Zero sample sound signal identification method based on attribute matching
Through the method based on attribute matching, the attribute reconstruction and generation learning are used to use artificial auditory descriptors to generate invisible category audio samples, solving the problem of redundancy and mismatch between semantic auditory descriptors, and improving the accuracy and generalization ability of zero-sample audio classification.
Patent Information
- Application Number
- CN202510854588.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-24
- Publication Date
- 2025-08-29
AI Technical Summary
In the existing zero-sample learning audio classification method, there is too much redundant information in semantic auditory descriptors, and the mismatch between modals leads to insufficient generalization of categories, especially when the distribution of invisible category feature and visible categories are largely different, the classification effect is poor.
Using an attribute matching method, the attribute reconstruction and generation learning are performed through artificial auditory descriptors, invisible category audio test samples are generated, and the optimal generation learning model and classifier are used to identify audio signals.
It realizes effective classification of invisible audio samples, improving the accuracy and generalization ability of zero-sample audio classification.
Smart Images
Figure CN120564751A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of audio signal processing, and in particular relates to a zero-sample sound signal recognition method based on attribute matching. Background Art
[0002] Audio classification refers to the analysis of audio signals and their classification into different categories. As an important tool in the field of audio signal processing, audio classification has a wide range of application scenarios, covering virtual assistants, environmental sound classification applications, music genre recognition and other fields. In addition, audio classification technology enables people to quickly and accurately identify sounds, further improving the convenience of daily life. For example, smartphone applications can use audio classification functions to identify the crying of a baby and remind parents in time. Audio classification technology can also be used to quickly identify different types of music and realize convenient music streaming playback. In addition, audio classification also has important applications in the security field, such as detecting intruders in homes or commercial places; in customer service, it can realize automatic response to customer inquiries and improve service efficiency.
[0003] Zero-shot learning (ZSL) is an important research topic in machine learning. Unlike traditional supervised and unsupervised learning, ZSL can classify samples of unknown categories by learning from existing categories and assisted by inter-class information. Zero-shot learning-based audio classification applies ZSL to audio classification to alleviate the problem of incomplete audio datasets.
[0004] Most current zero-shot learning-based audio classification tasks transfer information between classes by introducing semantic information from class labels into the model. This involves using semantic auditory descriptors as prior knowledge. Semantic auditory descriptors are vectorized descriptions of sample class characteristics based on human language. These descriptions contain multi-dimensional feature information about the sample class, such as color, structure, composition, and purpose. However, the descriptive information about sound features accounts for a very limited portion of the total information contained in the semantic auditory descriptor.
[0005] Therefore, the method of using semantic auditory descriptors brings two major challenges to the audio zero-shot classification model: one is the existence of a large amount of redundant information, which may not be directly relevant to the audio classification task; the other is the mismatch problem between modalities, that is, there may be a mismatch between the information in the auditory descriptor and the characteristics of the audio data itself. Summary of the Invention
[0006] Aiming at the problems of excessive redundant information and mismatch between modalities in semantic auditory descriptors in the prior art, the present invention proposes a zero-sample acoustic signal recognition method based on attribute matching.
[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0008] A zero-sample acoustic signal recognition method based on attribute matching, comprising a data acquisition step, a model optimization step, and an application step;
[0009] The data acquisition step first selects the ESC-50 dataset, which includes several audio samples and the category label corresponding to each sample; divides the ESC-50 dataset into a training set consisting of visible category samples and a test set consisting of unseen category samples; each sample has its corresponding unique category label.
[0010] The model optimization and application steps involve sequentially executing the following steps:
[0011] Step 1: Extract the corresponding audio features from the audio samples in the training set and the test set. The classes in the training set are called visible classes, and the classes in the test set are called invisible classes. F dimensional original features, and regularize the original features to obtain N (S) The normalized features corresponding to the visible category samples And the regularized features corresponding to the unseen category samples;
[0012] Step 2: By performing nonlinear mapping on the auditory descriptor after normalization of the artificial auditory confusion matrix, the semantic embedding prototype of the visible category is obtained. The context is called attribute. where c (S) is the number of visible categories, n A is the attribute dimension, and k is the nonlinear mapping coefficient.
