Generative adversarial network-based voice abnormal state detection method for language programs
Through the method based on the generative adversarial network, the adversarial training of generator and discriminator is used to solve the limitations of traditional autoencoder in abnormal sound detection, the accuracy and efficiency of detection are improved, and the implementation of the in many sound environments is achieved. The ability to accurately identify and distinguish abnormal sounds.
Patent Information
- Application Number
- CN202410970955.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-26
AI Technical Summary
Traditional autoencoders are difficult to effectively distinguish between normal sound and abnormal sound in abnormal sound detection, especially in terms of nuanced differences in sound samples, and cannot fully capture the inherent distribution characteristics of sound, resulting in inaccuracy and inefficient detection.
Using a method based on generative adversarial network, a generative adversarial network model is constructed through adversarial training of generators and discriminators. Multiple iterative training deeply mines and learns the inherent distribution characteristics of normal data samples to generate highly natural and realistic data. Sample, distinguish and identify abnormal sounds.
Improve the accuracy and efficiency of abnormal sound detection. Through the adversarial training of the generative adversarial network model, abnormal sound can be accurately identified and distinguished in various sound environments, achieving efficient and accurate detection, avoiding complex sampling and inference in traditional models. process.
Smart Images

Figure CN119028376A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of sound detection, and specifically to a method for detecting human voice anomalies in language programs based on a generative adversarial network. Background Art
[0002] With the rapid development of the radio and television industry, language programs such as news broadcasts, talk shows, educational lectures, etc. are diverse in form, and they have extremely high requirements for sound quality. The radio and television industry not only pursues the richness and professionalism of content, but also pays attention to the clarity and purity of sound to ensure the accurate transmission of information and the good experience of the audience, especially the distinction between normal and abnormal sounds. When there is a significant similarity between normal and abnormal sounds, traditional autoencoders often have difficulty working effectively because they rely on reconstruction errors for distinction. The limitation of this technology is that it cannot fully capture the intrinsic distribution characteristics of sound samples, especially in terms of subtle differences, which is crucial for accurately distinguishing normal from abnormal sounds. Summary of the invention
[0003] The purpose of the present invention is to provide a method for detecting abnormal human voice in language programs based on a generative adversarial network. The present invention utilizes the characteristics of high freedom of design, high training efficiency, and high sample generation efficiency of the generative adversarial network model to effectively solve the limitations of traditional autoencoders in abnormal sound detection and improve the accuracy and efficiency of detection.
[0004] The technical solution of the present invention is a method for detecting abnormalities in human voice in language programs based on a generative adversarial network, which specifically comprises the following steps:
[0005] Step 1: Obtain audio samples of human voices in language programs, and perform data preprocessing on the audio samples to obtain normal data samples and abnormal data samples;
[0006] Step 2: Extract features from normal data samples to obtain a training set, mix abnormal data samples into normal data samples and extract features to obtain a test set;
[0007] Step 3: construct a generative adversarial network model, and use a training set and a test set to train and test the generative adversarial network model respectively; wherein the generative adversarial network model includes a generator and a discriminator, the generator is used to encode and decode the original data after data preprocessing to obtain reconstructed data, and the discriminator encodes the reconstructed data using the same encoding method as the generator to distinguish between real normal data samples and abnormal data samples;
[0008] Step 4: Use the trained generative adversarial network model to detect anomalies in human voices in language programs.
[0009] In the aforementioned method for detecting human voice anomalies in language programs based on a generative adversarial network, in step one, the data preprocessing includes Fourier transform and normalization processing; the Fourier transform is to convert audio samples from time domain data to frequency domain data; the normalization processing is to calculate the maximum and minimum values of the amplitude spectrum in the frequency domain data, and then apply a logarithmic transform to each amplitude value in the amplitude spectrum to compress its dynamic range, and finally perform linear normalization to map the logarithmic transformed value to a specific range.
