A context-aware masking feature enhancement method for detecting synthesized speech

Abstract

The application discloses a kind of context-aware mask feature enhancement synthetic speech detection methods, it is related to speech signal processing technical field, method includes: to original audio data pre-processing, the speech data after pre-processing is input to SincNet filter extraction speech timing feature;The speech timing feature is input to CAM-ResNet extraction context-aware local feature;Context-aware local feature is input L-Transformer network calculation global feature;Again through fully connected layer and sigmoid activation function classification;After iteration training, select the best model for synthetic speech detection.CAM-ResNet network in the application can effectively weaken irrelevant features, focus on context-related features, thereby improving the perception of context-effective features, capturing the local details of speech context-aware features;Again after having L-Transformer deep network calculation global feature, to improve the detection accuracy of synthetic speech.

Description

Technical Field

[0001] This invention relates to the field of speech signal processing technology, and more specifically, to a synthetic speech detection method enhanced by context-aware mask features. Background Technology

[0002] With the rapid development of AI-generated content, synthetic speech technology has achieved significant breakthroughs. Modern synthetic speech systems include text-to-speech (TTS) and voice-to-text (VC) technologies. These technologies can generate natural, fluent, and near-realistic speech, accurately mimicking the individual's intonation, rhythm, and emotions. These advancements have led to their application in various fields, including intelligent voice assistants, intelligent voice customer service systems, virtual reality, and entertainment games. However, the illegal misuse of synthetic speech technology is becoming increasingly serious, raising a series of grave security and privacy issues.

[0003] However, with the continuous updates and iterations of speech synthesis technology, existing systems have limitations in dealing with traditional known spoofing attack types and are ineffective against new and unknown spoofing methods. At the same time, speech samples are often affected by environmental noise, and these irrelevant factors significantly affect detection performance. Therefore, synthetic speech detection technology must keep pace with the times, strengthen detection algorithms to deal with various new spoofing attacks and reduce the influence of irrelevant factors, so as to effectively distinguish between real and spoofed speech. This can not only prevent personal identity from being impersonated and tampered with, but also provide effective protection for information security in the Internet and Internet of Things environments, thereby improving users' sense of security and promoting the reliability and practicality of biometric technology in various application scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention discloses a synthetic speech detection method enhanced by context-aware mask features, aiming to achieve efficient synthetic speech detection through a multi-layered structure of context-aware feature extraction and global feature analysis.

[0005] This invention provides a context-aware mask feature-enhanced synthetic speech detection method, comprising the following steps:

[0006] Step S1: Obtain the raw audio data, perform data preprocessing on the raw audio data, and use the preprocessed speech data as the input of the synthesized speech detection framework;

[0007] Step S2: Input the preprocessed speech data into the SincNet filter module to extract speech temporal features;

[0008] Step S3: Input the speech temporal features into a ResNet module with context-aware masking to extract context-aware local features;

[0009] Step S4: Input the context-aware local features into a Transformer network module with Llama style to calculate global features;

[0010] Step S5: Process the global features using a fully connected layer, classify them using the Sigmoid activation function, and output the detection results of the synthesized speech.

[0011] In this scheme, in step S1, the data preprocessing includes normalization processing, pruning and padding processing, and application of pre-emphasis processing. The synthesized speech detection framework includes a SincNet filter module, a ResNet module with a context-aware mask module, and a Transformer network module with Llama style.

[0012] In this scheme, in step S2, the preprocessed speech signal x[n] is filtered using the SincNet filter module to obtain the speech temporal features y[n]. In the SincNet filter, convolution is defined as:

[0013] y[n] = x[n] * g[n,f1,f2] · w[n]

[0014] Where g represents the bandpass filter, n represents the timestamp, f1 and f2 represent the learnable parameters, representing the low cutoff frequency and high cutoff frequency of the bandpass filter, respectively, and w[n] represents the Hamming window function.