[0013] Step 3: Normalize the feature X according to the visible category (S) and the category labels of their corresponding samples and visible category attributes Through the attribute reconstruction model training process, the attributes of visible category samples can be reconstructed And the optimal attribute reconstruction model, then use the optimal attribute reconstruction model combined with the unseen class attributes to obtain the reconstruction attributes of the unseen class samples
[0014] Step 4: Use the regularized features X (S) and the category label y of the corresponding sample (S) , and the reconstruction properties of visible category samples Perform generative learning by training the generator f G (·) and the discriminator f D(·,·), solve the optimal generative learning model f corresponding to the generator G (·), and then use the optimal generative learning model to reconstruct the prototype in combination with the unseen categories Generate features to generate samples of unseen categories;
[0015] Step 5: Send the synthetic features of the samples generated by the unseen category into the classifier for training to obtain the optimal classifier. Then extract features of the unseen category audio samples for testing and use the regularization method described in step 1 to obtain the regularized features x of the unseen category test audio samples. (U) , and use the optimal classifier to make classification decisions.
[0016] The method of regularization processing described in step 1 is as follows:
[0017] The characteristic column vector of any sample among all audio samples before normalization is x (0) , where N (S) The training sample set composed of the feature column vectors of visible category training samples is set up for The jth characteristic element of ;
[0018] For any sample's feature column vector x (0) , feature j corresponds to the element The calculation formula for the regularization process is:
[0019]
[0020] in Represents X (0) The largest element in row j, Represents X (0) The smallest element in row j; x ·j for The result after normalization processing;
[0021] Calculate all elements in any sample according to formula (1) to obtain the normalized feature column vector x = [x ·1 , x ·2 ,…,x ·n ] T , where the normalized feature vectors of the audio signal samples belonging to the visible category training sample set constitute the normalized feature vector of the training sample
[0022] The properties described in step 2 can be achieved by performing nonlinear mapping on the auditory descriptor after normalization of the confusion matrix. The process is as follows:
[0023] There is a built-in auditory confusion matrix in the ESC-50 dataset n is the total number of categories, and the matrix element c is ij is the element in row i and column j of C, which indicates the number of times the sample with true category i is labeled as category j. First, normalize the confusion matrix row by row so that the sum of the elements in each row is 1. The processing process is as follows:
[0024]
[0025] in, is the normalized element, ensuring Reflects the probability distribution of samples of category i being mislabeled as other categories.
[0026] Then, the normalized matrix C is processed by nonlinear mapping to obtain the visible and invisible category attributes, and the elements inside the matrix C are The calculation formula of the nonlinear mapping process is:
[0027]
[0028] Where k (k>0) is a nonlinear mapping coefficient. When k takes different values, the values in different intervals will be magnified or reduced to different degrees.
[0029] Get the visible category attributes corresponding to the training set, expressed as c to be predicted in the test sample set (U) Unseen category attributes, represented as
[0030] The optimal attribute reconstruction model described in step 3 is
[0031]
[0032] where ψ(·) represents the attribute reconstruction mapping on the original attributes, and J(·, ·) represents the alignment loss function between the corresponding columns of the two parameters, which is implemented by support vector regression. is a visible category example; the linear mapping matrix The linear dimension reduction of visible category samples is performed through principal component analysis PCA; visible categories, using column orientation The normalized features of the visible category samples Get the first Centers of visible categories As examples of each visible category.
[0033] Attributes for visible and invisible categories and Use the optimal attribute reconstruction model ψ(·) to obtain the reconstruction attributes of the visible and invisible categories respectively and Audio sample labels y according to visible categories (S) ,Will Assigned to the corresponding visible category audio samples, and obtain the visible category reconstruction attributes corresponding to all visible category audio samples
[0034] The generator f described in step 4 G (·) is used to learn the optimal generative learning model and reconstruct attributes according to any category to obtain the characteristics of the corresponding category generated samples; suppose any visible category reconstructs attributes Its network structure is:
[0035]
[0036] Among them, f G1 (·,·) is the output of the first layer, and its formula is:
[0037]
[0038] Among them, n q The dimensional noise q follows a normal distribution, ReeLU(·) and LeakyReLU(·) are the rectified linear unit ReLU and Leaky ReLU activation functions respectively. and represents the linear weights of the generator network, and is the corresponding bias, and the hidden layer includes n G nodes.
