Speech authentication method and system based on residual attention network

Abstract

The present disclosure provides a voice authentication method and system based on a residual attention network, which comprises: obtaining audio data to be detected and performing corresponding preprocessing; performing feature extraction on the preprocessed audio data, and performing frame processing on the extracted voice feature data to obtain voice signal feature data with a fixed frame length; based on the voice signal feature data, using a pre-trained residual attention network model to obtain enhanced feature data; wherein the residual attention network model comprises sequentially connected convolution modules, multi-scale residual modules, shrinkage excitation units, attention pooling modules and a fully connected layer; and inputting the enhanced feature data into a pre-trained classifier to obtain a voice authentication result.

Description

Technical Field

[0001] This disclosure belongs to the field of speech detection and deep learning technology, and in particular relates to a speech authentication method and system based on residual attention networks. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Voice not only conveys the content information that needs to be expressed, but also contains the unique identity characteristics of the speaker, which can be used to verify the speaker's identity. However, automatic speaker verification systems are vulnerable to voice spoofing attacks, posing a threat to people's interests and social security. Therefore, research on voice spoofing detection technology is of great significance.

[0004] The inventors discovered that although residual neural networks have been widely used in the field of speech forgery detection and have achieved good performance, by introducing residual connections into convolutional neural networks and standardizing residual blocks, the residual network structure can train deep networks to achieve highly competitive recognition performance. However, when the input feature space becomes more complex, simply increasing the depth and width of the residual neural network cannot fully realize its performance potential and cannot meet the requirements for speech forgery detection accuracy in such scenarios. Summary of the Invention

[0005] To address the aforementioned issues, this disclosure provides a speech forgery detection method and system based on residual attention networks. The scheme employs a residual attention network structure to explicitly model the interdependencies between channels, adaptively recalibrates the feature responses in the channel directions, and assigns different weights to speech features in different time frames. This effectively focuses on the forged parts of the speech, and by training the model based on a single-classification loss function, it effectively improves the accuracy of speech forgery detection.

[0006] According to a first aspect of the embodiments of this disclosure, a speech authentication method based on a residual attention network is provided, comprising:

[0007] Acquire the audio data to be detected and perform corresponding preprocessing;

[0008] Feature extraction is performed on the preprocessed audio data, and the extracted speech feature data is then processed into frames to obtain speech signal feature data with a fixed frame length.

[0009] Based on the speech signal feature data, enhanced feature data is obtained using a pre-trained residual attention network model. The residual attention network model includes sequentially connected convolutional modules, multi-scale residual modules, a contraction activation unit, an attention pooling module, and fully connected layers. The contraction activation unit receives the deep feature vector output by the multi-scale residual module and sequentially passes it through a global average pooling operation, a fully connected layer, a ReLU activation function, another fully connected layer, and a sigmoid function to obtain the attention score of the channel feature vector. The output of the contraction activation unit is obtained by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module.

[0010] The enhanced feature data is input into a pre-trained classifier to obtain the speech authentication results.

[0011] Furthermore, the multi-scale residual module receives the feature vector obtained through convolution processing. The feature vector is then processed by a convolutional layer to uniformly divide the feature map into several subsets according to the channel dimension. The remaining subsets, excluding the first subset, are processed by a convolutional filter. Starting from the third subset, the output of the subset is added to the output of the previous subset and then input into the convolutional filter to obtain the output value corresponding to each subset. The output values ​​of each subset are concatenated and input into the convolutional layer to obtain the processing result. The processing result is then residually concatenated with the output of the convolutional module to obtain the deep feature vector.

[0012] Furthermore, the processing flow of the contraction excitation unit specifically includes:

[0013] The output of the multi-scale residual module is subjected to global average pooling, and the entire spatial dimension is compressed along the channel dimension for feature aggregation mapping to obtain... The feature vector, where C represents the number of channels;

[0014] The feature vector is input into the first fully connected layer, which compresses the C channels into C / R channels, and then passes through a ReLU activation function layer to obtain the channel-compressed feature vector, where R represents the compression ratio.

