Voice data processing method based on traditional Chinese medicine sound diagnosis symptom feature privacy protection
Through the combination of generator and discriminator networks, voice data features are extracted and converted, and the problem of health attribute exposure in voice data is solved, the privacy protection of voice data is realized, and the user's health information is not recognized by the traditional Chinese medicine voice diagnosis system.
Patent Information
- Application Number
- CN202411340010.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2025-01-10
AI Technical Summary
The prior art fails to effectively protect the health attributes in the user's voice data, resulting in the exposure of user health information to risks, especially when sharing voice data on social networks.
Using the speech data processing method based on the privacy protection of the characteristics of Chinese medicine sound diagnosis symptoms, the characteristics in the speech data are extracted and converted through the combination of the generator network and the discriminator network, and healthy speech data that cannot be inferred from the Chinese medicine sound diagnosis system are generated. .
While retaining other attributes of voice data, users' privacy information is successfully protected to ensure that the voice data shared by users cannot determine the user's health syndrome elements based on the traditional Chinese medicine audio diagnosis theory.
Smart Images
Figure CN119296577A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and in particular to a voice data processing method based on privacy protection of traditional Chinese medicine voice diagnosis symptom characteristics. Background Art
[0002] At present, with the development of TCM voice diagnosis technology, by extracting characteristic information from voice signals and comparing and matching the characteristic information of the pre-diagnosis case with the characteristic information of the pre-stored case library, a diagnosis report of the pre-diagnosis case can be output; while this TCM voice diagnosis technology provides medical assistance, it also causes that when the voice shared by users on social networks is maliciously collected, the health attributes of users can be obtained through TCM voice diagnosis methods, thereby exposing the health attributes of users to risks.
[0003] However, current research on speech conversion technology is mostly focused on processing the emotional attributes and voiceprint attributes of speech, and there is no research on the risk of health attributes being exposed in traditional Chinese medicine voice diagnosis. Therefore, how to convert the health attributes in speech data to protect the patient's sensitive information while ensuring the integrity of the speaker's speech data is an urgent problem to be solved in the field of speech privacy protection. Summary of the invention
[0004] In view of the above analysis, the present invention aims to provide a voice data processing method based on the privacy protection of TCM voice diagnosis symptom characteristics, so as to protect the user's voice data from inferring health attributes through the TCM voice diagnosis system without affecting the normal use of the voice data.
[0005] The purpose of the present invention is mainly achieved through the following technical solutions:
[0006] The present invention discloses a voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics, comprising:
[0007] Step S1, acquiring user voice data as training corpus from an acoustic device; the user voice data includes voice data to be converted that can obtain various symptom characteristics through traditional Chinese medicine acoustic diagnosis and target voice data under healthy conditions that cannot obtain symptom characteristics;
[0008] Step S2: extract the FO s ,AP s and 40-dimensional MCEP s Features, and FO in the target speech data p and 40-dimensional MCEP p feature;
[0009] Step S3: FO of the voice data to be transferred s , MCEP sFeatures and FO of target speech data p , MCEP p The features are input into the privacy-protected speech model of TCM acoustic diagnosis symptom features for network training; the trained speech model is used to input the FO of the speech data to be converted containing symptom features. s and MCEP s Feature conversion to FO of healthy speech * and MCEP * feature;
[0010] Step S4: FO of healthy voice * and MCEP * Features and FO in the voice data to be converted s The user's voice is synthesized to obtain privacy-protected features of TCM acoustic diagnosis symptoms.
[0011] Furthermore, when the user voice data used as training corpus is classified after TCM voice diagnosis, each voice data to be converted and target voice data are also labeled with data categories; among them, the voice data to be converted is labeled as "S" and the target voice data is labeled as "P".
[0012] Furthermore, the FO in the voice data to be converted is performed by the WORD vocoder. s ,AP s and 40-dimensional MCEP s Features, and FO in the target speech data p and 40-dimensional MCEP p Feature extraction.