[0039] Furthermore, the discriminator f in step 4 D (·,·) is used to distinguish X (S) An audio sample of any visible category or a generated sample feature x (S) , and the corresponding reconstruction attributes The formula is:
[0040]
[0041] Among them, f D1 (·,·) is the output of the hidden layer of the discriminator, and its formula is
[0042]
[0043] in and Represented as the linear weight of the discriminator network, b D1 and b D2is the corresponding bias, and n is used in the hidden layer. D nodes.
[0044] Furthermore, the process of solving the optimal generative learning model corresponding to the generator described in step 4 includes training the generator and the discriminator, and the loss function used is:
[0045]
[0046] Among them, the weight λ3>0; Features representing samples generated from visible categories In the corresponding label The classification error rate under Represents the visible category c (S) N G The visible category labels of the generated samples, each category N G Generate samples; Cls(·,·) is the audio sample X in the visible category (S) Softmax classifier trained on , using negative log-likelihood NLL loss;
[0047] Wasserstein Generative Adversarial Network (WGAN) Loss L WGAN Use Wasserstein to generate adversarial network gradient penalty WGAN-GP loss:
[0048]
[0049] The real sample features of the visible category Visible category generation sample features The weight λ1>0 is used to balance the relationship between real data and generated data. It can be seen that the characteristics of the class synthetic samples The weight λ2 follows a 0-1 uniform distribution.
[0050] The classifier described in step 5 is a supervised learning classifier, and the supervised learning classifier is a Softmax classifier:
[0051] Generating synthetic features for samples from unseen categories using an optimal generative learning model Using the logarithmic Softmax layer, the input synthetic features Converted to logarithmic probability distribution, the logarithmic Softmax function used is
[0052]
[0053] Among them, W is the weight matrix of the linear layer, b is the bias vector, and c is (U) is the number of unseen categories, and finally we get o iIt represents the log probability that the generated sample of the unseen category of the input belongs to the i-th category.
[0054] Beneficial effects: such as Figure 1 As shown, the present invention discloses a zero-sample acoustic signal recognition method based on attribute matching, which realizes the classification and recognition of invisible category audio samples by using artificial auditory descriptors to reconstruct attributes and generate invisible category audio test samples. The specific contents are as follows: audio features are extracted from visible category audio samples through a pre-training model, and the visible category descriptors are used as attributes, and combined with the visible category audio sample labels, an optimal attribute reconstruction model is trained, and then the visible category and invisible category descriptors are combined to obtain new attributes of the visible category and invisible category; then, audio features are extracted from visible category audio through a pre-training model, and the visible category audio sample labels and the new attributes of the visible category are combined to obtain the optimal generative learning model, and then the new attributes of the invisible category are combined to generate synthetic features of the invisible category samples; finally, the features of the invisible category generated samples are used to train the optimal classifier, and the labels of the tested invisible category audio samples are predicted. In terms of zero-sample audio classification, the present invention proposes a classification method for invisible category audio based on auditory descriptor matching, which can effectively distinguish the category of the audio signal.
[0055] There are two problems in the existing zero-shot audio classification methods: the semantic auditory descriptor contains a relatively limited amount of descriptive information about sound features, which brings a large amount of redundant information to the zero-shot classification process and causes the problem of mismatch between modalities; there is a problem of insufficient generalization during category transfer, especially when the feature distribution of the unseen category is very different from that of the visible category. The present invention provides a zero-shot sound signal recognition method based on attribute matching, which uses artificial auditory descriptors as attributes and adopts a zero-shot classification method based on attribute reconstruction and generative learning to generate unseen category audio test samples, thereby realizing the classification and recognition of unseen category audio samples. Descriptor matching zero-shot audio classification method
[0056] Experiments have shown that the present invention proposes a zero-sample audio classification method based on descriptor matching in audio classification, which can effectively distinguish the categories of audio samples of unseen categories. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 This is a flow chart of a zero-sample acoustic signal recognition method based on attribute matching according to the present invention. DETAILED DESCRIPTION
[0058] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0059] like Figure 1As shown, the method of the present invention first obtains artificial auditory descriptors as attributes by performing nonlinear mapping processing on the artificial classification confusion information. Then, in the attribute reconstruction stage, the audio features of the visible category audio samples are extracted through the pre-trained model, and the visible category descriptor attributes are used in combination with the visible category audio sample labels to train the optimal attribute reconstruction model. New attributes of the visible category and the invisible category are obtained through the optimal attribute reconstruction model; in the generative learning stage, the audio features of the visible category audio samples are used in combination with the visible category audio sample labels and the new attributes of the visible category to train the optimal generative learning model, and then the new attributes of the invisible category are combined to generate the features of the invisible category samples; in the supervised learning stage, the features of the invisible category generated samples are used to train the optimal classifier, and the category judgment is performed on the tested invisible category audio samples.
