Audio identification model training method and device, audio identification method and device and storage medium

Abstract

The invention provides an audio identification model training method and device, an audio identification method and device and a storage medium, and the method comprises the steps: obtaining an audio sample data set, and carrying out the preprocessing of the audio in the sample data set, and obtaining a training data set; inputting the audio in the training data set into a preset feature extraction network to obtain a fusion feature vector; inputting the fused feature vector into a preset heterogeneous graph attention network to obtain a target feature vector; obtaining a corresponding loss value by learning a loss function based on the target feature vector; and updating network parameters in a preset feature extraction network and a preset heterogeneous graph attention network based on the loss value, and determining a target feature extraction network and a target heterogeneous graph attention network obtained through training as a target audio identification model. The generalization ability of the model and the accuracy of the identification result are improved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a training method for an audio identification model, an audio identification method, an apparatus, and a storage medium. Background Technology

[0002] AI technology can generate highly realistic speech, and synthesized audio has been widely used in various fields, such as education. However, synthesized audio is difficult to distinguish from human speech, which can pose a threat to social security. Therefore, it is necessary to authenticate synthesized audio to effectively protect against attacks using AI-generated audio.

[0003] In related technologies, end-to-end AI-based synthetic audio discrimination models based on deep learning (such as RawNet2) can directly extract features from the original audio waveform, avoiding the bias of manual feature design, and improve discrimination performance by learning high-level semantic features hierarchically through deep neural networks. However, this model does not sufficiently extract artifact information from synthetic audio, resulting in missing discriminative information and thus reducing the accuracy of the discrimination results. Furthermore, while the model is trained using the SoftMax loss function, this function is suitable for tasks with balanced training and test sets, whereas synthetic speech discrimination tasks are diverse and unknown, with a wider range of types, thus limiting the model's generalization ability. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this disclosure provides a training method for an audio discrimination model, an audio discrimination method, an apparatus, and a storage medium.

[0005] According to a first aspect of the present disclosure, a method for training an audio discrimination model is provided, the method comprising: Obtain an audio sample dataset and preprocess the audio in the sample dataset to obtain a training dataset; The audio from the training dataset is input into a preset feature extraction network to obtain a fused feature vector; The fused feature vector is input into a preset heterogeneous graph attention network to obtain the target feature vector; Based on the target feature vector, the corresponding loss value is obtained by learning the loss function; The network parameters in the preset feature extraction network and the preset heterogeneous graph attention network are updated based on the loss value, and the target feature extraction network and the target heterogeneous graph attention network obtained through training are determined as the target audio discrimination model.

[0006] Optionally, in this embodiment of the disclosure, the preprocessing of the audio in the sample dataset to obtain the training dataset includes: The audio in the sample dataset is subjected to speech activity detection and unification processing, and the processed audio is determined as the training dataset.

[0007] Optionally, in this embodiment of the disclosure, the preset feature extraction network includes a speech representation learning model and a ViT model; the step of inputting audio from the training dataset into the preset feature extraction network to obtain a fused feature vector includes: In the speech representation learning model, features are extracted from the audio in the training dataset to obtain a first feature vector; The audio is converted to obtain the corresponding Mel spectrogram; The ViT model extracts features from the Mel spectrogram to obtain a second feature vector; The first feature vector and the second feature vector are fused to obtain a fused feature vector.

[0008] Optionally, in this embodiment of the disclosure, the ViT model performs feature extraction on the Mel spectrogram to obtain a second feature vector, including: The Mel spectrogram is segmented to obtain multiple image blocks; Perform linear projection embedding on the image patch to generate a corresponding first embedding vector; Add positional encoding to the first embedding vector to generate a corresponding second embedding vector, wherein the positional encoding includes time axis encoding and frequency axis encoding; The second embedding vector is input into the Transformer encoder to obtain the second feature vector.

[0009] Optionally, in this embodiment of the disclosure, the preset heterogeneous graph attention network includes a residual processing layer, a max pooling layer, a graph module layer, a heterogeneous superimposed graph attention layer, and an output layer; the step of inputting the fused feature vector into the preset heterogeneous graph attention network to obtain the target feature vector includes: The residual processing layer processes the fused feature vector to obtain a high-level feature map; The max pooling layer separates the high-level feature map to obtain a spectral representation vector and a temporal representation vector; The graph module layer is constructed based on the spectral representation vector and the time representation vector to obtain the spectral input graph and the time input graph; The heterogeneous overlay map attention layer combines the spectral input map and the temporal input map to obtain a heterogeneous spectral-time map; The output layer performs result prediction based on the heterogeneous spectrum-time map to obtain the target feature vector and the predicted audio identification result.

[0010] Optionally, in this embodiment of the disclosure, the graph module layer includes a spectral graph module and a temporal graph module, wherein the spectral graph module includes a spectral graph attention network and a spectral graph pooling layer, and the temporal graph module includes a temporal graph attention network and a temporal graph pooling layer; the graph module layer is constructed based on the spectral representation vector and the temporal representation vector to obtain a spectral input graph and a temporal input graph, including: The spectral attention network processes the spectral representation vector to obtain the spectral feature vector; The spectral pooling layer processes the spectral feature vector to obtain the spectral input map; The time-map attention network processes the time representation vector to obtain a time feature vector; The time-map pooling layer processes the time feature vector to obtain a time input map.

