Counterfeit voice detection method adopting two-dimensional map attention

Abstract

The invention relates to a forged voice detection method adopting two-dimensional map attention, and the method comprises the steps: carrying out the pre-emphasis processing of an original voice, and extracting the linear frequency cepstrum coefficient features of each frame; gMM modeling is carried out on real voice and forged voice by using a BP algorithm according to linear frequency cepstrum coefficient features, iterative operation is carried out on training data by using an expectation maximization algorithm until convergence, the number of voice frames is counted, the LGP of each frame of voice feature in each Gaussian component is calculated, a constant term is subtracted, and standardization processing is carried out; according to the method, the Gaussian mixture model is modeled by adopting a bidirectional subdivision method (BP), so that the difference of mean vectors of adjacent Gaussian components is relatively small, the dependency relationship between the Gaussian components can be effectively captured by two-dimensional convolution, a clearer and tighter feature relationship is provided for a graph attention module, and the image attention is more accurate. According to the method, the feature space structure is optimized from local to global, and the performance of forged voice detection is improved.

Description

Technical Field

[0001] The invention relates to a forged speech detection method using dual-dimensional graph attention, and belongs to the technical field of forged speech detection. Background Art

[0002] Fake speech detection technology is used to determine the authenticity of the speech to be tested to prevent malicious attacks. It has important application value in fields such as Automatic Speaker Verification (ASV) and forensic examination. Existing methods extract the log-Gaussian probability (LGP) of fake speech on the Gaussian mixture model (GMM) components as features for detection. However, such methods only perform one-dimensional convolution in the time dimension and fail to fully explore the relationship between Gaussian components.

[0003] The present invention uses bidirectional partitioning (BP) to model GMM so that the difference in mean vectors of adjacent Gaussian components is small, captures the dependency between Gaussian components through two-dimensional convolution processing, and introduces a graph attention module to aggregate information of time dimension and Gaussian component dimension, thereby improving the accuracy of forged speech detection. Summary of the invention

[0004] In order to overcome the deficiency of existing forged speech detection methods that only one-dimensional convolution is performed on the time dimension and the relationship between Gaussian components cannot be fully explored, the present invention provides a forged speech detection method using two-dimensional graph attention, which can more deeply capture the correlation information between the Gaussian component dimension and the time dimension, so that the model can more comprehensively analyze the structural characteristics of the input data, thereby improving the model's decision ability.

[0005] A method for detecting forged speech using dual-dimensional graph attention, comprising the following steps:

[0006] Step 1: First, pre-emphasize the original speech to enhance the details of the high-frequency waveform, then perform windowing and framing, and extract the Linear Frequency Cepstrum Coefficients (LFCC) features of each frame;

[0007] Step 2: Based on the LFCC features, the BP algorithm is used to perform GMM modeling on the real speech and the fake speech respectively. The expectation maximization (EM) algorithm is used to iteratively operate the training data until convergence. The number of speech frames is counted and the LGP of each frame of speech features in each Gaussian component is calculated respectively. Then, the constant term is subtracted and normalization is performed.

[0008] Step 3: Input the LGP feature matrix into six residual block modules. Each residual block consists of two 2D convolutional layers, a SeLU activation function, and a batch normalization layer. A 2×2 maximum pooling layer is added at the end of the first four residual blocks to reduce complexity and improve generalization ability.

[0009] Step 4: Input the deep features generated by the residual block into the temporal attention module and the Gaussian component attention module respectively to obtain graph nodes of the same dimension but different numbers. Both attention modules contain a graph attention network (GAT) and a graph pooling module (GPM). Then, concatenate the graph nodes obtained by the two networks and input them into the Gaussian component-temporal joint graph attention module to obtain the final graph nodes.

[0010] Step 5: The input features LGP of the fake and real speech are processed in parallel, the nodes of both sides are flattened and concatenated as the input of the fully connected layer, the confidence of the two categories of true / false is generated, the confidence is converted into probability through the softmax function, and the weighted cross entropy loss function (Weighted Cross Entropy, WCE) is used to calculate the error between the actual label and the predicted label, and the loss is minimized to optimize the classification performance of the model and improve the accuracy of the judgment of the authenticity of the speech.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] After the deep features of the Gaussian component and the time dimension are processed by independent graph attention modules, the present invention fuses the features of the two through a splicing operation, and further performs joint feature extraction through a Gaussian component-time joint graph attention module. This joint feature extraction method can capture the correlation information between the Gaussian component dimension and the time dimension more deeply, enabling the model to more comprehensively analyze the structural characteristics of the input data, thereby improving the model's judgment ability.