[0010] In the aforementioned method for detecting human voice anomalies in language programs based on a generative adversarial network, the formula for applying logarithmic transformation is:
[0011] log_F i (ω)=log_(F i (ω) / F min (ω));
[0012] Where: F i (ω) is the corresponding amplitude value in the amplitude spectrum; F min (ω) is the minimum amplitude value in the amplitude spectrum.
[0013] The linear normalization formula is:
[0014] F n_val =(log_F i (ω)-min[log_F i (ω)]) / (max[log_F i (ω))]-min[log_F i (ω)]);
[0015] Among them, min[log_F i (ω)] and max[log_F i (ω)] are the minimum and maximum values of the amplitude spectrum after logarithmic transformation.
[0016] In the aforementioned method for detecting human voice anomalies in language programs based on a generative adversarial network, in step 2, the feature extraction includes the extraction of low-level features and high-level features; the low-level features include the spectrum and energy of the audio sample; the high-level features include the Mel-frequency cepstral coefficients and pitch of the audio sample.
[0017] In the aforementioned method for detecting human voice anomalies in language programs based on a generative adversarial network, in step three, the generator includes an encoder and a decoder; the encoder includes 4 convolutional layers, wherein the first layer contains 64 convolutional kernels, the second layer contains 128 convolutional kernels, the third layer contains 256 convolutional kernels, and the fourth layer contains 512 convolutional kernels, and an activation function is used after each convolutional layer to increase nonlinearity; the decoder includes an upsampling layer and an output layer, the upsampling layer uses a deconvolution operation to restore the low-dimensional vector to the high-dimensional space of the original data, and an activation function is used after each deconvolution layer; the output layer is used to convert the output of the decoder into a format that is the same or similar to the original data.
[0018] In the aforementioned method for detecting human voice anomalies in language programs based on a generative adversarial network, a cross entropy loss function is used to minimize the difference between the reconstructed data and the original data; the cross entropy loss function is defined as:
[0019]
[0020] Where N is the number of audio samples, y i is the true label of the i-th audio sample; is the model's predicted probability for the i-th audio sample.
[0021] In the aforementioned method for detecting human voice anomalies in language programs based on generative adversarial networks, in step 3, based on the low-level features and high-level features of the training set, the corresponding sub-loss is calculated, the feature extraction of the original data and the feature extraction of the reconstructed data are compared, and the similarity between them is calculated; the sub-loss is divided into a time domain similarity loss and a frequency domain similarity loss, the time domain similarity loss is used to measure the similarity between the original data and the reconstructed data in the time domain, and the loss is represented by t1 or t2; the frequency domain similarity loss is used to measure the similarity between the inverse of the spectral feature MFCC of the original data and the inverse of the spectral feature MFCC of the reconstructed data, and the loss is represented by m1 or m2; according to each sub-loss, the target loss corresponding to the low-level features and the high-level features is calculated in a weighted manner; the goal of the discriminator D is to maximize the following objective function, and the objective function of the discriminator D is expressed as:
[0022] D1maxV(D,G)=E t1~pdata(t1) [log(D(t1))]+E 2~p(t2) [log(1-D(G(t2)))];
[0023] D2max V(D,G)=E m1~pdata(m1) [log(D(m1))]+E m2~p(m2) [log(1-D(G(m2)))];
[0024] Among them, t1 and m1 are audio samples sampled from the real data distribution pdata(x), t2 and m2 are random noise vectors sampled from the Gaussian distribution p(x); G(x) is the sample generated by the generator G based on the reconstructed data feature extraction; the first term E t1~pdata(t1) [log(D(t1))] encourages D to output high probability for the real sample t1; the second term E t2~p(t2) [log(1-D(G(t2)))] encourages D to output low probability for the generated sample G(t2);
[0025] The goal of the generator G is to minimize the following objective function, which is expressed as:
[0026] G1min V(D, G)=-E t2~p(t2) [log(D(G(t2)))];
[0027] G2min V(D, G)=-E m2~p(m2) [log(D(G(m2)))].