[0011] Optionally, in this embodiment of the disclosure, the output layer includes a readout layer and a fully connected layer; the output layer performs result prediction based on the heterogeneous spectral-time map to obtain a target feature vector and a predicted audio discrimination result, including: The readout layer performs feature readout on the heterogeneous spectrum-time map to obtain a first feature vector, a second feature vector, a third feature vector, a fourth feature vector, and a fifth feature vector; The first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector are concatenated to obtain the target feature vector; The fully connected layer predicts the result based on the target feature vector to obtain the predicted audio identification result.

[0012] Optionally, in this embodiment of the disclosure, the readout layer performs feature readout on the heterogeneous spectrum-time map to obtain a first feature vector, a second feature vector, a third feature vector, a fourth feature vector, and a fifth feature vector, including: The readout layer performs a maximum value operation on the spectral nodes in the heterogeneous spectrum-time graph to obtain a first feature vector. The spectral nodes in the heterogeneous spectrum-time plot are averaged to obtain the second feature vector. The third eigenvector is obtained by performing a maximum value operation on the time nodes in the heterogeneous spectrum-time plot. The time nodes in the heterogeneous spectrum-time plot are averaged to obtain the fourth feature vector. The feature vector in the stack node is determined as the fifth feature vector.

[0013] According to a second aspect of the present disclosure, an audio identification method is provided, the method comprising: Obtain the audio to be judged; The audio to be judged is preprocessed to obtain the processed audio; The processed audio is input into the target audio discrimination model to obtain the audio discrimination result.

[0014] According to a third aspect of the present disclosure, a training apparatus for an audio discrimination model is provided, the apparatus comprising: The preprocessing module is used to acquire an audio sample dataset and preprocess the audio in the sample dataset to obtain a training dataset. The feature extraction module is used to input the audio from the training dataset into a preset feature extraction network to obtain a fused feature vector; An audio identification module is used to input the fused feature vector into a preset heterogeneous graph attention network to obtain a target feature vector. The processing module is used to obtain the corresponding loss value based on the target feature vector by learning a loss function; The update module is used to update the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value, and to determine the target feature extraction network and target heterogeneous graph attention network obtained through training as the target audio discrimination model.

[0015] According to a fourth aspect of the present disclosure, an audio identification device is provided, the device comprising: The acquisition module is used to acquire the audio to be judged; The preprocessing module is used to preprocess the audio to be judged to obtain the processed audio; The audio discrimination module is used to input the processed audio into the target audio discrimination model to obtain the audio discrimination result.

[0016] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method described in the first or second aspect above.

[0017] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, the computer storage medium storing computer-executable instructions; the computer-executable instructions, when executed by a processor, are capable of implementing the method described in the first or second aspect above.

[0018] According to a seventh aspect of the present disclosure, a computer program product is provided, including a computer program that, when executed by a processor, can implement the method described in the first or second aspect described above.

[0019] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects: This disclosure proposes a training method, device, and storage medium for an audio discrimination model. The method includes acquiring an audio sample dataset and preprocessing the audio in the sample dataset to obtain a training dataset; inputting the audio from the training dataset into a preset feature extraction network to obtain a fused feature vector; inputting the fused feature vector into a preset heterogeneous graph attention network to obtain a target feature vector; obtaining a corresponding loss value based on the target feature vector through a learning loss function; updating the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value; and determining the trained target feature extraction network and target heterogeneous graph attention network as the target audio discrimination model. Therefore, this disclosure can fully extract features from the audio through a preset feature extraction network to obtain a fused feature vector, and use the fused feature vector for model training, making the model's discrimination results more accurate and improving the accuracy of the discrimination results. Simultaneously, by using a learning loss function to compress the embedded features of real speech and effectively recognize synthetic audio, the generalization ability of the model is improved.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0021] Figure 1 This is a flowchart illustrating a training method for an audio discrimination model according to some embodiments of the present disclosure; Figure 2 This is a flowchart illustrating a training method for an audio discrimination model according to some embodiments of the present disclosure; Figure 3 This is a flowchart illustrating a training method for an audio discrimination model according to some embodiments of the present disclosure; Figure 4 This is a flowchart illustrating an audio identification method according to some embodiments of the present disclosure; Figure 5 This is a flowchart illustrating an audio identification method according to some embodiments of the present disclosure; Figure 6This is a schematic diagram of the structure of a training device for an audio discrimination model according to some embodiments of the present disclosure; Figure 7 This is a schematic diagram of the structure of an audio identification device according to some embodiments of the present disclosure; Figure 8 This is a block diagram illustrating an electronic device suitable for performing a training method for an audio discrimination model or an audio discrimination method, according to some embodiments of the present disclosure. Detailed Implementation

[0022] Some embodiments of this disclosure will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. Various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent upon understanding this disclosure. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but can be changed as will become apparent upon understanding this disclosure, except for operations that must be performed in a particular order. Furthermore, for clarity and brevity, descriptions of features known in the art may be omitted.

[0023] The embodiments described in the following examples of this disclosure are not representative of all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0024] Currently, synthetic audio attacks can be categorized based on the techniques used to create fake audio. Among these, attack types can include replay attacks, specifically text-to-speech (TTS) and voice-to-speech (VC).