[0015] The channel-compressed feature vectors are input into the second fully connected layer to restore C / R channels back to C channels, and the attention scores of the channel feature vectors are obtained by using the sigmoid function.

[0016] The output of the contraction excitation unit is obtained by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module.

[0017] Furthermore, the processing flow of the attention pooling module specifically includes:

[0018] The deep feature vector of each time frame output by the shrinking excitation unit is multiplied by a preset linear transformation matrix. All time frames share the same parameters, and the attention score is obtained by using the tanh activation function.

[0019] The attention scores are normalized and scaled to [0,1] using the softmax function to obtain the attention weights.

[0020] The first-order and second-order statistical features of the weighted feature vector are concatenated and then linearly transformed through a fully connected layer to obtain the final global feature vector.

[0021] Furthermore, the convolution module includes sequentially connected convolutional layers, batch normalization layers, and ReLU activation functions.

[0022] Furthermore, the step of extracting features from the preprocessed audio data and performing frame segmentation on the extracted speech feature data specifically involves: extracting linear frequency cepstral coefficient features from the preprocessed audio data using a preset sampling frequency, preset window length, preset number of FFT points, and preset number of filters, along with their increments and double increment coefficients; and performing frame segmentation on the extracted speech feature data to obtain speech signal feature data of a preset frame length.

[0023] Furthermore, the training of the residual attention network model specifically involves:

[0024] Obtain speech feature data from the training and validation sets;

[0025] The speech feature data from the training set is input into the convolution module to obtain feature vectors;

[0026] The feature vector is input into the multi-scale residual module to obtain the deep feature vector;

[0027] Deep features are input into the contraction activation unit to obtain a channel attention-enhanced feature vector.

[0028] The channel attention-enhanced feature vector is input into the attention pooling module to generate a global feature vector, which is then input into the fully connected layer for classification.

[0029] Train using a single-class loss function and save the model parameters;

[0030] The validation set data is input into the model for validation. The model parameters are adjusted based on the validation results to obtain the trained attention residual network model.

[0031] According to a second aspect of the present disclosure, a speech authentication system based on a residual attention network is provided, comprising:

[0032] The data acquisition unit is configured to acquire the audio data to be detected and perform corresponding preprocessing.

[0033] The feature extraction unit is configured to extract features from the preprocessed audio data and perform frame segmentation on the extracted speech feature data to obtain speech signal feature data of fixed frame length.

[0034] A feature enhancement unit is configured to obtain enhanced feature data based on the speech signal feature data using a pre-trained residual attention network model. The residual attention network model includes sequentially connected convolutional modules, a multi-scale residual module, a contraction activation unit, an attention pooling module, and a fully connected layer. The contraction activation unit receives the deep feature vector output by the multi-scale residual module and sequentially passes it through a global average pooling operation, a fully connected layer, a ReLU activation function, another fully connected layer, and a sigmoid function to obtain the attention score of the channel feature vector. The output of the contraction activation unit is obtained by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module.

[0035] The anti-spoofing unit is used to input the enhanced feature data into a pre-trained classifier to obtain the speech anti-spoofing results.

[0036] According to a third aspect of the present disclosure, a computer-readable storage medium is provided having a program stored thereon that, when executed by a processor, implements a speech authentication method based on a residual attention network as described above.

[0037] According to a fourth aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement a speech authentication method based on a residual attention network as described above.

[0038] Compared with the prior art, the beneficial effects of this disclosure are:

[0039] This disclosure provides a speech forgery detection method and system based on residual attention networks. The scheme adopts a residual attention network structure to explicitly model the interdependence between channels, adaptively recalibrates the feature responses of the channel directions, and assigns different weights to the speech features of different time frames, thereby effectively focusing on the speech forgery parts. By training the model based on a single classification loss function, the accuracy of speech forgery detection is effectively improved.

[0040] Advantages of this disclosure in additional aspects will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this disclosure. Attached Figure Description

[0041] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0042] Figure 1 This is a flowchart of the training process for the residual attention network model described in this disclosure.

[0043] Figure 2 This is a flowchart of the shrinkage excitation unit processing described in this disclosure.