[0028] In the aforementioned method for detecting human voice anomalies in language programs based on generative adversarial networks, in step 4, the trained generator and discriminator are used to process the test set. By comparing the output of the discriminator with the true label of the test set, the detection performance of the training model for abnormal data samples is evaluated to meet the requirements of the State Administration of Radio, Film and Television for the distortion of language programs.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] The present invention deeply mines and learns the inherent distribution characteristics of normal data samples through multiple iterative training of the generative adversarial network, so that the network can accurately capture the subtle differences in normal sound patterns. In the testing phase of the present invention, the generative adversarial network model can generate data samples with high naturalness and fidelity through adversarial training of the generator and the discriminator. The generator strives to generate samples that are real enough to deceive the discriminator, while the discriminator attempts to distinguish between real samples and generated samples. This adversarial mechanism prompts the generator to continuously improve its generation ability, so that it can generate data that is very close to the real world. The normal data samples of the present invention can be reconstructed by the network almost without error, showing a high degree of fidelity, and the training process is relatively simple and efficient. Through the adversarial training of the generator and the discriminator, the generative adversarial network model can quickly learn the distribution characteristics of the data and generate corresponding data samples. This training method avoids the complex sampling and inference process in the traditional generative model, thereby improving the training efficiency. The abnormal data samples in the present invention do not conform to the learned normal patterns, and will show obvious deviations or distortions during the reconstruction process, especially in certain frequency bands. This significant difference in reconstruction effect provides a reliable basis for the model to distinguish and identify abnormal sounds, thereby achieving efficient and accurate abnormal sound detection. The present invention effectively solves the limitations of traditional autoencoders in abnormal sound detection and improves the accuracy and efficiency of detection. The present invention can accurately identify and distinguish abnormal sounds in a variety of sound environments by accurately learning and reconstructing the patterns of data. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 is a flow chart of the method of the present invention;
[0032] Figure 2 is the training set step diagram;
[0033] Figure 3 This is the test set step diagram. DETAILED DESCRIPTION
[0034] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.
[0035] Embodiment: A method for detecting human voice anomalies in language programs based on a generative adversarial network, such as Figure 1-3 As shown, the specific steps are as follows:
[0036] Step 1: Obtain audio samples of human voices in language programs, and perform data preprocessing on the audio samples to obtain normal data samples and abnormal data samples;
[0037] In this embodiment, the N audio samples are first subjected to Fourier transformation to convert them from the time domain f(t) to the frequency domain F(ω). Then, normalization is performed to convert the Fn Extract the amplitude spectrum F from (ω) n (ω), calculate the amplitude of each frequency component. Normalize the Fourier transformed data so that all audio samples are compared on the same scale. Find the minimum value F in the amplitude spectrum min (ω) and the maximum value F max (ω), for each amplitude value F in the amplitude spectrum i (ω)(i=1,2,3·····n) applies logarithmic transformation to compress its dynamic range. The logarithmic normalization formula is:
[0038] log_F i (ω)=log_(F i (ω) / F min (ω));
[0039] The logarithmic transformed values are further mapped to a specific range, so linear normalization is performed. The formula for linear normalization is:
[0040] F n_val =(log_F i (ω)-min[log_F i (ω)]) / (max[log_F i (ω))]-min[log_F i (ω)]);
[0041] Among them, min[log_F i (ω)] and max[log_F i (ω)] are the minimum and maximum values of the amplitude spectrum after logarithmic transformation.
[0042] From the normalized data, normal data samples (sound frequency is 82-1.1KHz) and abnormal data samples (sound frequency is 20-82Hz, 1.1-20KHz) are distinguished. Reference range: male host bass to treble range: roughly between 82Hz (bass) and 698Hz (treble). Female host bass to treble range: roughly between 82Hz (bass) and 1.1KHz (treble). Under normal circumstances, the sound frequency range that the human ear can hear is between 20Hz and 20KHz. This is the normal working range of the human ear hearing system, and it is most sensitive to sounds from 1KHz to 3KHz.