[0025] Traditional audio identification methods rely on signal processing techniques and machine learning algorithms. These methods typically include the following steps: preprocessing, feature extraction, feature selection, classifier training, and identification. The preprocessing process is cumbersome: it usually requires multiple steps of preprocessing the audio signal, including noise reduction, silence removal, and frame segmentation. These steps are not only time-consuming but also prone to introducing errors, affecting the accuracy of subsequent feature extraction. Feature extraction and selection are complex: traditional methods rely on manually designed features, such as Mel-Frequency Cepstral Coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC), making it difficult to fully capture the subtle features of synthetic audio. The classifiers have limitations: traditional methods usually use traditional machine learning algorithms such as Support Vector Machines (SVM) and Random Forests for classification. These algorithms have limited performance in handling high-dimensional feature spaces and nonlinear relationships, making them unsuitable for complex synthetic audio identification tasks. Furthermore, traditional methods have poor robustness: they are sensitive to changes in environmental noise and audio quality, easily affecting their identification performance.

[0026] Furthermore, in practical applications, the quality of synthesized audio varies depending on the recording environment, and traditional methods struggle to maintain efficient discrimination capabilities under various conditions. Moreover, the generalization ability of traditional machine learning models is insufficient to meet the challenges of an increasing number of synthesized audio methods.

[0027] Figure 1 This is a flowchart illustrating a training method for an audio discrimination model according to some embodiments of the present disclosure, such as... Figure 1 As shown, the method may include the following steps: Step 101: Obtain the audio sample dataset and preprocess the audio in the sample dataset to obtain the training dataset.

[0028] It should be noted that the main body executing the above-mentioned audio discrimination model training method is the audio discrimination model training device, which can be implemented by software and / or hardware. In this embodiment, the audio discrimination model training device can be configured in an electronic device.

[0029] In this example embodiment, the electronic device may include devices such as terminal devices and servers, and this embodiment does not limit the electronic device.

[0030] In this embodiment of the disclosure, the audio sample dataset includes real speech data and AI (Artificial Intelligence) synthesized speech data.

[0031] In this embodiment of the disclosure, the above-mentioned AI-synthesized speech data can be generated using various methods such as EmotiVoice and MeloTTS.

[0032] Furthermore, in this embodiment of the disclosure, after obtaining the audio sample dataset through the above steps, the audio in the sample dataset can be preprocessed to obtain the training dataset.

[0033] Specifically, in this embodiment of the present disclosure, the method of preprocessing the audio in the sample dataset to obtain the training dataset may include: performing speech activity detection processing and unification processing on the audio in the sample dataset, and determining the processed audio as the training dataset.

[0034] In this embodiment of the disclosure, the audio in the sample dataset can be processed by a WebRTC speech activity detector to detect speech activity and remove silent segments from the detection results, thereby completing the speech activity detection processing of the audio in the sample dataset, so that the processed audio is a pure audio segment without silent parts.

[0035] Furthermore, in this embodiment of the disclosure, the audio processed by voice activity detection can be subjected to volume normalization processing to avoid the impact of volume on the recognition effect. Specifically, the pydub library can be used to perform audio volume normalization processing, such as normalizing the audio volume to -20 dBFS.

[0036] In this embodiment of the disclosure, the audio processed by the above-described speech activity detection and unification is determined as the training dataset.

[0037] Step 102: Input the audio from the training dataset into the preset feature extraction network to obtain the fused feature vector.

[0038] In this embodiment of the disclosure, after obtaining the training dataset through the above steps, the audio in the training dataset can be input into a preset feature extraction network to obtain a fused feature vector, thereby giving the obtained fused feature vector higher discrimination accuracy and robustness.

[0039] In this embodiment of the disclosure, the aforementioned preset feature extraction network may include a speech representation learning model and a ViT (Vision Transformer) model.

[0040] Furthermore, in this embodiment of the present disclosure, the method of inputting audio from the training dataset into a preset feature extraction network to obtain a fused feature vector may include: extracting features from the audio in the training dataset in the speech representation learning model to obtain a first feature vector; converting the audio to obtain a corresponding Mel spectrogram; extracting features from the Mel spectrogram using the ViT model to obtain a second feature vector; and fusing the first feature vector and the second feature vector to obtain a fused feature vector.

[0041] Step 103: Input the fused feature vector into the preset heterogeneous graph attention network to obtain the target feature vector.

[0042] In this embodiment of the disclosure, after obtaining the fused feature vector through the above steps, the fused feature vector can be input into a preset heterogeneous graph attention network to obtain the target feature vector, so as to obtain the predicted audio identification result through the target feature vector.

[0043] In this embodiment, the aforementioned preset heterogeneous graph attention network may include a residual processing layer, a max pooling layer, a graph module layer, a heterogeneous overlay graph attention layer, and an output layer. Specifically, in this embodiment, the input fused feature vector can be processed through the residual processing layer, max pooling layer, graph module layer, heterogeneous overlay graph attention layer, and output layer to obtain the target feature vector and the predicted audio discrimination result. This will be described in detail in subsequent embodiments.

[0044] Step 104: Based on the target feature vector, obtain the corresponding loss value by learning the loss function.

[0045] In this embodiment of the disclosure, after determining the target feature vector through the above steps, the corresponding loss value can be obtained based on the target feature vector by learning the loss function.

[0046] In this embodiment of the disclosure, a compact embedding space can be trained for real speech during audio identification. However, training a compact embedding space for AI-synthesized audio attacks may lead to overfitting to known attacks, thereby reducing the model's generalization ability. Therefore, this embodiment of the disclosure introduces two different boundaries through a learning loss function to better compress real speech and isolate AI-synthesized audio attacks, thus avoiding model overfitting.