[0044] Figure 3 This is a flowchart of the attention pooling module processing described in this disclosure. Detailed Implementation

[0045] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0047] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0048] Where there is no conflict, the embodiments and features described herein can be combined with each other.

[0049] Example 1:

[0050] The purpose of this embodiment is to provide a speech authentication method based on residual attention networks.

[0051] A speech authentication method based on residual attention networks includes:

[0052] Acquire the audio data to be detected and perform corresponding preprocessing;

[0053] Feature extraction is performed on the preprocessed audio data, and the extracted speech feature data is then processed into frames to obtain speech signal feature data with a fixed frame length.

[0054] Based on the speech signal feature data, enhanced feature data is obtained using a pre-trained residual attention network model. The residual attention network model includes sequentially connected convolutional modules, multi-scale residual modules, a contraction activation unit, an attention pooling module, and fully connected layers. The contraction activation unit receives the deep feature vector output by the multi-scale residual module and sequentially passes it through a global average pooling operation, a fully connected layer, a ReLU activation function, another fully connected layer, and a sigmoid function to obtain the attention score of the channel feature vector. The output of the contraction activation unit is obtained by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module.

[0055] The enhanced feature data is input into a pre-trained classifier to obtain the speech authentication results.

[0056] Furthermore, the multi-scale residual module receives the feature vector obtained through convolution processing. The feature vector is then processed by a convolutional layer to uniformly divide the feature map into several subsets according to the channel dimension. The remaining subsets, excluding the first subset, are processed by a convolutional filter. Starting from the third subset, the output of the subset is added to the output of the previous subset and then input into the convolutional filter to obtain the output value corresponding to each subset. The output values ​​of each subset are concatenated and input into the convolutional layer to obtain the processing result. The processing result is then residually concatenated with the output of the convolutional module to obtain the deep feature vector.

[0057] Furthermore, the processing flow of the contraction excitation unit specifically includes:

[0058] The output of the multi-scale residual module is subjected to global average pooling, and the entire spatial dimension is compressed along the channel dimension for feature aggregation mapping to obtain... The feature vector, where C represents the number of channels;

[0059] The feature vector is input into the first fully connected layer, which compresses the C channels into C / R channels, and then passes through a ReLU activation function layer to obtain the channel-compressed feature vector, where R represents the compression ratio.

[0060] The channel-compressed feature vectors are input into the second fully connected layer to restore C / R channels back to C channels, and the attention scores of the channel feature vectors are obtained by using the sigmoid function.

[0061] The output of the contraction excitation unit is obtained by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module.

[0062] Furthermore, the processing flow of the attention pooling module specifically includes:

[0063] The deep feature vector of each time frame output by the shrinking excitation unit is multiplied by a preset linear transformation matrix. All time frames share the same parameters, and the attention score is obtained by using the tanh activation function.

[0064] The attention scores are normalized and scaled to [0,1] using the softmax function to obtain the attention weights.

[0065] The first-order and second-order statistical features of the weighted feature vector are concatenated and then linearly transformed through a fully connected layer to obtain the final global feature vector.

[0066] Furthermore, the convolution module includes sequentially connected convolutional layers, batch normalization layers, and ReLU activation functions.

[0067] Furthermore, the step of extracting features from the preprocessed audio data and performing frame segmentation on the extracted speech feature data specifically involves: extracting linear frequency cepstral coefficient features from the preprocessed audio data using a preset sampling frequency, preset window length, preset number of FFT points, and preset number of filters, along with their increments and double increment coefficients; and performing frame segmentation on the extracted speech feature data to obtain speech signal feature data of a preset frame length.

[0068] Furthermore, the training of the residual attention network model specifically involves:

[0069] Obtain speech feature data from the training and validation sets;

[0070] The speech feature data from the training set is input into the convolution module to obtain feature vectors;

[0071] The feature vector is input into the multi-scale residual module to obtain the deep feature vector;

[0072] Deep features are input into the contraction activation unit to obtain a channel attention-enhanced feature vector.