[0015] In this scheme, in step S3, the ResNet module with context-aware mask is referred to as the CAM-ResNet module;

[0016] Using a CAM-ResNet module with the same 6-layer structure, the speech temporal features y[n] output by the SincNet filter module are extracted using two convolutional blocks with a Conv-ReLU-BatchNorm structure. The computation of each convolutional block is represented as: Y local =BatchNorm(ReLU(Conv(Y)));

[0017] The extracted features are subjected to a context-aware mask. Different weights are assigned to the extracted features using attention statistical pooling. The pooled features are then processed using a sigmoid activation function to obtain a context-aware mask. This mask reduces the influence of irrelevant features and focuses on the context-aware feature representation. The context-aware mask M is expressed as follows:

[0018] M = σ(W·S + b)

[0019] Where W represents the weight matrix of the linear layer, b represents the bias term, and σ represents the Sigmoid function, which is used to generate a context-aware mask between 0 and 1;

[0020] Context-aware masks are used to enhance speech-related features, suppress environmental noise, and extract more discriminative local features for perceiving people's presence or absence. These local features are represented as follows:

[0021] Y cam =Y res +Y local ⊙M

[0022] Among them, Y res Represents the residual.

[0023] In this scheme, in step S4, two layers of Llama-style Transformer modules are set up for global feature extraction. The Llama-style Transformer modules are referred to as L-Transformer modules.

[0024] The context-aware local features are imported into the L-Transformer module, which divides the context-aware local features into multiple adjacent element pairs, introduces rotation angles for position encoding, determines position information, and performs RMSNorm normalization on the position-encoded features to stabilize their numerical range.

[0025] The normalized features are imported into a multi-head attention layer. The output of the single-head attention layer is obtained through a self-attention mechanism. The outputs of all attention heads are concatenated, and the output of the multi-head attention layer is obtained through a linear layer transformation.

[0026] The output of the multi-head attention layer is combined with the position-encoded features. The combined features are then fed into a feedforward neural network. After RMSNorm normalization and linear transformation, global features are obtained using the SwiGLU activation function.

[0027] In this scheme, the two-layer L-Transformer module is set as a dual-path structure to train on the speech data of real speech tags and the speech data of synthesized speech tags;

[0028] After location encoding, the speech-related features are used to obtain global feature embedding vectors for the real speech label path and the synthesized speech label path, respectively. The global feature embedding vectors are concatenated and input into a fully connected layer for detection.

[0029] In this scheme, in step S6, the global features are input into a fully connected layer, and the Sigmoid activation function is used in the fully connected layer to perform category classification, and the probability distribution of real speech labels and synthesized speech labels is output as the detection result.

[0030] In this scheme, the synthetic speech detection framework is trained using training data, specifically as follows:

[0031] Based on the speech-related corpus, we extract corpus samples and match category labels for each corpus sample using the descriptive information of each corpus sample, thereby obtaining a dataset of real speech labels and a dataset of synthetic speech labels.

[0032] The dataset for each label is randomly divided into training set, validation set and test set according to the proportion. The network framework parameters and learning rate are initialized. The corpus samples in the training set are selected as the input of the synthetic speech detection framework. The classification output value is obtained through detection operation.

[0033] The loss function is constructed by calculating the cross-entropy based on the membership function between the corpus samples and the category labels and the classification output value. The network parameters of the synthesized speech detection framework are iteratively updated based on the forward propagation loss. The detection results are verified using a validation set in each iteration.

[0034] If the detection performance of the synthesized speech detection framework meets the preset standard, the current network parameters are retained, and then the test set is used for testing. If the test results meet the preset test standard, the current network parameters are used as the structural parameters of the synthesized speech detection framework, and the synthesized speech detection framework is output.

[0035] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0036] This invention introduces CAM-ResNet and L-Transformer networks, combining context-aware masks and deep feature modeling. The CAM-ResNet network can effectively reduce irrelevant features and focus on context-related features, thereby improving the perception of effective context features and capturing local details of speech context-aware features. Then, global features are calculated through an L-Transformer deep network to improve the detection accuracy of synthesized speech. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples 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 according to these drawings without creative effort.