[0047] In this embodiment of the disclosure, the learning loss function can be a single-class learning loss function. Specifically, in this embodiment, the learning loss function is: ) Wherein, the weight vector w0 represents the optimization direction of the target class embedding. To determine the number of audio files in the training dataset, It is a scaling factor. It is the target feature vector. These are the true labels of the samples, with the two corner margins used to define w0 and x. i The angle between them, using Indicates. When hour, Used to make Less than arccosm0, when hour, Used to make It is greater than arccosm1. Also, a smaller arccosm0 can concentrate the target class around the weight vector w0, while a larger arccosm1 can push non-target data away from w0.

[0048] In this embodiment of the disclosure, by using the above-mentioned learning loss function and setting different corner margin constraints, the embedding features of real speech are compressed, thereby effectively isolating AI-synthesized audio attacks and achieving better performance in AI audio identification tasks.

[0049] Step 105: Update the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value, and determine the target feature extraction network and the target heterogeneous graph attention network obtained through training as the target audio discrimination model.

[0050] In this embodiment of the disclosure, after obtaining the loss value through the above steps, the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network can be updated based on the loss value until the number of times the network parameters are updated reaches a preset number, thereby obtaining the trained target feature extraction network and target heterogeneous graph attention network, and determining the target feature extraction network and target heterogeneous graph attention network as the target audio discrimination model.

[0051] In this embodiment of the disclosure, the preset number of times can be set as needed, such as 100.

[0052] This disclosure proposes a training method for an audio discrimination model. The method includes acquiring an audio sample dataset and preprocessing the audio in the dataset to obtain a training dataset; inputting the audio from the training dataset into a pre-defined feature extraction network to obtain a fused feature vector; inputting the fused feature vector into a pre-defined heterogeneous graph attention network to obtain a target feature vector; obtaining a corresponding loss value based on the target feature vector through a learning loss function; updating the network parameters in the pre-defined feature extraction network and the pre-defined heterogeneous graph attention network based on the loss value; and determining the trained target feature extraction network and target heterogeneous graph attention network as the target audio discrimination model. Therefore, this disclosure can fully extract features from the audio through a pre-defined feature extraction network to obtain a fused feature vector, and use the fused feature vector for model training, making the model's discrimination results more accurate and improving the accuracy of the discrimination results. Simultaneously, by using a learning loss function to compress the embedded features of real speech and effectively recognize synthetic audio, the generalization ability of the model is improved.

[0053] In this embodiment of the disclosure, as a detailed explanation of step 102, such as Figure 2 As shown, it may also include: Step 201: In the speech representation learning model, features are extracted from the audio in the training dataset to obtain the first feature vector.

[0054] In this embodiment of the disclosure, the speech representation learning model includes a convolutional neural network (CNN) and a Transformer.

[0055] In this embodiment of the disclosure, the method for extracting features from the audio in the training dataset to obtain a first feature vector in the above-mentioned speech representation learning model may include: extracting local acoustic features from the audio through a convolutional neural network to obtain a hidden feature sequence, and processing the hidden feature sequence through a Transformer to obtain a first feature vector.

[0056] Step 202: Convert the audio to obtain the corresponding Mel spectrogram.

[0057] In this embodiment of the disclosure, the method for converting audio to obtain the corresponding Mel spectrogram may include: converting the audio signal X into a Mel spectrogram M (e.g., 224x224) through a short-time Fourier transform (STFT).

[0058] In this embodiment of the disclosure, the Mel-scale can simulate the non-linear perception of frequency by the human ear, thus making it suitable for speech tasks.

[0059] Step 203: The ViT model extracts features from the Mel spectrogram to obtain the second feature vector.

[0060] In this embodiment of the disclosure, after obtaining the Mel spectrogram through the above steps, the Mel spectrogram can be input into the ViT model, and the ViT model can be used to extract features from the Mel spectrogram to obtain the second feature vector.

[0061] Specifically, in this embodiment of the disclosure, the method for extracting features from the Mel spectrogram using the ViT model to obtain the second feature vector may include the following steps: Step 2031: Perform image segmentation on the Mel spectrogram to obtain multiple image blocks.

[0062] In this embodiment of the disclosure, the Mel spectrogram can be segmented as needed to obtain multiple image blocks.

[0063] In this embodiment of the disclosure, the 224x224 Mel spectrogram is divided into 16x16 image blocks, that is, a total of 196 image blocks are obtained. In this embodiment of the disclosure, the dimension of the image block is (16x16=256 dimensions).

[0064] Step 2032: Perform linear projection embedding on the image patch to generate the corresponding first embedding vector.

[0065] In this embodiment of the disclosure, after obtaining multiple image blocks through the above steps, linear projection embedding can be performed on the image blocks to generate a corresponding first embedding vector.

[0066] In this embodiment of the disclosure, each image patch is linearly projected and mapped to a first embedding vector, for example, mapping 256 dimensions to 768 dimensions.

[0067] Step 2033: Add position encoding to the first embedding vector to generate the corresponding second embedding vector, wherein the position encoding includes time axis encoding and frequency axis encoding.

[0068] In this embodiment of the disclosure, after obtaining the first embedding vector through the above steps, position encoding can be added to the first embedding vector to generate the corresponding second embedding vector, so as to enhance the model's understanding of the time series and frequency domain.