[0073] The channel attention-enhanced feature vector is input into the attention pooling module to generate a global feature vector, which is then input into the fully connected layer for classification.

[0074] Train using a single-class loss function and save the model parameters;

[0075] The validation set data is input into the model for validation. The model parameters are adjusted based on the validation results to obtain the trained attention residual network model.

[0076] Specifically, for ease of understanding, the following detailed description of the solution in this embodiment is provided in conjunction with the accompanying drawings:

[0077] To address the problems existing in the prior art, this embodiment provides a speech authentication method based on residual attention networks, including:

[0078] S1: Read audio file;

[0079] S2: Preprocess the voice data; wherein, the preprocessing includes voice signal operations such as framing and windowing;

[0080] S3: Extract features from the preprocessed data, and perform frame segmentation on the extracted speech feature data to obtain speech signal feature data with a fixed frame length.

[0081] S4: Input the extracted speech signal feature data into the residual attention network to obtain enhanced feature data. The residual attention network mainly consists of a convolutional module, a multi-scale residual module, a shrinking excitation unit, and an attention pooling module.

[0082] S5: Input the enhanced feature data into the classifier to perform speech forgery detection and determine whether the speech data is real speech.

[0083] Furthermore, in S3, feature extraction is performed on the preprocessed data, specifically including: extracting linear frequency cepstral coefficient features with a sampling frequency of 16kHz, a window length of 20ms, 512 FFT points and 20 filters and their increments and double increment coefficients; performing frame segmentation on the extracted speech feature data to obtain 750 frames of speech signal feature data.

[0084] Furthermore, such as Figure 1 As shown, the training process of the residual attention network model is as follows:

[0085] (1) Obtain speech feature data of training set and validation set.

[0086] (2) Input the speech feature data of the training set into the convolution module to obtain the feature vector, wherein the convolution module includes a convolutional layer. A batch normalization layer, a ReLU activation function, and a convolutional kernel size of [missing value]. Step size is The number of convolution kernels is 16.

[0087] (3) Input the feature vector into the multi-scale residual module to obtain deep features, specifically:

[0088] (3.1) The multi-scale residual module first goes through a... convolutional layers The feature map is evenly divided into 8 subsets according to the channel dimension. These 8 subsets are then used... express;

[0089] (3.2) Do nothing for the first subset and output the value. ;

[0090] remove In addition, each subset is processed Convolutional Filter The processing starts from the third subset, adding the output of the subset to the output of the previous subset, and then inputting the result into the convolutional filter. The output value is processed in the middle.

[0091] (3.3) Concatenate all output values ​​using the cat function, and input the concatenated output values ​​into the second... convolutional layers middle;

[0092] (3.4) The first convolutional layer Output With the second Convolutional layer Output Perform residual joins as follows:

[0093]

[0094] The above settings can effectively solve the gradient explosion problem caused by excessively deep network layers.

[0095] (4) such as Figure 2 As shown, the deep feature vector The input is fed into the contraction excitation unit to obtain a channel attention-enhanced feature vector. Specifically:

[0096] (4.1) Perform global average pooling on the output u of the multi-scale residual module, compress the entire spatial dimension along the channel dimension, and perform feature aggregation mapping to obtain... (C represents the number of channels, which is set to 1 in this embodiment) ) eigenvectors;

[0097] (4.2) Input the feature vector into the first fully connected layer, compress the C channels into C / R channels, where R is the compression ratio (set to 8 in this embodiment), and pass it through the ReLU activation function layer to obtain the channel-compressed feature vector;

[0098] (4.3) Input the channel compressed feature vector into the second fully connected layer to restore C / R channels back to R channels, and use the sigmoid function to obtain the attention score of the channel feature vector;

[0099] (4.4) Multiply the attention score of the channel feature vector with the output of the multi-scale residual module to obtain the output of the shrinking excitation unit, explicitly model the interdependence between channels, and adaptively recalibrate the channel feature response.