[0038] Figure 1 A flowchart of the synthetic speech detection method using context-aware mask feature enhancement according to the present invention is shown;

[0039] Figure 2 An architecture diagram of the speech detection framework according to an embodiment of the present invention is shown;

[0040] Figure 3 The algorithm flowchart of the CAM-ResNet module in an embodiment of the present invention is shown;

[0041] Figure 4 The algorithm flowchart of the L-Transformer module in an embodiment of the present invention is shown;

[0042] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0043] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0045] Figure 1 Here is a flowchart of a context-aware mask feature-enhanced synthetic speech detection method provided in Embodiment 1 of the present invention;

[0046] like Figure 1 As shown, this invention provides a synthetic speech detection method with context-aware mask feature enhancement, comprising the following steps:

[0047] Step S1: Obtain the raw audio data, perform data preprocessing on the raw audio data, and use the preprocessed speech data as the input of the synthesized speech detection framework;

[0048] Step S2: Input the preprocessed speech data into the SincNet filter module to extract speech temporal features;

[0049] Step S3: Input the speech temporal features into a ResNet module with context-aware masking to extract context-aware local features;

[0050] Step S4: Input the context-aware local features into a Transformer network module with Llama style to calculate global features;

[0051] Step S5: Process the global features using a fully connected layer, classify them using the Sigmoid activation function, and output the detection results of the synthesized speech.

[0052] It should be noted that the data preprocessing includes standardization, trimming and padding, and pre-emphasis processing to avoid interference from excessive differences in the input data. Trimming and padding the original audio data yields a speech signal within a stationary time range, which facilitates subsequent audio data processing. Synthetic speech exists at the sub-band level, and synthesized speech is typically detected using the high-frequency components exhibited within the same sub-band. Since the human vocal system suppresses the high-frequency components of the speech signal, pre-emphasis is needed to enhance these components. Pre-emphasis processing can better highlight high-frequency formants.

[0053] like Figure 2 As shown, the synthesized speech detection framework includes a SincNet filter module, a ResNet module with a context-aware mask module (CAM-ResNet module), and a Transformer network module with Llama style (L-Transformer module).

[0054] The preprocessed speech signal x[n] is filtered using the SincNet filter module to obtain the speech temporal features y[n]. In the SincNet filter, the first convolutional layer is considered as a set of trainable finite impulse response (FIR) filters. Each convolution is defined as y[n] = x[n] * h[n], where h[n] represents a filter of length L. h[n] is predefined as a function g, and function g has only a few learnable parameters θ. The convolution formula is transformed into y[n] = x[n] * g[n, θ]. In this scheme, function g is defined as a bandpass filter, which can be expressed as the difference between the results of two low-pass filters rect, expressed as... f represents the parameters of the bandpass filter. After inverse Fourier transform, the time-domain representation of the bandpass filter g is obtained: g[n,f1,f2]=2f2sinc(2πf2n)-2f1sinc(2πf1n), where the sinc function is defined as: sinc(x)=sin(x) / x; to reduce the spectral leakage effect, a window function is added during the convolution process. Therefore, the windowed bandpass filter is represented as: g w [n,f1,f2]=g[n,f1,f2]·w[n], therefore the convolution process of the SincNet filter module is defined as:

[0055] y[n] = x[n] * g[n,f1,f2] · w[n]

[0056] Where g represents the bandpass filter, n represents the timestamp, f1 and f2 represent the learnable parameters, representing the low cutoff frequency and high cutoff frequency of the bandpass filter, respectively, and w[n] represents the Hamming window function.

[0057] Figure 3 The algorithm flowchart of the CAM-ResNet module in an embodiment of the present invention is shown.

[0058] It should be noted that the ResNet module with context-aware mask is referred to as the CAM-ResNet module. It further extracts local features from the temporal features of the speech after SincNet filtering and improves the generalization ability of the model by using context-aware mask to enhance the features.