[0013] The present invention uses the bidirectional segmentation method (BP) to model the Gaussian mixture model, so that the mean vector difference of adjacent Gaussian components is small. This strategy not only enables the two-dimensional convolution to more effectively capture the dependency between Gaussian components, but also provides a clearer and closer feature relationship for the graph attention module, greatly enhancing the attention mechanism's ability to characterize global features. This method optimizes the feature space structure from local to global, greatly improving the performance of forged speech detection.

[0014] The present invention establishes Gaussian mixture models (GMM) of real speech and fake speech respectively, extracts their features, and inputs them into models with the same structure but different parameters in parallel. The features of real and fake speech are fused in the fully connected layer, making full use of the relationship between the two, and further improving the accuracy of the model in distinguishing real / fake speech. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 The flowchart and principle block diagram of the forged speech detection method using dual-dimensional graph attention of the present invention;

[0017] Figure 2 A schematic diagram of the GMM model generation and training principle of the forged speech detection method using dual-dimensional graph attention of the present invention;

[0018] Figure 3 This is a structural diagram of a residual block module of a forged speech detection method using dual-dimensional graph attention in the present invention;

[0019] Figure 4 This is the process and principle block diagram of the time attention module and Gaussian component attention module of the forged speech detection method using dual-dimensional graph attention in the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Reference Figure 1, a forged speech detection method using dual-dimensional graph attention, comprising the following steps:

[0022] Step 1: First, pre-emphasize the original speech to enhance the details of the high-frequency waveform, then perform windowing and framing, and extract the Linear Frequency Cepstrum Coefficients (LFCC) features of each frame;

[0023] Step 1 specifically includes:

[0024] The input speech signal is first pre-emphasized, and a first-order high-pass filter is used to enhance the high-frequency components and compensate for the natural attenuation of high-frequency energy. Subsequently, the signal is divided into frames of 20-30 milliseconds in length, and the frame shift is set to 50% of the frame length to achieve overlapping of adjacent frames and retain the correlation between signal frames. Each frame of the speech signal is then windowed (such as a Hamming window) to reduce edge discontinuities.

[0025] Perform a Fast Fourier Transform (FFT) on each frame of speech to convert the time domain signal into a frequency domain signal and obtain the spectrum information. Then, use a set of triangular filters with linear frequency intervals to filter the spectrum. Each filter covers a specific frequency band, and the energy of the filter output is calculated.

[0026] The natural logarithm of the energy output by the filter is taken to make the data more consistent with the human ear's perception of loudness, and the multiplication operation is converted into an addition operation. Next, the obtained logarithmic energy sequence is subjected to a discrete cosine transform (DCT) to remove the correlation between features, compress the data dimension, and obtain the linear frequency scale cepstral coefficient (LFCC) feature, where each coefficient represents the spectral feature of the speech signal.

[0027] In order to capture the dynamic changes of speech signals, the first-order and second-order differential coefficients (i.e., Δ and ΔΔ coefficients) are further calculated. These differential coefficients reflect the changing trend of features over time and improve the model's ability to model time series information.

[0028] Step 2: Based on the LFCC features, the BP algorithm is used to perform GMM modeling on the real speech and the fake speech respectively. The expectation maximization (EM) algorithm is used to iteratively operate the training data until convergence. The number of speech frames is counted and the LGP of each frame of speech features in each Gaussian component is calculated respectively. Then, the constant term is subtracted and normalization is performed.

[0029] Step 2 specifically includes:

[0030] In the present invention, the BP algorithm is combined with the GMM to train, and the Gaussian distribution of a single component is gradually evolved into a Gaussian model of multiple components. The method first uses the global mean and variance to initialize the GMM of a single Gaussian component, and then gradually increases the number of Gaussian components to adapt to the complexity of the data. The real and fake speech are modeled by GMM respectively, and λ is used to represent the parameter set composed of the mean vector, covariance matrix and mixing weight of each Gaussian component, that is:

[0031] λ={ω i ,μ i ,∑ i},i=1,2,…,512 (1)

[0032] Among them, ω i is the mixing weight and satisfies μ i and∑ i denote the mean vector and covariance matrix of the ith Gaussian component respectively. The model parameter set λ is obtained by the EM algorithm.

