic pattern recognition. Its overall process is as follows: Figure 4 As shown, it includes a feature extraction module and a feature learning and output module.

[0033] Feature extraction module: Using Mel spectrograms as input, it extracts more refined spectral features step by step through a multi-level convolution-max pooling alternating structure. Each convolutional layer is followed by ReLU activation, with negative values ​​set to zero and positive values ​​retained, introducing nonlinearity while maintaining computational efficiency. As the network deepens, the number of convolutional kernels gradually increases, enabling the classification network to capture increasingly abstract and complex acoustic cues, thereby achieving a comprehensive representation of potential tampering traces.

[0034] Feature learning and output module: Extracted high-order features are expanded into one-dimensional vectors by a Flatten layer and then fed into two consecutive fully connected layers for nonlinear transformation and information integration. The number of neurons decreases layer by layer, progressively refining the feature representation. The final output layer contains 5 neurons, corresponding to five categories: "unmodified / copy-paste / deleted / homogeneous concatenation / heterogeneous concatenation". Softmax is used to convert the output into a probability distribution, and the category with the highest probability is used as the final prediction result. This design achieves accurate discrimination of multiple tampering types while maintaining lightweight design, providing reliable labels for subsequent intent analysis.

[0035] It should be noted that, in the actual implementation process, the detection network and the classifier network are trained and optimized separately.

[0036] During the training phase of the detection network, the network parameters are first randomly initialized. Then, the Mel Spectrum Image (MSF) and its corresponding Ground Truth Metric (GTM) are simultaneously input into the detection network to obtain the predicted localization mask (PLM). The parameters are optimized by calculating the binary cross-entropy (BCE) loss between the predicted PLM and GTM. in, The total number of samples, Indicates the first The true label of the frame (0: not tampered, 1: tampered). The detection network predicts the probability that a frame belongs to the tampered category. Through multiple iterations, the detection network gradually learns the optimal mapping relationship between the MSF and PLM, thus ensuring the final detection network has excellent performance. It should be noted that the detection network ultimately outputs a localization mask through the Sigmoid activation function. The output value of each pixel in the mask represents the probability that the location belongs to the tampered area. The essence of the localization mask is the pixel-level tampering probability distribution, and the visualized black and white effect is based on the probability judgment result.

[0037] The classification network is also trained using Mel-spectral Facility (MSF) as input, loaded with the corresponding true class labels (integer indices). The optimization objective is the Sparse Categorical Cross-Entropy (SCCE) loss: in, Indicates that the sample belongs to the first The true label of the class, For the network to the first Predicted probability of class The total number of categories is 5 in this paper. After multiple rounds of iterative updates, the classification network gradually converges and fully adapts to the classification task, ultimately achieving the optimal average classification accuracy on the validation set.

[0038] The effectiveness of the embodiments of the present invention will be verified through the following verification experiments: Verify test settings: Training data: The THCHS-30 and AISHELL datasets were mainly used as the original Chinese audio datasets; LibriSpeech and Common Voice were used as the original English audio datasets to create various types of tampered samples for network training.

[0039] Test data: The THCHS-30 and AISHELL datasets were mainly used as the original Chinese audio datasets; LibriSpeech and Common Voice were used as the original English audio datasets to create various types of tampered samples for model testing.

[0040] Evaluation metrics: Accuracy, precision, recall, and F1 score were used as the main evaluation metrics. Additionally, the average localization error was used to evaluate the model's localization error.

[0041] Experimental environment: The model was trained on an NVIDIA GeForce RTX 2070 GPU using the Adam optimizer with a learning rate of [missing information]. .

[0042] Experimental results: This method achieves the best performance compared to several related methods, and even outperforms methods specializing in single-item tamper detection, demonstrating its strong detection performance. The experimental results of the detection network are shown in Table 1, where DE represents deletion, CM represents copy-paste, Hos represents homologous splicing, Hes represents heterologous splicing, and UN represents untampered samples.

[0043] Table 1. Results of network testing The LE metrics of the detection network are shown in Table 2, which compares the performance of several methods with localization capabilities. It can be seen that the localization error of this method is within 0.1 seconds, which is better than the comparison methods.

[0044] Table 2. LE values ​​(average, in seconds) for different methods The experimental results of the classification network are shown in Table 3. Due to the lack of relevant research, only the metrics of this method are shown. The metrics also show that the classification network has achieved high accuracy on both Chinese and English datasets.

[0045] Table 3. Experimental Results of Classification Network This invention also provides a general audio tampering detection system, which is used to perform the general audio tampering detection method as described above.

[0046] This invention also provides a computer-readable storage medium storing a computer program for general audio tampering detection, wherein the computer program causes a computer to execute the general audio tampering detection method as described above.