[0069] Specifically, in this embodiment of the disclosure, the Mel spectrogram is a time-frequency diagram, which can encode time and frequency positions respectively. Based on this, the above-mentioned position encoding can include time axis encoding and frequency axis encoding.

[0070] In this embodiment of the disclosure, the time axis encoding can be the timing information of the audio (such as the frame number), and the frequency axis encoding can be the Mel frequency band number (such as low frequency to high frequency).

[0071] Step 2034: Input the second embedding vector into the Transformer encoder to obtain the second feature vector.

[0072] In this embodiment of the disclosure, after obtaining the second embedding vector through the above steps, the second embedding vector can be input into the Transformer encoder to obtain the second feature vector.

[0073] Step 204: Perform feature fusion between the first feature vector and the second feature vector to obtain a fused feature vector.

[0074] In this embodiment of the disclosure, after obtaining the first feature vector and the second feature vector through the above steps, the first feature vector and the second feature vector can be fused to effectively integrate and enhance the relevant information in the obtained features, thereby obtaining a fused feature vector.

[0075] In this embodiment of the disclosure, the method for fusing the first feature vector and the second feature vector to obtain a fused feature vector may include: dynamically weighting the first feature vector and the second feature vector to obtain the fused feature vector. Furthermore, the weight coefficients corresponding to the first feature vector and the second feature vector can be set as needed, and the sum of the weight coefficients corresponding to the first feature vector and the second feature vector is 1, such as the weight coefficient of the first feature vector being 0.5 and the weight coefficient of the second feature vector being 0.5.

[0076] In this embodiment of the disclosure, the first feature vector obtained by the speech representation learning model and the second feature vector obtained by the ViT model can be fused through the above steps to obtain a fused feature vector, thereby giving the fused feature vector higher discrimination accuracy and robustness.

[0077] In this embodiment of the disclosure, as a detailed explanation of step 103, such as Figure 3 As shown, it may also include: Step 301: The residual processing layer processes the fused feature vector to obtain the advanced feature map.

[0078] In this embodiment of the disclosure, the residual processing layer can use a residual encoder, such as the residual encoder of RawNet2. Based on this, the fused feature vector can be processed by the large residual encoder in the residual processing layer to obtain a high-level feature map.

[0079] Step 302: The max pooling layer separates the high-level feature maps to obtain the spectral representation vector and the temporal representation vector.

[0080] In this embodiment of the disclosure, after obtaining the high-level feature map through the above steps, the high-level feature map can be separated by a max pooling layer to obtain a spectral representation vector and a temporal representation vector.

[0081] In this embodiment of the disclosure, information can be compressed on both the frequency axis (spectral dimension) and the time axis (temporal dimension) using a max-pooling layer to capture significant features in the frequency domain (such as formants and harmonic structures) to obtain a spectral representation vector. Additionally, key changes in the time domain (such as phoneme boundaries and energy abrupt changes) can be captured to obtain a temporal representation vector.

[0082] Step 303: The graph module layer is constructed based on the spectral representation vector and the time representation vector to obtain the spectral input graph and the time input graph.

[0083] In this embodiment of the disclosure, after obtaining the spectral representation vector and the time representation vector through the above steps, the spectral input graph and the time input graph can be constructed based on the spectral representation vector and the time representation vector through the graph module layer, thereby efficiently modeling the complex dependencies between the frequency domain and the time domain in the speech signal, and capturing non-local interaction features by utilizing the flexibility of the graph structure.

[0084] In this embodiment of the disclosure, the graph module layer may include a spectral graph module and a temporal graph module. The spectral graph module may include a spectral graph attention network and a spectral graph pooling layer, and the temporal graph module may include a temporal graph attention network and a temporal graph pooling layer.

[0085] Specifically, in this embodiment of the disclosure, the above-mentioned graph module layer is constructed based on spectral representation vectors and temporal representation vectors. The method for obtaining the spectral input graph and the temporal input graph may include the following steps: Step 3031: The spectral graph attention network processes the spectral representation vector to obtain the spectral feature vector.

[0086] In this embodiment of the disclosure, the spectral attention network can process the spectral representation vector, model the relationship between frequency points, and obtain the spectral feature vector.

[0087] Step 3032: The spectral pooling layer processes the spectral feature vector to obtain the spectral input map.

[0088] In this embodiment of the disclosure, after obtaining the spectral feature vector through the above steps, the spectral feature vector can be reduced in dimensionality and key frequency domain information can be retained by the spectral pooling layer to obtain the spectral input map.

[0089] Step 3033: The time-map attention network processes the time representation vector to obtain the time feature vector.

[0090] In this embodiment of the disclosure, the time-graph attention network models the dependencies between time steps based on the time representation vector to obtain the time feature vector.

[0091] Step 3034: The time-map pooling layer processes the time feature vector to obtain the time input map.

[0092] In this embodiment of the disclosure, after obtaining the time feature vector through the above steps, the time feature vector can be processed by the time graph pooling layer to obtain the time input graph.

[0093] Step 304: The heterogeneous overlay map attention layer combines the spectral input map and the temporal input map to obtain a heterogeneous spectral-time map.

[0094] In this embodiment of the disclosure, after obtaining the spectral input map and the time input map through the above steps, the spectral input map and the time input map can be combined through the heterogeneous overlay map attention layer, thereby capturing the relationship between artifacts across the time and spectral domains and obtaining a heterogeneous spectral-time map.