[0033] In each split of the BP algorithm, each Gaussian component is divided into two. This split is achieved by adding or subtracting a small perturbation to the mean vector of the original component. Specifically, given the mean vector μ and covariance matrix Σ of a Gaussian component, the mean vectors of the two newly generated Gaussian components are μ+ò and μ-ò, respectively, where ò is a small perturbation vector whose value is δ×eigenvector(Σ). δ is a preset constant, and eigenvector(Σ) represents the main eigenvector of the covariance matrix Σ, indicating the direction of the maximum variation of the data, and the formula is as follows:

[0034] μ new1 = μ+δ×eigenvector(Σ) (2)

[0035] μ new2 = μ-δ×eigenvector(Σ) (3)

[0036] μ new1 and μ new2is the mean vector of the two newly generated Gaussian components, the covariance matrix Σ remains unchanged, and the initial weights of the two new Gaussian components are set to half of the original component weights to ensure that the sum remains consistent. After each split, the new GMM model contains more components, and an appropriate number of EM iterations are performed to estimate the new GMM parameters to ensure that the model parameters can fully fit the data after the introduction of new components. When the last split is completed, the EM iteration process continues until the change in model parameters is lower than a certain threshold, or the preset number of iterations is reached, and GMM models that fit real speech and fake speech are obtained respectively. The principles of generating the GMM model and the training process are as follows: Figure 2 shown.

[0037] After the GMM models of the real speech and the fake speech are established, the LFCC features of the speech are input into the two GMM models respectively to obtain the LGP features of each frame of speech features in the GMM. Assume x t Represents the LFCC feature vector of the t-th frame speech signal, and finds x t After taking the logarithm of the probability density of each component, we can get:

[0038]

[0039] where d is the vector x t The dimension of x is , and T represents the transposition operator. In order to focus on the discriminative features, we retain the features that directly reflect x t The probability of the items under different components. Therefore, only the items related to x in equation (4) are retained. t The relevant items are x t The logarithmic probability of the i-th Gaussian component is:

[0040]

[0041] After calculating the LGP of all speech frames in each Gaussian component t,i Afterwards, the data is standardized along the Gaussian component dimension to a distribution with a mean of 0 and a variance of 1, thereby obtaining the LGP feature matrices under the real speech GMM and the fake speech GMM respectively.

[0042] Step 3: Input the LGP feature matrix into six residual block modules. Each residual block consists of two 2D convolutional layers, a SeLU activation function, and a batch normalization layer. A 2×2 maximum pooling layer is added at the end of the first four residual blocks to reduce complexity and improve generalization ability.

[0043] The network architecture proposed in the present invention is based on a deep residual learning framework and includes two stages of residual block processing. Before the LGP feature is input into the residual network, it is first expanded or truncated and unified to 400 frames in the time dimension. In the first stage, the network consists of four residual blocks, each of which contains two two-dimensional convolutional layers with a convolution kernel size of (3,3) and 32 output channels. Batch normalization and SeLU activation functions are set between convolutional layers to enhance the nonlinearity and stability of the model. A 2×2 maximum pooling layer is also added after each residual block to reduce the feature dimension and reduce the subsequent computational burden.

[0044] In the second stage, the network contains two residual blocks, the number of output channels of the convolutional layer is increased to 64, and no pooling layer is added after each block. After six residual blocks, batch normalization and SeLU activation function are applied again to further stabilize the feature distribution and enhance the nonlinear expression ability, providing more discriminative feature input for the subsequent graph attention module.

[0045] This design effectively improves the expressiveness of features by gradually deepening the network depth and increasing the number of output channels. At the same time, the pooling layer is used to reduce the number of parameters and the amount of calculation, thereby improving the generalization ability of the model. The entire network aims to achieve efficient processing of complex data by balancing depth and width. The residual block module structure is as follows Figure 3 shown.

[0046] Step 4: Input the deep features generated by the residual block into the temporal attention module and the Gaussian component attention module respectively to obtain graph nodes of the same dimension but different numbers. Both attention modules contain a graph attention network (GAT) and a graph pooling module (GPM). Then, concatenate the graph nodes obtained by the two networks and input them into the Gaussian component-temporal joint graph attention module to obtain the final graph nodes.