[0013] Furthermore, the privacy-preserving speech model for TCM acoustic diagnosis symptom features includes a generator network and a discriminator network; wherein,
[0014] The input data of the generator network is the FO of the speech data to be converted s , MCEP s Features and FO of target speech data p , MCEP p The feature vector composed of features; the generator's FO for the speech data to be converted s , MCEP s After feature dimension reduction and attention mechanism, the healthy independent representation Z of the speech to be converted is obtained; for the FO of the target speech data p , MCEP p After feature dimension reduction and attention mechanism, the healthy embedding representation C of the target speech is obtained; and the healthy independent representation Z of the speech to be converted and the healthy embedding representation C of the target speech are deeply fused to obtain the FO of the healthy speech. * and MCEP *feature;
[0015] The input of the discriminator network is the FO of healthy speech * and MCEP * Features; FO of the discriminator network on the input healthy speech * and MCEP * The characteristics of TCM voice diagnosis are identified by the features. When it is identified as "S", the parameters of the generator network are iteratively updated. When it is identified as "P", the generator network is stopped from being updated and the FO of the healthy voice output by the generator network is replaced by the FO of the healthy voice output by the generator network. * and MCEP * Feature output.
[0016] Furthermore, the generator network includes an encoder network, an attention mechanism module and a decoder network connected in sequence; wherein,
[0017] The encoder network and the attention mechanism module are connected sequentially to convert the input speech data into FO s , MCEP s The feature dimension reduction is performed to obtain the healthy independent representation z of the speech to be converted, and the FO of the input target speech data is converted p , MCEP p The features are reduced in dimension to obtain a healthy embedding representation C of the target speech;
[0018] The decoder network performs a deep feature fusion concat operation on the healthy independent representation Z of the input speech to be converted and the healthy embedded representation C of the target speech, and outputs the FO* and MCEP* features of the healthy speech that are finally fused into a new language feature vector.
[0019] Furthermore, the encoder network is composed of a multi-scale convolutional neural network, which further reduces the dimension of speech features; an attention mechanism is added to make the speech features more focused on the category of speech data during the encoding process.
[0020] Furthermore, the encoder network includes 9 Conv layers and corresponding Bathnorm layers and 2 Linear layers; the Conv layer includes three categories, among which the first Conv layer and the second Conv layer are Conv5-512; the third to the eighth Conv layers are Conv3-64; and the ninth Conv layer is Conv1-512; the corresponding Bathnorm layers are also three categories, among which the first Bathnorm layer and the second Bathnorm layer are Bathnorm1d-512; the third to the eighth Bathnorm layers are Bathnorm1d-64; and the ninth Bathnorm layer is Bathnorm1d-512; the Linear layer includes the first Linear layer Linear-256 and the second Linear layer Linear-512.
[0021] Furthermore, the total framework loss function of the privacy-preserving speech model of TCM acoustic diagnosis symptom features is:
[0022] L TCM =λ rec L rec +λ C L (C)
[0023] Among them, λ rec , C is a hyperparameter, L rec is the reconstruction loss of the generator network, L (C) is the loss function of the encoder network.
[0024] Furthermore, the loss function L of the encoder network is (C) for:
[0025]
[0026] Among them, x is the input feature vector, z is the healthy independent representation of the speech to be converted output by the encoder network; y i，j To indicate whether the true label belongs to the first j categories of the i-th sample, p i，j represents the probability that the model predicts that the i-th sample belongs to the j-th category, i = 1, ..., n, j = 1, ..., k, n represents the number of samples, and k is the number of categories;
[0027] The reconstruction loss L of the generator network rec for:
[0028]
[0029] Among them, E θ (x) and E e(x) refers to the network output of the encoder network encoding the features of category “S” and “P” in the feature vector x; D(·, ·) is the distance function; is the expected value function, and ||·||1 is the L1 norm.
[0030] Further, the discriminator network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first Bathnorm layer, a Linear layer, and a second Bathnorm layer connected sequentially;
[0031] Among them, the first, second and third convolutional layers are Conv1-1536, Conv1-128, Conv1-1536 respectively; the first Bathnorm layer is Bathnorm1d-3072; the Linear layer is Linear-192; and the second Bathnorm layer is Bathnorm1d-3072.