[0095] In this embodiment, the Heterogeneous Stacked Graph Attention Layer (HS-GAL) may include stacked nodes and an attention mechanism. Furthermore, in this embodiment, the interaction between the temporal and spectral domains is explicitly modeled through stacked nodes as an intermediary; the weights between nodes are calculated using multi-head heterogeneous attention within the attention mechanism, and the information from the stacked nodes is aggregated using these weights to obtain a heterogeneous spectral-temporal graph.

[0096] In this embodiment of the disclosure, the stack node can be a global node to aggregate global information and transmit it between multiple layers in the multi-layer HS-GAL, so that each layer can retain and transmit global time-frequency information.

[0097] Step 305: The output layer predicts the results based on the heterogeneous spectrum-time map to obtain the target feature vector and the predicted audio identification result.

[0098] In this embodiment of the disclosure, after obtaining the heterogeneous spectrum-time map through the above steps, the output layer can perform result prediction based on the heterogeneous spectrum-time map to obtain the target feature vector and the predicted audio identification result.

[0099] In this embodiment of the disclosure, the output layer may include a readout layer and a fully connected layer.

[0100] In this embodiment of the disclosure, the method for the output layer to predict the result based on the heterogeneous spectrum-time map and obtain the target feature vector and the predicted audio identification result may include the following steps: Step 3051: The readout layer performs feature readout on the heterogeneous spectrum-time map to obtain the first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector.

[0101] In this embodiment of the disclosure, the method for the readout layer to perform feature readout on the heterogeneous spectrum-time map to obtain a first feature vector, a second feature vector, a third feature vector, a fourth feature vector, and a fifth feature vector may include the following steps: Step 1: The readout layer performs a maximum value operation on the spectral nodes in the heterogeneous spectrum-time plot to obtain the first eigenvector; Step 2: Average the spectral nodes in the heterogeneous spectrum-time plot to obtain the second eigenvector; Step 3: Perform a maximum value operation on the time nodes in the heterogeneous spectrum-time plot to obtain the third eigenvector; Step 4: Average the time nodes in the heterogeneous spectrum-time plot to obtain the fourth eigenvector; Step 5: Determine the fifth feature vector in the stack node.

[0102] In this embodiment of the disclosure, the stack node is a global node in the attention layer of the heterogeneous overlay graph.

[0103] Step 3052: Concatenate the first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector to obtain the target feature vector.

[0104] Step 3053: The fully connected layer predicts the result based on the target feature vector to obtain the predicted audio identification result.

[0105] In this embodiment of the disclosure, after obtaining the target feature vector through the above steps, the result can be predicted based on the target feature vector through a fully connected layer to obtain the predicted audio identification result.

[0106] In this embodiment of the disclosure, the predicted audio identification result can be real or synthesized. Specifically, when the predicted audio identification result is real, it means that the audio is real audio and not synthesized audio; when the predicted audio identification result is synthesized, it means that the audio is synthesized audio and not real audio.

[0107] In this embodiment of the disclosure, multiple features are effectively fused through a pre-set heterogeneous graph attention network to obtain the predicted audio identification result. This allows for learning of multiple speech types and improves the model's generalization ability.

[0108] Figure 4 This is a flowchart illustrating an audio identification method according to some embodiments of the present disclosure, such as... Figure 4 As shown, the method may include the following steps: Step 401: Obtain the audio to be judged.

[0109] In this embodiment of the disclosure, the audio input by the user to be judged can be obtained.

[0110] Step 402: Preprocess the audio to be judged to obtain the processed audio.

[0111] In this embodiment of the disclosure, after obtaining the audio to be judged through the above steps, the audio to be judged can be preprocessed to obtain the processed audio, so that the audio judgment result can be obtained through the processed audio in the future.

[0112] In this embodiment of the disclosure, the method for preprocessing the audio to be judged to obtain processed audio may include: performing speech activity detection processing and unification processing on the audio to be judged to obtain processed audio. For details on this part, please refer to the detailed description in the above embodiments; this disclosure will not repeat it here.

[0113] Step 403: Input the processed audio into the target audio discrimination model to obtain the audio discrimination result.

[0114] In this embodiment of the disclosure, after obtaining the processed audio through the above steps, the processed audio can be input into the target audio discrimination model to obtain the audio discrimination result.

[0115] This disclosure proposes an audio discrimination method, which includes acquiring the audio to be discriminated; preprocessing the audio to be discriminated to obtain processed audio; and inputting the processed audio into a target audio discrimination model to obtain the audio discrimination result.

[0116] Therefore, this disclosure can use the target audio discrimination model to perform audio discrimination on the audio to be judged, obtain audio discrimination results, and improve the accuracy of the discrimination results.

[0117] Based on the above description Figure 5 This is a flowchart illustrating an audio identification method according to some embodiments of the present disclosure, such as... Figure 5As shown, the method may include: acquiring audio data to be discriminated; inputting the audio data to be discriminated into a target audio discrimination model; and obtaining audio discrimination results through the target audio discrimination model. Specifically, the target audio discrimination model is obtained by training a preset audio discrimination model. The method for training the preset audio discrimination model may include: acquiring an audio sample dataset; preprocessing the audio in the audio sample dataset to obtain a training dataset; extracting features from the training data using a speech representation learning model and a ViT model to obtain a fused feature vector; inputting the fused feature vector into a preset heterogeneous graph attention network to obtain a target feature vector; obtaining the corresponding loss value based on the target feature vector using a single-class learning loss function, and updating the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value, until the target audio discrimination model is obtained.