[0060] The following experimental method compares the unweighted accuracy (UA) recognition rate and macro F1-score of the method of the present invention with the existing zero-shot learning method.
[0061] The experiment uses audio signals in the ESC-50 (Dataset for Environmental Sound Classification) dataset and manually classified confusing information to verify the effectiveness of the method of the embodiment of the present invention.
[0062] The ESC-50 dataset is an audio dataset for environmental sound classification. The dataset contains 2,000 5-second audio samples. The dataset has five categories, namely "Animals", "Natural soundscapes & water sounds", "Human, non-speech sounds", "Interior / domestic sounds", and "Exterior / urban noises". Each category contains 10 categories, for a total of 50 categories, and each category has 40 audio samples. In view of the fact that the ESC-50 dataset has five independent categories, this experiment set up a 5-fold cross-validation. During training, the training set consists of four of the five categories, that is, the visible categories, and the remaining category is used as the test set, that is, the unseen category. The experimental results take the average of the 5-fold cross-validation.
[0063] In this experiment, when processing the spectrogram, a Hamming window with a window length of 882 sampling points was selected, and there was an overlap of 662 points between adjacent windows. It was then input into the pre-trained model ResNet50 to extract the original audio features. The dimension n F= 2048. When obtaining the descriptor, the artificial auditory descriptor is obtained as an attribute by performing nonlinear mapping on the artificial classification confusion information. The dimension n A =50, and the nonlinear mapping coefficient k value is {1, 1.5, 2, 2.5, 3}.
[0064] In the experiment, in order to reflect the effect of the method of the present invention, the methods used for comparison are: EXEM (EXEMplar synthesis) and FGN (Feature Generating Network) recognition models. The zero-sample acoustic signal recognition method based on attribute matching of the present invention is represented as: the method of the present invention (Example 1).
[0065] This experiment uses 4-fold cross validation to select the optimal parameters. As Example 1, the specific range of selected parameters is: the regularization coefficient is {2 -3 ,2 -2 ,…,2 3}, v value is {2 -8 ,2 -7 ,…,2 0}, the kernel scale parameter is {2 -4 ,2 -3 ,…,2 4}, the dimensions after principal component analysis are {1, 2, ..., 50}. The parameters in the generative learning process are set as follows: Adam optimization operator is used, and the initial learning rate is 10 -4 , the maximum number of training rounds is 30, the batch size is 64; in the generator and discriminator, n G =4096,n D =1024,n q =312; For the Softmax classifier in Cls(·,·), the Adam optimization operator is used with an initial learning rate of 10 -3 , the maximum number of training rounds is 20, the batch size is 100; the weights λ1 and λ3 are set to 10 and 1 respectively, and λ2 is randomly selected according to the uniform distribution; by using the trained generator, N G Set to {300, 600, 900}; for the classifier of the generated unseen class samples for supervised learning, the learning rate of the Adam optimization operator is set to 10 -3 , with a batch size of 25.
[0066] When the number of samples generated for the unseen class is 600 and the nonlinear mapping coefficient k is {1, 1.5, 2, 2.5, 3}, the optimal UA average results of these state-of-the-art zero-shot audio classification methods on the ESC-50 dataset are shown in Table 1.
[0067] Table 1
[0068] Adoption Method K＝1 K＝1.5 K＝2 K＝2.5 K＝3 Comparative Example 1 EXEM 0.2275 0.2330 0.2310 0.2305 0.2310 Comparative Example 2 FGN 0.2715 0.2720 0.2650 0.2635 0.2645 Example 1 Method of the present invention 0.2435 0.2760 0.2680 0.2745 0.2650
[0069] When the number of samples generated for the unseen class is 600 and the nonlinear mapping coefficient k is {1, 1.5, 2, 2.5, 3}, the F1-score results of the zero-shot audio classification method used on the ESC-50 dataset are shown in Table 2.