[0032] The beneficial effects of the present invention are as follows:
[0033] The voice data processing method based on privacy protection of TCM voice diagnosis symptom characteristics of the present invention successfully protects the privacy information of the patient's voice data after it is released while also retaining other attributes of the voice data. The final effect is that the user shares his own voice data, and the original attributes are maintained in terms of hearing and content, but the corresponding symptom elements of the user cannot be judged based on the theoretical knowledge of TCM voice diagnosis, thereby protecting the privacy of the user. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.
[0035] Figure 1 This is a flow chart of a method for processing voice data based on privacy protection of TCM acoustic diagnosis symptom characteristics according to an embodiment of the present invention;
[0036] Figure 2 Schematic diagram of the encoder network structure of an embodiment of the present invention;
[0037] Figure 3 A schematic diagram of a discriminator network structure according to an embodiment of the present invention;
[0038] Figure 4 This is a schematic diagram of the privacy protection speech model training process for TCM acoustic diagnosis symptom features according to an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.
[0040] This embodiment discloses a voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics. Figure 1 As shown, including:
[0041] Step S1, acquiring user voice data as training corpus from an acoustic device; the user voice data includes voice data to be converted that can obtain various symptom characteristics through traditional Chinese medicine acoustic diagnosis and target voice data under healthy conditions that cannot obtain symptom characteristics;
[0042] Step S2: extract the FO s ,AP s and 40-dimensional MCEP s Features, and FO in the target speech data p and 40-dimensional MCEP p feature;
[0043] Step S3: FO of the voice data to be transferred s , MCEP s Features and FO of target speech data p , MCEP p The features are input into the privacy-protected speech model of TCM acoustic diagnosis symptom features for network training; the trained speech model is used to input the FO of the speech data to be converted containing symptom features. s and MCEP s Feature conversion to FO of healthy speech * and MCEP * feature;
[0044] Step S4: FO of healthy voice * and MCEP * Features and FO in the voice data to be converted s The user's voice is synthesized to obtain privacy-protected features of TCM acoustic diagnosis symptoms.
[0045] Specifically, in step S1, when the user voice data used as training corpus is classified after TCM voice diagnosis, each voice data to be converted and target voice data are also labeled with data categories; wherein the voice data to be converted is labeled as "S" and the target voice data is labeled as "P";
[0046] Among them, the method of TCM voice diagnosis can adopt the existing objective means of voice diagnosis method, which can not only achieve the correct classification of corpus data, but also avoid the wrong classification caused by subjective consciousness. Specific objective means of TCM voice diagnosis method can be found in existing public documents, and the specific method is not limited here.
[0047] Specifically, in step S2, the FOs ,AP s and 40-dimensional MCEP s Features, and FO in the target speech data p and 40-dimensional MCEP p Features; AP s It contains the spoken content, so it is the final speech synthesis task; the FO of the speech data to be converted s , MCEP s Features and FO of target speech data p , MCEP p The features are used as training data for model training in step S3.
[0048] Specifically, feature extraction can be performed using a WORD vocoder.
[0049] FO feature is the fundamental frequency feature; the fundamental frequency determines the rhythm and structure information of the speech, corresponding to the periodic pulse sequence generated by the vibration of the vocal cords. The WORLD vocoder uses the DIO algorithm to extract the fundamental frequency. The algorithm filters the original signal through a low-pass filter and calculates the standard deviation of the possible fundamental frequency period. Finally, the inverse of the period with the smallest standard deviation is selected as the fundamental frequency.
[0050] AP features are non-periodic signal parameter features; non-periodic signal parameters determine the unvoiced part of speech, corresponding to the non-periodic pulse sequence in the mixed excitation. The WORLD vocoder uses the D4C algorithm to extract non-periodic signals, which estimates non-periodic signals based on fundamental frequency, spectral envelope and waveform;
[0051] The MCEP feature is the Mel-Cepstral coefficient feature; the Mel-Cepstral coefficient is extracted through the Mel scale and cepstrum analysis, which can reflect the energy distribution of the signal at different Mel frequencies.