[0118] Figure 6 This is a training apparatus for an audio discrimination model shown according to some embodiments of the present disclosure, such as... Figure 6 As shown, the device may include: The preprocessing module 601 is used to acquire an audio sample dataset and preprocess the audio in the sample dataset to obtain a training dataset. Feature extraction module 602 is used to input audio from the training dataset into a preset feature extraction network to obtain a fused feature vector; The audio identification module 603 is used to input the fused feature vector into a preset heterogeneous graph attention network to obtain the target feature vector; Processing module 604 is used to obtain the corresponding loss value based on the target feature vector by learning the loss function; The update module 605 is used to update the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value, and to determine the target feature extraction network and the target heterogeneous graph attention network obtained through training as the target audio discrimination model.

[0119] In this embodiment of the disclosure, the preprocessing module 601 is specifically used for: The audio in the sample dataset is processed for speech activity detection and unification, and the processed audio is used as the training dataset.

[0120] Optionally, in this embodiment of the disclosure, the preset feature extraction network includes a speech representation learning model and a ViT model; the feature extraction module 602 is specifically used for: In the speech representation learning model, features are extracted from the audio in the training dataset to obtain the first feature vector; The audio is converted to obtain the corresponding Mel spectrogram; The ViT model extracts features from the Mel spectrogram to obtain the second feature vector. The first feature vector and the second feature vector are fused to obtain the fused feature vector.

[0121] Optionally, in this embodiment of the disclosure, the feature extraction module 602 is further used for: Image segmentation is performed on the Mel spectrogram to obtain multiple image blocks; Perform linear projection embedding on the image patch to generate the corresponding first embedding vector; Add positional encoding to the first embedding vector to generate the corresponding second embedding vector, wherein the positional encoding includes time axis encoding and frequency axis encoding; The second embedding vector is input into the Transformer encoder to obtain the second feature vector.

[0122] Optionally, in this embodiment of the present disclosure, the aforementioned preset heterogeneous graph attention network includes a residual processing layer, a max pooling layer, a graph module layer, a heterogeneous overlay graph attention layer, and an output layer; the aforementioned audio discrimination module 603 is specifically used for: The residual processing layer processes the fused feature vectors to obtain high-level feature maps; The max pooling layer separates the high-level feature maps, yielding spectral representation vectors and temporal representation vectors; The graph module layer is constructed based on spectral representation vectors and temporal representation vectors to obtain spectral input graphs and temporal input graphs; The heterogeneous overlay map attention layer combines the spectral input map and the temporal input map to obtain a heterogeneous spectral-time map; The output layer predicts the results based on heterogeneous spectral-time maps, obtaining the target feature vector and the predicted audio identification results.

[0123] Optionally, in this embodiment of the disclosure, the graph module layer includes a spectral graph module and a temporal graph module, wherein the spectral graph module includes a spectral graph attention network and a spectral graph pooling layer, and the temporal graph module includes a temporal graph attention network and a temporal graph pooling layer; the audio discrimination module 603 is further configured to: The spectral attention network processes the spectral representation vector to obtain the spectral feature vector; The spectral pooling layer processes the spectral feature vectors to obtain the spectral input map; The time-plot attention network processes the time representation vector to obtain the time feature vector; The temporal graph pooling layer processes the temporal feature vectors to obtain the temporal input graph.

[0124] Optionally, in this embodiment of the disclosure, the output layer includes a readout layer and a fully connected layer; the audio discrimination module 603 is further used for: The readout layer performs feature readout on the heterogeneous spectrum-time map to obtain the first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector; The first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector are concatenated to obtain the target feature vector; The fully connected layer predicts the result based on the target feature vector, thus obtaining the predicted audio identification result.

[0125] Optionally, in this embodiment of the disclosure, the audio identification module 603 is further used for: The readout layer performs a maximum value operation on the spectral nodes in the heterogeneous spectrum-time plot to obtain the first eigenvector. The second eigenvector is obtained by averaging the spectral nodes in the heterogeneous spectrum-time plot. The third eigenvector is obtained by performing a maximum value operation on the time nodes in the heterogeneous spectrum-time plot; The fourth eigenvector is obtained by averaging the time nodes in the heterogeneous spectrum-time plot. The feature vector in the stack node is determined as the fifth feature vector.

[0126] Figure 7 This is an audio identification device illustrated according to some embodiments of the present disclosure, such as... Figure 7 As shown, the device may include: The acquisition module 701 is used to acquire the audio to be judged; Preprocessing module 702 is used to preprocess the audio to be judged to obtain processed audio; The audio discrimination module 703 is used to input the processed audio into the target audio discrimination model to obtain the audio discrimination result.

[0127] like Figure 8 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. The RAM 803 may also store various programs and data required for the operation of the electronic device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0128] Multiple components in electronic device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of displays, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows electronic device 800 to exchange information / data with other electronic devices through computer networks such as the Internet and / or various telecommunications networks.

[0129] The computing unit 801 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the training method for an audio discrimination model. For example, in some embodiments, the training method for an audio discrimination model can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the training method for the audio discrimination model described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform the training method for the audio discrimination model by any other suitable means (e.g., by means of firmware).