[0043] Step 2: Extract features from normal data samples to obtain a training set, mix abnormal data samples into normal data samples and extract features to obtain a test set;
[0044] In this embodiment, feature extraction is performed on audio samples at different levels of audio data. The first feature extraction extracts low-level features such as spectrum and energy of audio samples, while the second feature extraction extracts high-level features such as Mel-frequency cepstral coefficients (MFCC) and pitch of audio samples. The normal data samples are subjected to the first feature and the second feature extraction process to obtain a feature set of normal data samples, which is used as training data for the generator and the discriminator. The abnormal data samples are mixed into the normal data samples, and feature extraction including the first feature and the second feature is performed together to obtain a test set including normal data samples and abnormal data samples.
[0045] Step 3: construct a generative adversarial network model, and use a training set and a test set to train and test the generative adversarial network model respectively; wherein the generative adversarial network model includes a generator and a discriminator, the generator is used to encode and decode the original data after data preprocessing to obtain reconstructed data, and the discriminator encodes the reconstructed data using the same encoding method as the generator to distinguish between real normal data samples and abnormal data samples;
[0046] In this embodiment, the generative adversarial network model consists of a generator and a discriminator. The generator includes an encoder and a decoder, and a convolutional autoencoder is used to encode and decode the training set of normal data samples. The encoder encodes the input data into a low-dimensional vector (i.e., a human voice vector) and performs feature extraction. The encoder includes 4 convolutional layers to extract features from the sample data. The first layer contains 64 convolution kernels, the second layer contains 128 convolution kernels, the third layer contains 256 convolution kernels, and the fourth layer contains 512 convolution kernels, which are used to capture features of different frequencies and time scales, and an activation function is used after each convolution layer to increase nonlinearity.
[0047] The decoder reconstructs the original data from the low-dimensional vector. The decoder includes an upsampling layer and an output layer. The upsampling layer uses a deconvolution operation to restore the low-dimensional vector to the high-dimensional space of the original data. Similarly, an activation function is used after each deconvolution layer. The output layer is used to convert the output of the decoder into a format that is the same or similar to the original data.
[0048] Define the input data set X, expressed as X = {x1, x2, ..., x n}, where x i is a sample of input data. (The goal is to learn an encoding function f and a decoding function g so that the reconstructed data is as close to the original data as possible.)
[0049] The encoder is represented as:
[0050] h = f(X);
[0051] Among them, h is the encoded feature representation; the encoder part is implemented using convolution operations and pooling operations, expressed as:
[0052] h=σ h (W X+b);
[0053] Among them, W is the convolution kernel, b is the bias term, σ h is the ReLU function.
[0054] The decoder part is expressed as:
[0055] X′=g(h);
[0056] Among them, X′ is the reconstructed data; the decoder part is implemented using convolution operations and upsampling operations, expressed as:
[0057] X′=σ h’ (W′h+b′);
[0058] Where W′ is the convolution kernel of the decoder, b′ is the bias term of the decoder, and σ h’ is the ReLU function.
[0059] In order to minimize the difference between the reconstructed data and the original data, a loss function is defined, using the cross entropy loss function, and stopping after 1000 iterations.
[0060]
[0061] Where N is the number of audio samples, y i is the true label (0 or 1) of the i-th audio sample; is the model's predicted probability for the i-th audio sample (between 0 and 1).
[0062] The discriminator encodes the reconstructed data of the generator using the same encoding method as the generator. The goal of the discriminator is to distinguish between real normal data samples and the reconstructed data of the generator. Through training, the discriminator will learn the characteristics of all normal human voice samples.