[0047] Step 4 specifically includes:

[0048] In the graph attention module, the dimension of the input feature Q is (64, 25, 32), where 64 represents the number of feature channels, 25 and 32 represent the number of time dimensions and the number of Gaussian component dimensions, respectively. First, the input data is processed by the Gaussian component attention module to capture the features on the Gaussian component dimension. A maximum pooling operation is performed along the time dimension to obtain the feature representation on the Gaussian component dimension, resulting in a feature map of size (64, 1, 32). Then, the feature map is dimensionally transformed to obtain a representation of dimension (64, 32), and input into the graph attention network (GAT) to generate attention-weighted graph features of dimension (32, 32). Subsequently, the graph features are pooled using the graph pooling module (GPM), and the most representative nodes are selected to obtain graph features of dimension (32, 18).

[0049] The structure of the temporal attention module remains unchanged, and it is mainly used to capture feature information in the temporal dimension. First, the input feature is subjected to the maximum pooling operation in the Gaussian component dimension to obtain a temporal feature representation with a dimension of (64, 25, 1). Next, the dimension is transformed to obtain a feature representation of (64, 25), which is then input into GAT to generate a graph feature with a dimension of (32, 25). Subsequently, GPM is used to pool the graph nodes and select key nodes to obtain a representation with a dimension of (32, 14), providing graph features in the temporal dimension for the subsequent joint attention module.

[0050] In the Gaussian component-time joint attention module, the graph features on the Gaussian component and time dimensions are concatenated as input and processed by this module to capture the joint information between the two. The concatenated feature dimension is (64,32), which is input into the joint graph attention network for further feature extraction, and the output dimension is (16,32). After that, the dimension is reduced to (16,16) through GPM. Next, the feature of size (16,16) is flattened into a 256-dimensional feature vector. The feature vectors of the real speech and the fake speech are concatenated into a 512-dimensional vector and input into the fully connected layer (FC layer) to generate confidence scores for the two categories of true / false. Figure 4 The flow and principle block diagram of the graph attention module.

[0051] Then, the two scores are mapped to the range of (0,1) through softmax, and the sum of the two values ​​is 1, which represents the probability of belonging to each category. Finally, the cross entropy loss function is used to calculate the error between the actual label and the predicted label. By minimizing the cross entropy loss, the classification performance of the model is optimized and the accuracy of the judgment on the authenticity of the input speech is improved.

[0052] The graph attention network (GAT) includes:

[0053] In GAT, the core operation is to aggregate the information of a node and its neighboring nodes through the self-attention mechanism to capture the relationship between nodes and generate more discriminative features. The input of the GAT layer is a graph G = (N, E, H), where N represents the node set, E represents the edge set, and H is the feature representation of each node. First, for each node n, traverse its neighboring node set M(n) and use w map Calculate the attention coefficient a between node n and its neighbor node u u,n :

[0054]

[0055] Among them, h n and h u Represent the original features of itself and neighbor nodes respectively, ⊙ represents the multiplication of the corresponding elements of the features, and w map It is a weight vector with the same dimension as the node feature, which is used to weight each element to obtain the attention score that measures the strength of the relationship between nodes.

[0056] Next, the aggregated feature m of the central node n is obtained by weighted summation n :

[0057]

[0058] Then introduce the learnable mapping matrix W att and W res , used to convert m n and h n Projected to the same target dimension d' (32 in this invention), the new feature representation o is obtained after addition. n :

[0059] o n =SeLU(BN(W att m n +W res h n )) (8)

[0060] Where SeLU represents the activation function and BN represents batch normalization. This structure is achieved by fusing the aggregated features m n and its own characteristics n , so that o n It also contains local neighborhood information and its own inherent information, which further improves the discriminative ability of node representation, and combines SeLU and BN to enhance the stability and expression ability of the model.

[0061] The graph pooling module (GPM) specifically includes:

[0062] The main function of GPM is to select the most representative nodes to improve the discriminative power of the model. The implementation process is as follows:

[0063] When the input graph contains N nodes, the node feature matrix is ​​H∈R N×d , where d is the dimension of the node feature. GPM first passes a learnable projection vector p∈R d , calculate the projection score s for each node i :

[0064]

[0065] h i is the feature representation of node i, p T represents the transpose of the projection vector. Then according to the projection score s i Sort all nodes, select the first k nodes with the largest projection scores, and get the corresponding node index set idh. The node feature matrix is ​​updated to H'∈R k×d :

[0066] H'=H idh ×sigmoid(s idh ) (10)

[0067] Among them, H idh and idh They represent the features of the selected nodes and the corresponding projection scores respectively, and the unselected nodes will be discarded. While reducing the computational complexity, GPM retains important structural information, which helps to improve the classification performance of the model.