[0070] Table 2
[0071] Adoption Method K＝1 K＝1.5 K＝2 K＝2.5 K＝3 Comparative Example 1 EXEM 0.17344 0.19086 0.19123 0.19405 0.19788 Comparative Example 2 FGN 0.20325 0.20656 0.20419 0.20233 0.19493 Example 1 Method of the present invention 0.18867 0.20915 0.21002 0.20287 0.20573
[0072] As shown in Tables 1 and 2, the method of Example 1 achieves better UA performance and F1-score results for classifying unseen classes in audio signals than other comparative methods. Table 3 shows the optimal average UA and F1-score results for the method of Example 1, for different numbers of generated unseen class samples.
[0073] Table 3
[0074]
[0075] In summary, the method of the present invention adopted in Example 1 first reconstructs the attributes obtained from the descriptor, then obtains the optimal generative learning model through training, generates unseen category samples, and finally trains a classifier for the unseen category through supervised learning, thereby achieving better performance in the zero-sample audio classification problem.
[0076] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A zero-sample acoustic signal recognition method based on attribute matching, characterized in that: The method includes: first, performing nonlinear mapping on artificial auditory descriptors, then extracting audio features from visible category audio samples, and simultaneously using the visible category descriptors as semantic embedding prototypes, i.e., attributes, and combining them with visible category audio sample labels to train an optimal attribute reconstruction model, and then combining visible category and invisible category descriptors to obtain new attributes of visible and invisible categories; then, combining visible category audio features with visible category audio sample labels and the new attributes of visible categories to train an optimal generative learning model, and then combining the new attributes of invisible categories to generate invisible category samples; finally, using the synthesized audio features of the generated invisible category samples to train an optimal classifier to classify the tested invisible category audio samples. The method specifically comprises the following steps: Step 1: Extract the corresponding audio features from the audio samples in the training set and the test set. The classes in the training set are called visible classes, and the classes in the test set are called invisible classes. F dimensional original features, and regularize the original features to obtain N (S) The normalized features corresponding to the visible category samples And the regularized features corresponding to the unseen category samples; Step 2: By performing nonlinear mapping on the auditory descriptor after normalization of the artificial auditory confusion matrix, the semantic embedding prototype of the visible category is obtained. The context is called attribute. where c (S) is the number of visible categories, n A is the attribute dimension, and k is the nonlinear mapping coefficient. Step 3: Normalize the feature X according to the visible category (S) and the category labels of their corresponding samples and visible category attributes Perform attribute reconstruction model training to obtain the reconstructed attributes of visible category samples And the optimal attribute reconstruction model, and then use the optimal attribute reconstruction model to combine the unseen category attributes to obtain the reconstruction attributes of the unseen category samples Step 4: Use the normalized feature X of the visible category audio (S) and the category label y of the corresponding sample (S) , and the reconstruction properties of visible category samples Perform generative learning by training the generator f G (·) and the discriminator f D (·,·), solve the optimal generative learning model f corresponding to the generator G (·), then use the optimal generative learning model to reconstruct the prototype in combination with the unseen categories Generate audio features of samples generated from unseen categories; Step 5: Send the synthetic features of the samples generated by the unseen category into the classifier for training to obtain the optimal classifier. Then extract features of the unseen category audio samples for testing and use the regularization method described in step 1 to obtain the regularized features x of the unseen category test audio samples. (U) , and use the optimal classifier to make classification decisions.
2. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The method of regularization processing described in step 1 is as follows: The characteristic column vector of any sample among all audio samples before normalization is x (0) , where N (S) The training sample set composed of the feature column vectors of visible category training samples is set up for The jth characteristic element of . For any sample's characteristic column vector x (0) , feature j corresponds to the element The calculation formula for the regularization process is: in Represents X (0) The largest element in row j, Represents X (0) The smallest element in row j; x ·j for The result after normalization processing; Calculate all elements in any sample according to formula (1) to obtain the normalized feature column vector x = [x ·1 , x ·2 ,…,x ·n ] T , where the normalized feature vectors of the audio signal samples belonging to the visible category training sample set constitute the normalized feature vector of the training sample 3. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The properties described in step 2 can be achieved by performing nonlinear mapping on the auditory descriptor after normalization of the confusion matrix. The process is as follows: There is a built-in auditory confusion matrix in the ESC-50 dataset n is the total number of categories, and the matrix element c is ij The element in row i and column j of C represents the number of times that the sample with true category i is labeled as category j. First, normalize the confusion matrix row by row so that the sum of the elements in each row is 1. The processing process is as follows: in, is the normalized element, ensuring Reflects the probability distribution of samples of category u being mislabeled as other categories. Then, the normalized matrix C is processed by nonlinear mapping to obtain the visible and invisible category attributes, and the elements inside the matrix C are The calculation formula of the nonlinear mapping process is: Where k (k>0) is a nonlinear mapping coefficient. When k takes different values, the values in different intervals will be magnified or reduced to different degrees. Get the visible category attributes corresponding to the training set, expressed as c to be predicted in the test sample set (U) Unseen category attributes, represented as 4. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The optimal attribute reconstruction model described in step 3 is where ψ(·) represents the attribute reconstruction mapping on the original attributes, and J(·, ·) represents the alignment loss function between the corresponding columns of the two parameters, which is implemented by linear support vector regression. is a visible category example; the linear mapping matrix The linear dimension reduction of the visible category samples is performed by principal component analysis PCA; for the cth visible category, the column-wise and The normalized features of the visible category samples Get the center of the cth visible category As examples of each visible category. Attributes for visible and invisible categories and Use the optimal attribute reconstruction model ψ(·) to obtain the reconstruction attributes of the visible and invisible categories respectively and Audio sample labels y according to visible categories (S) ,Will Assigned to the corresponding visible category audio samples, and obtain the visible category reconstruction attributes corresponding to all visible category audio samples 5. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The generator f described in step 4 G (·) It is used to learn the optimal generative learning model and reconstruct attributes according to any category to obtain the audio features of the generated samples of the corresponding category; Let any visible category reconstruction attribute Its network structure is: Among them, f G1 (·,·) is the output of the first layer, and its formula is: Among them, n q The dimensional noise q follows a normal distribution, ReLU(·) and LeakyReLU(·) are the rectified linear unit ReLU and Leaky ReLU activation functions, respectively. and represents the linear weights of the generator network, and is the corresponding bias, and the hidden layer includes n G nodes.
6. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The discriminator f in step 4 D (·,·) is used to distinguish X (S) An audio sample of any visible category in or a generated sample audio feature x (S) , and the corresponding reconstruction attributes The formula is: Among them, f D1 (·,·) is the output of the hidden layer of the discriminator, and its formula is in and Represented as the linear weight of the discriminator network, b D1 and b D2 is the corresponding bias, and n is used in the hidden layer. D nodes.
7. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The process of solving the optimal generative learning model corresponding to the generator described in step 4 includes training the generator and the discriminator, and the loss function used is: Among them, the weight λ3>0; Features representing samples generated from visible categories In the corresponding label The classification error rate under Represents the visible category c (S) N G The visible category labels of the generated samples, each category N G Generate samples; Cls(·,·)) is the audio sample X in the visible category (S) Softmax classifier trained on , using negative log-likelihood NLL loss; Wasserstein Generative Adversarial Network (WGAN) Loss L WGAN Use Wasserstein to generate adversarial network gradient penalty WGAN-GP loss: The real sample features of the visible category Visible category generation sample features The weight λ1>0 is used to balance the relationship between real data and generated data. It can be seen that the characteristics of the class synthetic samples The weight λ2 follows a 0-1 uniform distribution.
8. The zero-sample acoustic signal recognition method based on attribute matching according to claim 1, characterized in that: The classifier described in step 5 is a supervised learning classifier, and the supervised learning classifier is a Softnax classifier: Generating synthetic features for samples from unseen categories using an optimal generative learning model Using the logarithmic Softmax layer, the input synthetic features Converted to logarithmic probability distribution, the logarithmic Softmax function used is Among them, W is the weight matrix of the linear layer, b is the bias vector, and c is (U) is the number of unseen categories, and finally we get o i It represents the log probability that the generated sample of the unseen category of the input belongs to the i-th category.