[0063] In this embodiment, when training the model for voice anomaly detection, the target loss (total loss) includes different weight calculation methods for each sub-loss:
[0064] For each audio feature group (first audio feature (time domain feature) and second audio feature (frequency domain feature)), the corresponding sub-loss is calculated. The feature extraction of the original data and the feature extraction of the reconstructed data are compared, and the similarity between them is calculated. The similarity loss refers to the loss used to ensure that the generated reconstructed data is similar to the original data in specific features, including time domain similarity loss and frequency domain similarity loss.
[0065] The time domain similarity loss measures the similarity between the original data and the reconstructed data in the time domain, and the loss is represented by t1 or t2.
[0066] The frequency domain similarity loss measures the similarity between the inverse MFCC of the spectral feature of the original data and the inverse MFCC of the spectral feature of the reconstructed data, and the loss is represented by m1 or m2.
[0067] Based on each sub-loss, the target loss (total loss) corresponding to the audio feature group is calculated in a weighted manner. The weight can be adjusted according to the importance of the task and the importance of the feature.
[0068] The goal of the discriminator D is to maximize the following objective function, which measures the ability of D to correctly distinguish between real samples and generated samples. The objective function of the discriminator D is expressed as:
[0069] D1max V(D, G)=E t1~pdata(t1) [log(D(t1))]+E t2~p(t2) [log(1-D(G(t2)))];
[0070] D2max V(D,G)=E m1~pdata(m1) [log(D(m1))]+E m2~p(m2) [log(1-D(G(m2)))];
[0071] Among them, t1 and m1 are audio samples sampled from the real data distribution pdata(x), and t2 and m2 are random noise vectors sampled from the Gaussian distribution p(x). G(x) is the sample generated by the generator G based on the reconstructed data feature extraction; the first term E t1~pdata(t1) [log(D(t1))] encourages D to output a high probability (close to 1) for the true sample t1.
[0072] Item 2E t2~p(t2 )[log(1-D(G(t2)))] encourages D to output low probability (close to 0) for generating samples G(t2).
[0073] The goal of the generator G is to minimize the following objective function, which is expressed as:
[0074] G1min V(D, G)=-E t2~p(t2) [log(D(G(t2)))];
[0075] G2min V(D, G)=-E m2~p(m2) [log(D(G(m2)))];
[0076] Here, the generator G tries to maximize the likelihood that the discriminator D outputs true (i.e., high probability) for the sample G(z) it generates by adjusting its parameters.
[0077] In this embodiment, the test set is input into the generator, and encoding, decoding and batch normalization are performed to obtain reconstructed data. The discriminator encodes the reconstructed data of the generator using the same encoding method and outputs a probability value indicating whether the sample is a normal data sample or an abnormal data sample. By comparing the output of the discriminator with the true label of the test set, the detection performance of the model for abnormal sounds is evaluated.
[0078] Step 4: Use the trained generative adversarial network model to detect abnormalities in human voices in language programs;
[0079] During the test, the trained generator and discriminator are used to process the samples. Normal samples should be able to be reconstructed with a small error. The reconstruction error and the deformation of the frequency band are used to evaluate the effect of reconstruction. However, due to the difference in distribution between abnormal samples and normal samples, the reconstruction effect is usually poor, and obvious deformation may occur in some frequency bands. According to the reconstruction error and the deformation of the frequency band, a loss threshold is set to determine whether the sample is a normal data sample or an abnormal data sample.
[0080] According to the requirements of the State Administration of Radio, Film and Television on the distortion of language programs, the distortion scoring standards and control measures, combined with the distortion requirement for voice equipment to be <1%, the condition for the end of training should be that the target loss (total loss) should be <1%.
[0081] When the generator and discriminator perform well on the training set and the test set, the target loss reaches the distortion requirement of the State Administration of Radio, Film and Television for language programs, and can accurately identify normal data samples and abnormal data samples, the training model of the current training cycle is determined as the voice anomaly detection model.
[0082] Reference distortion scoring criteria: If the sound has no obvious distortion (such as overload, speed change, pitch change, freedom, jitter, etc.), the full score is 3 points. If the sound is distorted but does not affect the overall viewing effect, the score will be between 1 and 2 points. If the sound is severely distorted and affects the overall viewing effect, it will be scored 0 points.