[0100] (5) such as Figure 3 As shown, deep feature vectors enhanced with channel attention. The input is fed into the attention pooling module to generate a global feature vector, specifically:

[0101] (5.1) Multiply the deep feature vector of each time frame output by the shrinking excitation unit by a linear transformation matrix. All time frames share the same parameters. Use the tanh activation function to obtain the current time step. Other time steps Attention score;

[0102] (5.2) The attention score is normalized and scaled using the softmax function. Obtain attention weights;

[0103] (5.3) Calculate the summation statistical characteristics of the weighted eigenvectors Statistical characteristics of standard deviation The data is concatenated and then linearly transformed through a fully connected layer to obtain the final global feature vector. The fully connected layer has 256 neurons. Different weights are given to the speech features of different time frames, so as to effectively focus on the speech forgery part.

[0104] (6) Use a single-class loss function for training, save the loss function value for each batch, set the batch size to 64, set the initial learning rate to 0.0003, decay by 50% every 10 epochs, and save the model parameters.

[0105] (7) Input the validation set data into the model for validation, adjust the model parameters according to the validation results, and obtain the trained residual attention network model.

[0106] The scheme described in this embodiment selects the model with the lowest error rate on the validation set for testing, achieving an error rate of 1.87% and a cascaded detection cost function index of 0.05.

[0107] Example 2:

[0108] The purpose of this embodiment is to provide a speech authentication system based on residual attention networks.

[0109] A speech authentication system based on residual attention networks includes:

[0110] The data acquisition unit is configured to acquire the audio data to be detected and perform corresponding preprocessing.

[0111] The feature extraction unit is configured to extract features from the preprocessed audio data and perform frame segmentation on the extracted speech feature data to obtain speech signal feature data of fixed frame length.

[0112] A feature enhancement unit is configured to obtain enhanced feature data based on the speech signal feature data using a pre-trained residual attention network model. The residual attention network model includes sequentially connected convolutional modules, a multi-scale residual module, a contraction activation unit, an attention pooling module, and a fully connected layer. The contraction activation unit receives the deep feature vector output by the multi-scale residual module and sequentially passes it through a global average pooling operation, a fully connected layer, a ReLU activation function, another fully connected layer, and a sigmoid function to obtain the attention score of the channel feature vector. The output of the contraction activation unit is obtained by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module.

[0113] The anti-spoofing unit is used to input the enhanced feature data into a pre-trained classifier to obtain the speech anti-spoofing results.

[0114] Furthermore, the system described in this embodiment corresponds to the method described in Embodiment 1, and its technical details have been described in detail in Embodiment 1, so they will not be repeated here.

[0115] In further embodiments, the following is also provided:

[0116] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0117] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0118] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0119] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0120] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0121] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0122] The above embodiments provide a speech authentication method and system based on residual attention networks, which can be implemented and has broad application prospects.

[0123] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A voice authentication method based on a residual attention network, characterized in that, The method comprises the following steps: acquiring audio data to be detected and performing corresponding preprocessing; extracting features from the preprocessed audio data, performing frame processing on the extracted speech feature data, and obtaining speech signal feature data with a fixed frame length; based on the speech signal feature data, using a pre-trained residual attention network model to obtain enhanced feature data; wherein the residual attention network model comprises sequentially connected convolution modules, multi-scale residual modules, a shrinkage excitation unit, an attention pooling module, and a fully connected layer; the shrinkage excitation unit receives a deep feature vector output by the multi-scale residual module, sequentially undergoes a global average pooling operation, a fully connected layer, a ReLU activation function, a fully connected layer, and a sigmod function to obtain an attention score of a channel feature vector, and obtains an output of the shrinkage excitation unit by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module; inputting the enhanced feature data into a pre-trained classifier to obtain a speech authentication result.