[0052] Specifically, the Traditional Chinese Medicine Generative adversarial networks (TCMGAN) in step S3 includes a generator network and a discriminator network;
[0053] Among them, the input data of the generator network is the FO of the speech data to be converted s , MCEP s Features and FO of target speech data p , MCEP p The feature vector composed of features; the generator's FO for the speech data to be converted s , MCEP sAfter feature dimension reduction and attention mechanism, the healthy independent representation Z of the speech to be converted is obtained; for the FO of the target speech data p , MCEP p After feature dimension reduction and attention mechanism, the healthy embedding representation C of the target speech is obtained; and the healthy independent representation Z of the speech to be converted and the healthy embedding representation C of the target speech are deeply fused to obtain the FO of the healthy speech. * and MCEP * feature;
[0054] The input of the discriminator network is the FO of healthy speech * and MCEP * Features; FO of the discriminator network on the input healthy speech * and MCEP * The characteristics of TCM voice diagnosis are identified by the features. When it is identified as "S", the parameters of the generator network are iteratively updated. When it is identified as "P", the generator network is stopped from being updated and the FO of the healthy voice output by the generator network is replaced by the FO of the healthy voice output by the generator network. * and MCEP * Feature output.
[0055] Specifically, the generator network includes an encoder network, an attention mechanism module, and a decoder network connected sequentially;
[0056] The encoder network and the attention mechanism module are connected sequentially to convert the input speech data into FO s , MCEP s The feature dimension reduction is performed to obtain the healthy independent representation z of the speech to be converted, and the FO of the input target speech data is converted p , MCEP p The features are reduced in dimension to obtain a healthy embedding representation C of the target speech;
[0057] The encoder network is composed of a multi-scale convolutional neural network, which further reduces the dimension of speech features. The attention mechanism is added to make the speech features more focused on the category of speech data during the encoding process.
[0058] Specifically, Figure 2As shown, the encoder network includes 9 Conv layers and corresponding Bathnorm layers and 2 Linear layers; the Conv layer includes three categories, among which the first Conv layer and the second Conv layer are Conv5-512, 5 is the size of the convolution kernel, and 512 is the number of convolution kernels; the third to the eighth Conv layers are Conv3-64, 3 is the size of the convolution kernel, 64 is the number of convolution kernels, and the ninth Conv layer is Conv1-512; the corresponding Bathnorm layers are also three categories, among which the first Bathnorm layer and the second Bathnorm layer are Bathnorm1d-512; the third to the eighth Bathnorm layers are Bathnorm1d-64; the ninth Bathnorm layer is Bathnorm1d-512; the Linear layer includes the first Linear layer Linear-256 and the second Linear layer Linear-512;
[0059] The loss function of the encoder network is L (C) :
[0060]
[0061] Among them, x is the input feature vector, z is the healthy independent representation of the speech to be converted output by the encoder network; y i，j To indicate whether the true label belongs to the first j categories of the i-th sample, p i，j represents the probability that the model predicts that the i-th sample belongs to the j-th category, i = 1, ..., n, j = 1, ..., k, n represents the number of samples, and k is the number of categories;
[0062] The activation function used is the Sigmoid function:
[0063]
[0064] The x in the activation function is the potential representation of the model input.
[0065] The decoder network performs deep feature fusion concat operation on the healthy independent representation Z of the input speech to be converted and the healthy embedded representation C of the target speech, and outputs the FO of the healthy speech that is finally fused into a new language feature vector * and MCEP * feature;
[0066] The reconstruction loss L of the generator network in this example rec for:
[0067]
[0068] Among them, E θ (x) and E e(x) refers to the network output of the encoder network encoding the features of category “S” and “P” in the feature vector x; D(·, ·) is the distance function; is the expected value function, and ||·||1 is the L1 norm.
[0069] The loss function of the overall framework is:
[0070] L TCM =λ rec L rec +λ C L (C) ;
[0071] Among them, λ rec , C For the hyperparameter, choose a suitable value λ rec =10,λ C =1 not only ensures the sound quality of the converted voice, but also ensures that the user's true TCM characteristics are not leaked.