[0059] Using a CAM-ResNet module with the same 6-layer structure, the speech temporal features y[n] output by the SincNet filter module are extracted using two convolutional blocks with a Conv-ReLU-BatchNorm structure. The computation of each convolutional block is represented as: Y local =BatchNorm(ReLU(Conv(Y))); A context-aware mask is applied to the extracted speech-related features. The context-aware mask module (CAM) consists of three parts: attention statistical pooling, a linear layer, and a sigmoid activation function. The extracted features are assigned different weights using attention statistical pooling; for example, noisy frames are assigned smaller weights, and frames related to human voices are assigned larger weights. Attention statistical pooling helps the synthetic speech detection framework focus on and learn key frame-level features. It applies an attention mechanism to the frame-level features along the frame length T to obtain the importance of the frame-level features for each channel at different times. After a nonlinear transformation, it is converted into channel-dependent self-attention values ​​through a linear layer, obtaining the importance of the frame-level features for each channel. These importance values ​​are then mapped to weights, expressed as:

[0060]

[0061]

[0062] Where s t y represents the frame-level feature of frame t. t For the importance of each channel, W represents the weight, δ represents the activation function, p represents the bias, q represents the linear layer bias, and v represents the linear layer weight.

[0063] The importance of different frame-level features in each channel is weighted and calculated to obtain the weighted average and weighted standard deviation of the channel. The weighted average and weighted standard deviation of the channel are concatenated and output for subsequent processing.

[0064] The pooled features are processed using the Sigmoid activation function to obtain a context-aware mask. This mask reduces the influence of irrelevant features and focuses on the context-aware feature representation. The context-aware mask M is expressed as follows:

[0065] M = σ(W·S + b)

[0066] Where W represents the weight matrix of the linear layer, b represents the bias term, and σ represents the Sigmoid function, which is used to generate a context-aware mask between 0 and 1;

[0067] Context-aware masks are used to enhance speech-related features, suppress environmental noise, and extract more discriminative local features for perceiving people's presence or absence. These local features are represented as follows:

[0068] Y cam =Y res +Y local ⊙M

[0069] Among them, Y res Represents the residual.

[0070] Figure 4 The algorithm flowchart of the L-Transformer module in an embodiment of the present invention is shown.

[0071] It should be noted that two Llama-style Transformer modules are set up for global feature extraction. The Llama-style improvements include Rotation Position Embedding (RoPE), RMSnorm normalization, and SwiGLU activation function. The Llama-style Transformer module is referred to as the L-Transformer module.

[0072] The context-aware local features are imported into the L-Transformer module for Rotation-Based Position Encoding (RoPE). This process divides the context-aware local features into multiple adjacent element pairs, introduces rotation angles for position encoding, and determines positional information. This approach more effectively represents relative positional relationships. Assume the feature embedding vector for each position is X = [x1, x2, ..., x...]. d ], where d is the embedding dimension. For each pair of adjacent elements (x) in the vector 2i ,x 2i+1 The rotation transformation is obtained by performing rotation encoding as follows:

[0073]

[0074] Wherein, rotation angle θ i With position p and frequency parameter ω i Related:

[0075] θ i =p·ω i

[0076] Based on the property that the inner product satisfies linear superposition, any even-dimensional RoPE can be represented as a concatenation of two-dimensional cases, i.e.:

[0077]

[0078] The position-encoded features are normalized using RMSNorm to stabilize their numerical range. The normalized features are then introduced into a multi-head attention layer. The output of a single-head attention layer is obtained through a self-attention mechanism. The outputs of all attention heads are concatenated, and the output of the multi-head attention layer is obtained through a linear layer transformation.

[0079] The output of the multi-head attention layer is combined with the position-encoded features. These combined features are then fed into a feedforward neural network. After RMSnorm normalization and linear transformation, the global features are obtained using the SwiGLU activation function. The SwiGLU activation function is a combination of the Swish activation function and a gated linear unit. For an input vector x, a linear layer produces two branch outputs x1 and x2. The output value is obtained by multiplying the Swish-activated x1 and x2.