[0068] Step 5: The input features LGP of the fake and real speech are processed in parallel, the nodes of both sides are flattened and concatenated as the input of the fully connected layer, the confidence of the two categories of true / false is generated, the confidence is converted into probability through the softmax function, and the weighted cross entropy loss function (Weighted Cross Entropy, WCE) is used to calculate the error between the actual label and the predicted label, and the loss is minimized to optimize the classification performance of the model and improve the accuracy of the judgment of the authenticity of the speech.

[0069] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A forged speech detection method using dual-dimensional graph attention, characterized in that: The following steps are involved: Step 1: First, pre-emphasize the original speech to enhance the details of the high-frequency waveform, then perform windowing and framing, and extract the linear frequency cepstral coefficient features of each frame; Step 2: Based on the linear frequency cepstral coefficient features, the BP algorithm is used to perform GMM modeling on the real speech and the fake speech respectively. The expectation maximization algorithm is used to iteratively operate the training data until convergence. The number of speech frames is counted and the LGP of each frame of speech features in each Gaussian component is calculated respectively. Then, the constant term is subtracted and normalization is performed. Step 3: Input the LGP feature matrix into six residual block modules. Each residual block consists of two 2D convolutional layers, a SeLU activation function, and a batch normalization layer. A 2×2 maximum pooling layer is added at the end of the first four residual blocks to reduce complexity and improve generalization ability. Step 4: Input the deep features generated by the residual block into the temporal attention module and the Gaussian component attention module respectively to obtain graph nodes of the same dimension but different numbers. Both attention modules contain a graph attention network and a graph pooling module. Then, the graph nodes obtained by the two networks are concatenated and input into the Gaussian component-temporal joint graph attention module to obtain the final graph nodes. Step 5: The input features LGP of the fake and real speech are processed in parallel, the nodes of both sides are flattened and concatenated as the input of the fully connected layer, the confidence of the two categories of true / false is generated, the confidence is converted into probability through the softmax function, and the weighted cross entropy loss function is used to calculate the error between the actual label and the predicted label. The loss is minimized to optimize the classification performance of the model and improve the accuracy of the judgment of the authenticity of the speech.

2. The method for detecting forged speech using dual-dimensional graph attention according to claim 1, characterized in that: The pre-emphasis processing and extraction of linear frequency cepstrum coefficient features in step 1 specifically include: The input speech signal is first pre-emphasized, and a first-order high-pass filter is used to enhance the high-frequency components to compensate for the natural attenuation of high-frequency energy. The signal is divided into frames with a length of 20-30 milliseconds, and the frame shift is set to 50% of the frame length to achieve the overlap of adjacent frames and retain the correlation between signal frames. Each frame of speech signal is windowed to reduce edge discontinuity; Perform fast Fourier transform on each frame of speech, convert the time domain signal into a frequency domain signal, obtain spectrum information, and use a set of triangular filter banks with linear frequency intervals to filter the spectrum; The natural logarithm of the energy output by the filter is taken to make the data more consistent with the human ear's perception of loudness, and the multiplication operation is converted into an addition operation. The obtained logarithmic energy sequence is subjected to discrete cosine transform to remove the correlation between features and compress the data dimension to obtain the linear frequency scale cepstral coefficient feature, where each coefficient represents the spectral feature of the speech signal.