[0072] Specifically, the purpose of the discriminator network is to determine whether the speech conversion is successful, that is, whether the TCM voice diagnosis symptom characteristics after speech conversion can still be identified as the "S" category; if the discriminator judges that the category is still "S", it will continue to iteratively update the encoder encoder until the category is judged to be "P".
[0073] This embodiment here utilizes the characteristics of the generative adversarial network, that is, the adversarial relationship between the generator and the discriminator, and combines the four network structures of encoder-attention-decoder-discriminator to form a generative adversarial network for privacy protection of TCM voice diagnosis symptom features.
[0074] More specifically, the discriminator network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first Bathnorm layer, a Linear layer, and a second Bathnorm layer connected sequentially;
[0075] Among them, the first, second and third convolutional layers are Conv1-1536, Conv1-128, Conv1-1536 respectively; the first Bathnorm layer is Bathnorm1d-3072; the Linear layer is Linear-192; the second Bathnorm layer is Bathnorm1d-3072;
[0076] By training the discriminator network, the discriminator network can classify and output classification results based on the input FO and MCEP features. The classification results are "S" class that can obtain various symptom characteristics through TCM voice diagnosis and "P" class that cannot obtain symptom characteristics.
[0077] In step S4, the FO of the healthy voice after successfully hiding the TCM voice diagnosis symptom features in the voice data is * and MCEP * Features and FO in the voice data to be converted s The voice is synthesized in a vocoder to obtain the user's voice with privacy protection for the characteristics of TCM voice diagnosis symptoms.
[0078] The vocoder vocoder is a W0RD vocoder.
[0079] like Figure 4 As shown, it is a schematic diagram of the training process of the privacy-protected speech model of TCM acoustic diagnosis symptom features according to an embodiment of the present invention.
[0080] In summary, the voice data processing method based on privacy protection of TCM voice diagnosis symptom characteristics in the embodiment of the present invention successfully protects the privacy information of the patient's voice data after it is released while also retaining other attributes of the voice data. The final effect is that the user shares his own voice data, which maintains the original attributes in terms of hearing and content, but the corresponding symptom elements of the user cannot be judged based on the theoretical knowledge of TCM voice diagnosis, thereby protecting the user's privacy.
[0081] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics, characterized in that: include: Step S1, acquiring user voice data as training corpus from an acoustic device; the user voice data includes voice data to be converted that can obtain various symptom characteristics through traditional Chinese medicine acoustic diagnosis and target voice data under healthy conditions that cannot obtain symptom characteristics; Step S2: extract F0 from the voice data to be converted s ,AP s and 40-dimensional MCEP s Features, and F0 in the target speech data p and 40-dimensional MCEP p feature; Step S3: F0 of the voice data to be transferred s , MCEP s Features and F0 of target speech data p , MCEP p The features are input into the privacy-protected speech model of TCM acoustic diagnosis symptom features for network training; the trained speech model is made to input the F0 of the speech data to be converted containing symptom features s and MCEP s Feature conversion to F0 of healthy speech * and MCEP * feature; Step S4: F0 of the healthy voice * and MCEP * Features and F0 in the speech data to be converted s The user's voice is synthesized to obtain privacy-protected features of TCM acoustic diagnosis symptoms.
2. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 1 is characterized in that: After the user voice data used as training corpus passes through TCM voice diagnosis, when the corpus data is classified, the data category of each voice data to be converted and the target voice data is also labeled; among them, the voice data to be converted is labeled as "S" and the target voice data is labeled as "P".
3. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 2 is characterized in that: Use WORD vocoder to convert F0 in the voice data to be converted s 、AP s and 40-dimensional MCEP s Features, and F0 in the target speech data p and 40-dimensional MCEP p Feature extraction.
4. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 3 is characterized in that: The privacy-preserving speech model for TCM acoustic diagnosis syndrome features includes a generator network and a discriminator network; The input data of the generator network is the F0 of the speech data to be converted s , MCEP s Features and F0 of target speech data p , MCEP p The feature vector composed of features; the F0 of the generator for the speech data to be converted s , MCEP s After feature dimension reduction and attention mechanism, the healthy independent representation Z of the speech to be converted is obtained; for the F0 of the target speech data p , MCEP p After feature dimension reduction and attention mechanism, the healthy embedding representation C of the target speech is obtained; and the healthy independent representation Z of the speech to be converted and the healthy embedding representation C of the target speech are deeply fused to obtain the F0 of the healthy speech. * and MCEP * feature; The input of the discriminator network is the F0 of healthy speech * and MCEP * Features; F0 of the discriminator network for the input healthy speech * and MCEP * The characteristics of TCM voice diagnosis are used to identify the symptoms. When it is identified as "S", the parameters of the generator network are iteratively updated. When it is identified as "P", the generator network is stopped from being updated and the F0 of the healthy speech output by the generator network is * and MCEP * Feature output.
5. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 3 is characterized in that: The generator network includes a sequentially connected encoder network, an attention mechanism module, and a decoder network; wherein, The encoder network and the attention mechanism module are connected sequentially to convert the F0 of the input speech data to be converted. s , MCEP s The feature dimension reduction is performed to obtain the healthy independent representation z of the speech to be converted, and the F0 of the input target speech data is converted p , MCEP p The features are reduced in dimension to obtain a healthy embedding representation C of the target speech; The decoder network performs deep feature fusion concat operation on the healthy independent representation Z of the input speech to be converted and the healthy embedded representation C of the target speech, and outputs the F0 of the healthy speech that is finally fused into a new language feature vector * and MCEP * feature.
6. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 5 is characterized in that: The encoder network is composed of a multi-scale convolutional neural network, which further reduces the dimension of speech features. The attention mechanism is added to make the speech features more focused on the category of speech data during the encoding process.
7. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 6 is characterized in that: The encoder network includes 9 Conv layers and corresponding Bathnorm layers and 2 Linear layers; the Conv layers include three categories, among which the first Conv layer and the second Conv layer are Conv5-512; the third to the eighth Conv layers are Conv3-64; the ninth Conv layer is Conv1-512; the corresponding Bathnorm layers are also three categories, among which the first Bathnorm layer and the second Bathnorm layer are Bathnorm1d-512; the third to the eighth Bathnorm layers are Bathnorm1d-64; the ninth Bathnorm layer is Bathnorm1d-512; the Linear layer includes the first Linear layer Linear-256 and the second Linear layer Linear-512.
8. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 6 is characterized in that: The total framework loss function of the privacy-preserving speech model for TCM acoustic diagnosis symptom features is: L TCM =λ rec L rec +λ C L (C) Among them, λ rec , C is a hyperparameter, L rec is the reconstruction loss of the generator network, L (C) is the loss function of the encoder network.
9. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 8 is characterized in that: The loss function L of the encoder network (C) for: Among them, x is the input feature vector, z is the healthy independent representation of the speech to be converted output by the encoder network; y i,j To indicate whether the true label belongs to the first j categories of the i-th sample, p i,j represents the probability that the model predicts that the i-th sample belongs to the j-th category, i = 1, ..., n, j = 1, ..., k, n represents the number of samples, and k is the number of categories; The reconstruction loss L of the generator network rec for: Among them, E θ (x) and E e (x) refers to the network output of the encoder network encoding the features of category "S" and "P" in the feature vector x; D(·,·) is the distance function; is the expected value function, and ||·||1 is the L1 norm.
10. The voice data processing method based on privacy protection of TCM acoustic diagnosis symptom characteristics according to claim 4 is characterized in that: The discriminator network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first Bathnorm layer, a Linear layer, and a second Bathnorm layer connected sequentially; Among them, the first, second and third convolutional layers are Conv1-1536, Conv1-128, Conv1-1536 respectively; the first Bathnorm layer is Bathnorm1d-3072; the Linear layer is Linear-192; and the second Bathnorm layer is Bathnorm1d-3072.