[0083] Distortion control measures: Control the recording level: Generally speaking, the recording level of language programs needs to be controlled between -25 and 15dBFS to better improve the recording effect and avoid sound distortion. Reasonable use of filters: During the production process, filters can be used to filter out noise signals greater than 10kHz to improve the quality of audio signals.
[0084] The present invention deeply mines and learns the inherent distribution characteristics of normal data samples through multiple iterative training of the generative adversarial network, so that the network can accurately capture the subtle differences in normal sound patterns. In the testing phase of the present invention, the generative adversarial network model can generate data samples with high naturalness and fidelity through adversarial training of the generator and the discriminator. The generator strives to generate samples that are real enough to deceive the discriminator, while the discriminator attempts to distinguish between real samples and generated samples. This adversarial mechanism prompts the generator to continuously improve its generation ability, so that it can generate data that is very close to the real world. The normal data samples of the present invention can be reconstructed by the network almost without error, showing a high degree of fidelity, and the training process is relatively simple and efficient. Through the adversarial training of the generator and the discriminator, the generative adversarial network model can quickly learn the distribution characteristics of the data and generate corresponding data samples. This training method avoids the complex sampling and inference process in the traditional generative model, thereby improving the training efficiency. The abnormal data samples in the present invention do not conform to the learned normal patterns, and will show obvious deviations or distortions during the reconstruction process, especially in certain frequency bands. This significant difference in reconstruction effect provides a reliable basis for the model to distinguish and identify abnormal sounds, thereby achieving efficient and accurate abnormal sound detection. The present invention effectively solves the limitations of traditional autoencoders in abnormal sound detection and improves the accuracy and efficiency of detection. The present invention can accurately identify and distinguish abnormal sounds in a variety of sound environments by accurately learning and reconstructing the patterns of data.
[0085] In summary, the present invention utilizes the characteristics of high design freedom, high training efficiency, and high sample generation efficiency of the generative adversarial network model to effectively solve the limitations of traditional autoencoders in abnormal sound detection and improve the accuracy and efficiency of detection.
Claims
1. A method for detecting human voice anomalies in language programs based on a generative adversarial network, characterized in that: The following steps are involved: Step 1: Obtain audio samples of human voices in language programs, and perform data preprocessing on the audio samples to obtain normal data samples and abnormal data samples; Step 2: Extract features from normal data samples to obtain a training set, mix abnormal data samples into normal data samples and extract features to obtain a test set; Step 3: construct a generative adversarial network model, and use a training set and a test set to train and test the generative adversarial network model respectively; wherein the generative adversarial network model includes a generator and a discriminator, the generator is used to encode and decode the original data after data preprocessing to obtain reconstructed data, and the discriminator encodes the reconstructed data using the same encoding method as the generator to distinguish between real normal data samples and abnormal data samples; Step 4: Use the trained generative adversarial network model to detect anomalies in human voices in language programs.
2. The method for detecting human voice anomalies in language programs based on a generative adversarial network according to claim 1, characterized in that: In step one, the data preprocessing includes Fourier transform and normalization processing; the Fourier transform is to convert the audio sample from time domain data to frequency domain data; the normalization processing is to calculate the maximum and minimum values of the amplitude spectrum in the frequency domain data, and then apply a logarithmic transform to each amplitude value in the amplitude spectrum to compress its dynamic range, and finally perform linear normalization to map the logarithmic transformed value to a specific range.
3. The method for detecting human voice anomalies in language programs based on generative adversarial networks according to claim 2 is characterized in that: The formula for applying the logarithmic transformation is: log_F i ((i))=log_(F i (ω) / F min ((i))); Where: F i (ω) is the corresponding amplitude value in the amplitude spectrum; F min (ω) is the minimum amplitude value in the amplitude spectrum. The linear normalization formula is: F n_val =(log_F i (ω)-min[log_F i ((i))] / (max[log_F i (ω)))]-min[log_F i (oh)]); Among them, min[log_F i (ω)] and max[log_F i (ω)] are the minimum and maximum values of the amplitude spectrum after logarithmic transformation.