[0080] In the L-Transformer module, a two-layer L-Transformer module is configured as a dual-path structure to train on speech data of both real and synthesized speech labels. After positional encoding, the speech-related features are processed through both the real and synthesized speech label paths to obtain global feature embedding vectors for each path. These global feature embedding vectors are concatenated and input into a fully connected layer for detection. The global features are then fed into the fully connected layer, where a sigmoid activation function is used for category classification. The probability distributions of the real and synthesized speech labels are output as the detection results.

[0081] It should be noted that in the training of the synthetic speech detection framework, the CAM-ResNet and L-Transformer networks are pre-trained first. Fully connected layers with sigmoid outputs are temporarily added after each path in the L-Transformer module, and training is performed using the cross-entropy loss function. Then, the fully connected layers are removed, and the parameters of the CAM-ResNet and L-Transformer networks are frozen to train the fully connected layer classifier.

[0082] Speech samples are extracted from a speech-related corpus. Category labels are matched to the samples based on their descriptive information, resulting in datasets of real and synthetic speech labels. Each label's dataset is randomly divided into training, validation, and test sets according to a set ratio. Network framework parameters and learning rates are initialized. Speech samples from the training set are selected as input to the synthetic speech detection framework, and classification output values ​​z are obtained through detection operations. i,b Based on the dependency function f between the corpus samples and the category labels i,b The loss function L is constructed by calculating the cross-entropy of the classification output values, and expressed as:

[0083]

[0084] Where b represents the category label, i represents the corpus sample, M represents the total number of corpus samples, N represents the total number of category labels, and f m,i The function represents the relationship between the corpus sample and the category label. It takes the value 1 when the corpus sample i belongs to the analogy label b, and 0 otherwise.

[0085] The network parameters of the synthesized speech detection framework are iteratively updated based on the forward propagation loss. In each iteration, the detection results are verified using a validation set. If the detection performance of the synthesized speech detection framework meets the preset standard, the current network parameters are retained. Subsequently, the test set is used for testing. If the test results meet the preset test standard, the current network parameters are used as the structural parameters of the synthesized speech detection framework, and the synthesized speech detection framework is output.

[0086] In a preferred embodiment, the ASVspoof 2019 and ASVspoof 2021 logical access corpora are used for training and validation of the synthetic speech detection framework. The real samples in the corpora are derived from the VCTK corpus and contain 46 male speakers and 61 female speakers, totaling 107 speakers. The forged samples are generated using different TTS (text-to-speech) and VC (voice-to-speech) techniques than those used in the training and development sets. Compared to previous synthetic speech or logical access databases, the ASVspoof 2019 logical access corpus offers greater scale and diversity in attack algorithms. The ASVspoof 2021 logical access test dataset includes a collection of real and deceptive speech transmitted through various telephone systems, including Voice over IP (VoIP) and the Public Switched Telephone Network (PSTN). Transmitting all data through telephone systems introduces interference variations that can occur in a range of real-world logical access application scenarios.

[0087] The initial learning rate was 3e-4, the batch size was 16, the optimizer used was AdamW, and the number of training epochs was set to 100. During training, a combination of learning rate preheating and cosine annealing was used. In the first 10 training epochs, the learning rate preheating strategy was used to gradually increase the learning rate; after reaching the maximum period of cosine annealing, the learning rate began to decrease.

[0088] Finally, the trained synthetic speech detection framework was tested on the logical access corpus test set of ASVspoof2021. The equal error rate (EER) was used as the test indicator. The measured EER score was 3.35%, which exceeded the preset test standard based on the pre-prepared automatic speaker verification system.