3. The method for detecting forged speech using dual-dimensional graph attention according to claim 1, characterized in that: The GMM modeling and LGP feature extraction in step 2 specifically include: The BP algorithm is combined with GMM for training, and the Gaussian distribution of a single component is gradually evolved into a Gaussian model with multiple components. The global mean and variance are first used to initialize the GMM of a single Gaussian component, and then the number of Gaussian components is gradually increased to adapt to the complexity of the data. GMM is used to model the real and fake speech respectively, and λ is used to represent the parameter set composed of the mean vector, covariance matrix and mixing weight of each Gaussian component, that is: λ={ω i ,m i ,∑ i },i=1,2,…,512 (1) Among them, ω i is the mixing weight and satisfies μ i and∑ i They represent the mean vector and covariance matrix of the i-th Gaussian component respectively, and the model parameter set λ is obtained by the EM algorithm; In each split of the BP algorithm, each Gaussian component is divided into two. This split is achieved by adding or subtracting a small disturbance to the mean vector of the original component. Given a mean vector μ and covariance matrix Σ of a Gaussian component, the mean vectors of the two newly generated Gaussian components are μ+ò and μ-ò, respectively, where ò is a small disturbance vector whose value is δ×eigenvector(Σ), δ is a preset constant, and eigenvector(Σ) represents the main eigenvector of the covariance matrix Σ, indicating the direction of the maximum variation of the data. The formula is as follows: m new1 = μ+δ×eigenvector(Σ) (2) m new2 = μ-δ×eigenvector(Σ) (3) μ new1 and μ new2 is the mean vector of the two newly generated Gaussian components, the covariance matrix Σ remains unchanged, the initial weights of the two new Gaussian components are set to half of the original component weights to ensure that the sum remains consistent. After each split, the new GMM model contains more components, and an appropriate number of EM iterations are performed to estimate the new GMM parameters to ensure that the model parameters can fully fit the data after the introduction of new components. When the last split is completed, the EM iteration process continues until the change in model parameters is lower than a certain threshold, or the preset number of iterations is reached, and the GMM models that fit the real speech and the fake speech are obtained respectively; After the GMM models of real speech and fake speech are established, the LFCC features of speech are input into the two GMM models respectively to obtain the LGP features of each frame of speech features in GMM. Assuming x t Represents the LFCC feature vector of the t-th frame speech signal, and finds x t After taking the logarithm of the probability density of each component, we can get: where d is the vector x t The dimension of x is , T represents the transposition operator. In order to focus on the discriminative features, the direct reflection of x is retained. t The probability of the items under different components is only retained in formula (4) t The relevant items are x t The logarithmic probability of the i-th Gaussian component is: After calculating the LGP of all speech frames in each Gaussian component t,i Afterwards, the data is standardized along the Gaussian component dimension to a distribution with a mean of 0 and a variance of 1, and the LGP feature matrices under the real speech GMM and the fake speech GMM are obtained respectively.

4. The method for detecting forged speech using dual-dimensional graph attention according to claim 1, characterized in that: The step 4 specifically includes: In the module, the dimension of the input feature Q is (64, 25, 32), where 64 represents the number of feature channels, 25 and 32 represent the number of time dimensions and Gaussian component dimensions, respectively. The input data is processed by the Gaussian component attention module to capture the features on the Gaussian component dimension. The maximum pooling operation is performed along the time dimension to obtain the feature representation on the Gaussian component dimension, and a feature map of size (64, 1, 32) is obtained. The feature map is dimensionally transformed to obtain a representation of dimension (64, 32) and input into the graph attention network to generate attention-weighted graph features of dimension (32, 32). The graph features are pooled using the graph pooling module (GPM) to select the most representative nodes and obtain graph features of dimension (32, 18).

5. The method for detecting forged speech using dual-dimensional graph attention according to claim 4, characterized in that: The graph attention network in step 4 specifically includes: In the graph attention network, the information of nodes and their neighboring nodes is aggregated through the self-attention mechanism to capture the relationship between nodes and generate more discriminative features. The input of the graph attention network layer is a graph G = (N, E, H), where N represents the node set, E represents the edge set, and H is the feature representation of each node. For each node n, its neighbor node set M(n) is traversed and the neighbor node set M(n) is obtained through w map Calculate the attention coefficient a between node n and its neighbor node u u,n : Among them, h n and h u Represent the original features of itself and neighbor nodes respectively, ⊙ represents the multiplication of the corresponding elements of the features, and w map is a weight vector with the same dimension as the node feature, which is used to weight each element to obtain the attention score that measures the strength of the relationship between nodes; The aggregated feature m of the central node n is obtained by weighted summation n : Then, the learnable mapping matrix W is introduced. att and W res , used to convert m n and h n Projected to the same target dimension d', the new feature representation o is obtained after addition. n : o n =SeLU(BN(W att m n +W res h n )) (8) Where SeLU represents the activation function and BN represents batch normalization.

6. The method for detecting forged speech using dual-dimensional graph attention according to claim 4, characterized in that: The image pooling module in step 4 specifically includes: When the input graph contains N nodes, the node feature matrix is ​​H∈R N×d , where d is the dimension of the node feature, GPM first passes a learnable projection vector p∈R d , calculate the projection score s for each node i : s i =p T h i (9) h i is the feature representation of node i, p T Represents the transpose of the projection vector, and then according to the projection score s i Sort all nodes, select the first k nodes with the largest projection scores, obtain the corresponding node index set idh, and update the node feature matrix to H'∈R k×d : H'=H idh ×sigmoid(s idh ) (10) Among them, H idh and idh They represent the features of the selected nodes and the corresponding projection scores respectively. The unselected nodes will be discarded.

Images