4. The method for detecting human voice anomalies in language programs based on a generative adversarial network according to claim 1, characterized in that: In step 2, the feature extraction includes extraction of low-level features and high-level features; the low-level features include spectrum and energy of the audio sample; the high-level features include Mel-frequency cepstral coefficients and pitch of the audio sample.
5. The method for detecting human voice anomalies in language programs based on generative adversarial networks according to claim 1, characterized in that: In step three, the generator includes an encoder and a decoder; the encoder includes 4 convolutional layers, wherein the first layer contains 64 convolutional kernels, the second layer contains 128 convolutional kernels, the third layer contains 256 convolutional kernels, and the fourth layer contains 512 convolutional kernels, and an activation function is used after each convolutional layer to increase nonlinearity; the decoder includes an upsampling layer and an output layer, the upsampling layer uses a deconvolution operation to restore the low-dimensional vector to the high-dimensional space of the original data, and an activation function is used after each deconvolution layer; the output layer is used to convert the output of the decoder into a format that is the same or similar to the original data.
6. The method for detecting human voice anomalies in language programs based on generative adversarial networks according to claim 5, characterized in that: The cross entropy loss function is used to minimize the difference between the reconstructed data and the original data; the cross entropy loss function is defined as: Where N is the number of audio samples, yi is the true label of the i-th audio sample; is the model's predicted probability for the i-th audio sample.
7. The method for detecting human voice anomalies in language programs based on a generative adversarial network according to claim 4, characterized in that: In step 3, based on the low-level features and high-level features of the training set, the corresponding sub-loss is calculated, the feature extraction of the original data and the feature extraction of the reconstructed data are compared, and the similarity between them is calculated; the sub-loss is divided into time domain similarity loss and frequency domain similarity loss, the time domain similarity loss is used to measure the similarity between the original data and the reconstructed data in the time domain, and the loss is represented by t1 or t2; the frequency domain similarity loss is used to measure the similarity between the inverse of the spectral feature SCC of the original data and the inverse of the spectral feature SCC of the reconstructed data, and the loss is represented by m1 or m2; according to each sub-loss, the target loss corresponding to the low-level features and the high-level features is calculated in a weighted manner; the goal of the discriminator D is to maximize the following objective function, and the objective function of the discriminator D is expressed as: D1max V(D,G)=E t1~pdata(t1) [log(D(t1))]+E t2~p(t2) [log(1-D(G(t2)))]; D2max V(D,G)=E m1~pdata(m1) [log(D(m1))]+E m2~p(m2) [log(1-D(G(m2)))]; Among them, t1 and m1 are audio samples sampled from the real data distribution pdata(x), t2 and m2 are random noise vectors sampled from the Gaussian distribution p(x); G(x) is the sample generated by the generator G based on the reconstructed data feature extraction; the first term Et1~pdata(t1)[log(D(t1))] encourages D to output a high probability for the real sample t1; the second term E t2~p(t2) [log(1-D(G(t2)))] encourages D to output low probability for the generated sample G(t2); The goal of the generator G is to minimize the following objective function, which is expressed as: G1min V(D,G)=-E t2~p(t2) [log(D(G(t2)))]; G2min V(D,G)=-E m2~p(m2) [log(D(G(m2)))]。 8. The method for detecting human voice anomalies in language programs based on generative adversarial networks according to claim 1, characterized in that: In step 4, the trained generator and discriminator are used to process the test set. By comparing the output of the discriminator with the true label of the test set, the detection performance of the trained model for abnormal data samples is evaluated to meet the requirements of the State Administration of Radio, Film and Television for the distortion of language programs.