[0089] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] Alternatively, if the integrated modules of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0091] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A synthetic speech detection method enhanced with context-aware mask features, characterized in that, This includes the model building and training phase, as well as the detection and inference phase; The model building and training phase includes the following steps: S1: Construct a synthetic speech detection framework, which sequentially includes a SincNet filter module, a context-aware mask ResNet module, and a Llama-style dual-path Transformer module; the context-aware mask ResNet module is denoted as the CAM-ResNet module; the Llama-style Transformer module is denoted as the L-Transformer module; S2: Obtain a training dataset with real speech labels and synthesized speech labels; S3: Input the training data into the synthesized speech detection framework for iterative training, and perform the following operations during the forward propagation of each iteration: S31: Extract the temporal features of the input speech using the SincNet filter module; S32: The temporal features are processed by the CAM-ResNet module to generate context-aware local features; the CAM-ResNet module generates a speech context-aware mask through attention statistical pooling, and the context-aware mask... Represented as: ， in, This represents the weight matrix of the linear layer. Indicates the bias term. This represents the Sigmoid function, used to generate a context-aware mask between 0 and 1; Context-aware masks are used to enhance speech-related features, suppress environmental noise, and extract more discriminative local features that capture the sense of people moving up and down. Represented as: ， in, Represents the residual; Represents convolutional features; S33: Input the context-aware local features into the L-Transformer module; the L-Transformer module has parallel real speech paths and synthesized speech paths, and uses rotational position encoding, RMSNorm normalization and Swiglu activation function; in this step, the context-aware local features are input into the two paths respectively, and processed by the path-independent Transformer encoder to obtain two sets of global feature embedding vectors that correspond to the real speech and synthesized speech categories respectively and have discriminative differences; S34: Concatenate the two sets of global feature embedding vectors, input them into a fully connected layer, and output the prediction probability through the Sigmoid activation function; S4: Calculate the loss based on the predicted probability and the true label, and update the network parameters through backpropagation until the model converges; The detection inference phase includes the following steps: S5: Input the preprocessed speech to be detected into the trained synthetic speech detection framework, and process it sequentially through steps S31 to S34. Finally, the Sigmoid activation function outputs the probability that the speech to be detected is synthetic speech.

2. The synthetic speech detection method with context-aware mask feature enhancement according to claim 1, characterized in that, Input speech is extracted using the SincNet filter module. Temporal characteristics In the SincNet filter, convolution is defined as: ， in, Indicates a bandpass filter. Represents a timestamp. and These represent the learnable parameters, which are the low cutoff frequency and high cutoff frequency of the bandpass filter, respectively. This represents the Hamming window function.

3. The synthetic speech detection method with context-aware mask feature enhancement according to claim 1, characterized in that, The rotational position encoding adopts the RoPE method, which encodes position information into a rotational transformation of feature vectors.

4. The synthetic speech detection method with context-aware mask feature enhancement according to claim 1, characterized in that, The preprocessing of the speech data to be detected includes standardization, cropping and padding, and pre-emphasis.

5. The synthetic speech detection method with context-aware mask feature enhancement according to claim 1, characterized in that, The synthetic speech detection framework is trained using training data, specifically as follows: Based on the speech-related corpus, we extract corpus samples and match category labels for each corpus sample using the descriptive information of each corpus sample, thereby obtaining a dataset of real speech labels and a dataset of synthetic speech labels. The dataset for each label is randomly divided into training set, validation set and test set according to the proportion. The network framework parameters and learning rate are initialized. The corpus samples in the training set are selected as the input of the synthetic speech detection framework. The classification output value is obtained through detection operation. The loss function is constructed by calculating the cross-entropy based on the membership function between the corpus samples and the category labels and the classification output value. The network parameters of the synthesized speech detection framework are iteratively updated based on the forward propagation loss. The detection results are verified using a validation set in each iteration. If the detection performance of the synthesized speech detection framework meets the preset standard, the current network parameters are retained, and then the test set is used for testing. If the test results meet the preset test standard, the current network parameters are used as the structural parameters of the synthesized speech detection framework, and the synthesized speech detection framework is output.

Images