[0047] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the general audio tampering detection method as described above.

[0048] In summary, compared with existing technologies, it has the following beneficial effects: 1. In this embodiment of the invention, a dense convolutional block attention module and a self-attention module are introduced into the decoder of the detection network. This effectively alleviates gradient vanishing, information redundancy, and effectively integrates local and global features, enhancing the detection network's ability to perceive subtle tampering traces and significantly improving the detection network's sensitivity and generalization ability to minor tampering.

[0049] 2. The detection, localization and classification tasks are integrated into a unified deep learning framework, avoiding the complexity of traditional methods that require designing separate models for each type of tampering.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A general method for detecting audio tampering, characterized in that, The general audio tampering detection method utilizes a pre-constructed integrated detection framework based on deep neural networks. This framework includes a detection network that processes the Mel spectrogram of the audio to be detected and outputs a localization mask with the same size as the Mel spectrogram. A classification network outputs a tampering category label for the tampered audio's Mel spectrogram. The detection network comprises an encoder and a decoder. The encoder includes a cascaded channel-level high-pass filter module, a dense convolutional block attention module, and a self-attention module; The channel-level high-pass filtering module splits the Mel spectrum into three channels: H, W, and C. It applies three learnable high-pass convolution kernels to each channel independently, and after highlighting edges and gray-level abrupt changes, it splices the results to obtain the channel-level high-pass filtered features. The dense convolutional block attention module consists of a first dense block, a dual-branch attention unit, and a second dense block. The first dense block performs multi-scale feature extraction and reuse on the channel-level high-pass filtered features, outputting fine-grained detail features. The dual-branch attention unit performs channel attention and spatial attention processing on the fine-grained detail features, enhancing useful features and suppressing irrelevant noise, outputting filtered features enhanced with key features. The second dense block integrates the filtered features, outputting refined features that combine fine-grained and high-level semantics. The self-attention module includes a spatial attention submodule and a channel attention submodule. The spatial attention submodule generates position-sensitive weights by calculating the similarity between positions in the refined features, thereby amplifying the features of the tampered region and generating spatial attention features. The channel attention submodule strengthens the task-related semantic information in the refined features and generates channel attention features. The spatial attention features and channel attention features are fused and processed through a convolutional layer to output deep features containing global semantics. The decoder gradually restores the feature resolution by upsampling and feature concatenation, concatenating deep features with refined features and then with fine-grained features, and finally outputting a positioning mask.

2. The general audio tampering detection method as described in claim 1, characterized in that, The integrated detection framework also includes a classification network, which processes the Mel spectrogram of the audio to be detected and outputs a classification result.

3. The general audio tampering detection method as described in claim 1, characterized in that, The dual-branch attention includes channel branching and spatial branching; In the channel branch, fine-grained detail features are processed through parallel max pooling and average pooling layers. The processed features are then passed through a shared MLP layer. The output of the shared MLP layer is assigned different weight vectors. The features with different weights are fused and then passed through an activation layer to output channel features. After multiplying the channel features with the fine-grained detail features, the data is input into the spatial branch. The spatial branch performs max pooling, average pooling, convolution, and activation on the input data. The output features are then multiplied with the channel features to obtain the filtered features.

4. The general audio tampering detection method as described in any one of claims 1 to 3, characterized in that, The decoder includes a first module and a second module; The first module is used to perform upsampling, convolution, and VGG processing on the first spliced ​​feature and output the first feature. The first spliced ​​feature and the fine-grained detail feature are spliced ​​together and then input into the second module for sampling, convolution, and activation processing, and output a localization mask. The first spliced ​​feature refers to the feature after splicing the deep feature and the refined feature.

5. The general audio tampering detection method as described in any one of claims 1 to 3, characterized in that, The loss function during the training process of the detection network includes: in, The total number of samples, Indicates the first The true label of the frame, To detect the probability that the frame is tampered with.

6. The general audio tampering detection method as described in claim 2, characterized in that, The classification network includes a feature extraction module and a feature learning and output module, wherein, The feature extraction module is used to extract spectral features from the Mel spectrogram step by step through a multi-level convolution-max pooling alternating structure; The feature learning and output module is used to process spectral features and output a classification of tampering types.

7. The general audio tampering detection method as described in claim 2, characterized in that, The loss function during the training process of the classification network includes: in, Indicates that the sample belongs to the first The true label of the class, For the network to the first Predicted probability of class This represents the total number of categories.

8. A universal audio tampering detection system, characterized in that, The audio deepfake detection system is used to perform the general audio tampering detection method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program for general audio tampering detection, wherein the computer program causes the computer to perform the general audio tampering detection method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the general audio tampering detection method as described in any one of claims 1 to 7.

Images