2. The voice authentication method based on the residual attention network according to claim 1, wherein, The multi-scale residual module receives a feature vector obtained through convolution processing; the feature vector is uniformly divided into several subsets according to the channel dimension through a convolution layer; the remaining subsets except the first subset are processed through a convolution filter, and from the third subset, the output of the convolution filter is added to the output of the previous subset and then input into the convolution filter to obtain the output values corresponding to each subset; the output values of each subset are spliced and input into a convolution layer to obtain a processing result; the processing result is connected in residual with the output of the convolution module to obtain a deep feature vector. 3.The voice authentication method based on the residual attention network according to claim 1, wherein, The processing flow of the shrinkage excitation unit specifically comprises: The output of the multi-scale residual module is subjected to a global average pooling operation, and the entire spatial dimension is compressed along the channel dimension to obtain a feature vector where C represents the number of channels. inputting the feature vector into a first fully connected layer to compress C channels into C / R channels, and passing through a ReLU activation function layer to obtain a channel-compressed feature vector, wherein R represents a compression ratio; inputting the channel-compressed feature vector into a second fully connected layer to restore C / R channels to C channels, and using a sigmod function to obtain an attention score of a channel feature vector; multiplying the attention score of the channel feature vector with the output of the multi-scale residual module to obtain the output of the shrinkage excitation unit.

4. The voice authentication method based on the residual attention network according to claim 1, wherein, The processing flow of the attention pooling module specifically comprises: multiplying each time frame deep feature vector output by the shrinkage excitation unit with a preset linear transformation matrix, all time frames sharing parameters, and using a tanh activation function to obtain an attention score; using a softmax function to normalize and scale the obtained attention score to [0, 1] to obtain an attention weight; concatenating the first-order and second-order statistical features of the weighted feature vector, and then performing linear transformation through a fully connected layer to obtain a final global feature vector.

5. The voice authentication method based on the residual attention network according to claim 1, wherein, The convolution module comprises sequentially connected convolution layers, batch normalization layers, and ReLU activation functions.

6. The voice authentication method based on the residual attention network according to claim 1, wherein, The preprocessed audio data is subjected to feature extraction, and the extracted speech feature data is subjected to frame processing, specifically: the preprocessed audio data is subjected to extraction of linear frequency cepstrum coefficient features with a preset sampling frequency, a preset window length, a preset number of FFT points, and a preset number of filters and their increments and double-increment coefficients, and the extracted speech feature data is subjected to frame processing to obtain speech signal feature data of a preset frame length.

7. The voice authentication method based on the residual attention network according to claim 1, wherein, The training of the residual attention network model is specifically as follows: Obtain speech feature data of a training set and a validation set; Input the training set speech feature data into a convolution module to obtain a feature vector; Input the feature vector into a multi-scale residual module to obtain a deep feature vector; Input the deep feature into a shrinkage excitation unit to obtain a channel attention enhanced feature vector; Input the channel attention enhanced feature vector into an attention pooling module to generate a global feature vector, and input the feature vector into a fully connected layer for classification; Use a single classification loss function to train and save model parameters; Input the validation set data into the model for validation, adjust the model parameters according to the validation result, and obtain the trained attention residual network model.

8. A residual attention network based voice authentication system characterized in that, It includes: A data acquisition unit configured to acquire audio data to be detected and perform corresponding preprocessing; A feature extraction unit configured to extract features from preprocessed audio data and perform frame processing on extracted speech feature data to obtain speech signal feature data of a fixed frame length; A feature enhancement unit configured to obtain enhanced feature data based on the speech signal feature data using a pre-trained residual attention network model; wherein the residual attention network model includes sequentially connected convolution modules, multi-scale residual modules, shrinkage excitation units, attention pooling modules, and fully connected layers, the shrinkage excitation unit receives a deep feature vector output by the multi-scale residual module, and sequentially undergoes a global average pooling operation, a fully connected layer, a ReLU activation function, a fully connected layer, and a sigmod function to obtain an attention score of a channel feature vector, and by multiplying the attention score of the channel feature vector with the output of the multi-scale residual module, the output of the shrinkage excitation unit is obtained; A pseudo-detection unit for inputting the enhanced feature data into a pre-trained classifier to obtain a speech pseudo-detection result.

9. A computer-readable storage medium having stored thereon a program, characterized in that, The program is executed by the processor to implement a residual attention network-based speech pseudo-detection method according to any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement a residual attention network-based speech pseudo-detection method according to any one of claims 1-7.

Images