[0130] Various embodiments of the apparatuses and techniques described above herein can be implemented in digital electronic circuit devices, integrated circuit devices, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), device-on-a-chip (SoC) devices, payload-programmable logic electronic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable device including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage device, at least one input device, and at least one output device, and transmitting data and instructions to the storage device, the at least one input device, and the at least one output device.

[0131] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0132] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution device, apparatus, or electronic device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor device, apparatus, or electronic device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage electronics, magnetic storage electronics, or any suitable combination of the foregoing.

[0133] To provide interaction with a user, the apparatus and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of apparatus can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0134] The apparatus and techniques described herein can be implemented in computing devices that include backend components (e.g., as a data server), or computing devices that include middleware components (e.g., an application server), or computing devices that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the apparatus and techniques described herein), or computing devices that include any combination of such backend, middleware, or frontend components. The components of the apparatus can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

[0135] Computer devices can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability. Servers can also be distributed server devices or servers incorporating blockchain technology.

[0136] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0137] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for an audio discrimination model, characterized in that, The method comprises: acquiring an audio sample data set and pre-processing audio in the sample data set to obtain a training data set; inputting the audio in the training data set into a preset feature extraction network to obtain a fusion feature vector; inputting the fusion feature vector into a preset heterogeneous graph attention network to obtain a target feature vector; based on the target feature vector, obtaining a corresponding loss value through a learning loss function; updating network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value, and determining the target feature extraction network and the target heterogeneous graph attention network obtained through training as a target audio discrimination model.

2. The method of claim 1, wherein, The pre-processing of the audio in the sample data set to obtain the training data set comprises: performing speech activity detection processing and unification processing on the audio in the sample data set, and determining the processed audio as the training data set.

3. The method of claim 1, wherein, The preset feature extraction network comprises a speech representation learning model and a ViT model; and the inputting of the audio in the training data set into the preset feature extraction network to obtain the fusion feature vector comprises: performing feature extraction on the audio in the training data set in the speech representation learning model to obtain a first feature vector; converting the audio to obtain a corresponding mel spectrum graph; performing feature extraction on the mel spectrum graph in the ViT model to obtain a second feature vector; performing feature fusion on the first feature vector and the second feature vector to obtain the fusion feature vector.

4. The method of claim 3, wherein, The feature extraction on the mel spectrum graph in the ViT model to obtain the second feature vector comprises: performing image segmentation on the mel spectrum graph to obtain a plurality of image blocks; performing linear projection embedding on the image blocks to generate a corresponding first embedding vector; adding position encoding to the first embedding vector to generate a corresponding second embedding vector, wherein the position encoding comprises time axis encoding and frequency axis encoding; inputting the second embedding vector into a Transformer encoder to obtain the second feature vector.

5. The method of claim 1, wherein, The preset heterogeneous graph attention network comprises a residual processing layer, a maximum pooling layer, a graph module layer, a heterogeneous superimposed graph attention layer, and an output layer; and the inputting of the fusion feature vector into the preset heterogeneous graph attention network to obtain the target feature vector comprises: the residual processing layer processes the fusion feature vector to obtain a high-level feature mapping; the maximum pooling layer separates the high-level feature mapping to obtain a spectral representation vector and a time representation vector; the graph module layer is constructed based on the spectral representation vector and the time representation vector to obtain a spectral input graph and a time input graph; the heterogeneous superimposed graph attention layer combines the spectral input graph and the time input graph to obtain a heterogeneous spectral-time graph; the output layer predicts a result based on the heterogeneous spectral-time graph to obtain a target feature vector and a predicted audio discrimination result.

6. The method of claim 5, wherein, The graph module layer includes a spectral graph module and a temporal graph module. The spectral graph module includes a spectral graph attention network and a spectral graph pooling layer, and the temporal graph module includes a temporal graph attention network and a temporal graph pooling layer. The graph module layer is constructed based on the spectral representation vector and the temporal representation vector to obtain a spectral input graph and a temporal input graph, including: The spectral attention network processes the spectral representation vector to obtain the spectral feature vector; The spectral pooling layer processes the spectral feature vector to obtain the spectral input map; The time-map attention network processes the time representation vector to obtain a time feature vector; The time-map pooling layer processes the time feature vector to obtain a time input map.

7. The method of claim 5, wherein, The output layer includes a readout layer and a fully connected layer; the output layer performs result prediction based on the heterogeneous spectral-time map to obtain the target feature vector and the predicted audio discrimination result, including: The readout layer performs feature readout on the heterogeneous spectrum-time map to obtain a first feature vector, a second feature vector, a third feature vector, a fourth feature vector, and a fifth feature vector; The first feature vector, the second feature vector, the third feature vector, the fourth feature vector, and the fifth feature vector are concatenated to obtain the target feature vector; The fully connected layer predicts the result based on the target feature vector to obtain the predicted audio identification result.

8. The method of claim 7, wherein, The readout layer performs feature readout on the heterogeneous spectrum-time map to obtain a first feature vector, a second feature vector, a third feature vector, a fourth feature vector, and a fifth feature vector, including: The readout layer performs a maximum value operation on the spectral nodes in the heterogeneous spectrum-time graph to obtain a first feature vector. The spectral nodes in the heterogeneous spectrum-time plot are averaged to obtain the second feature vector. The third eigenvector is obtained by performing a maximum value operation on the time nodes in the heterogeneous spectrum-time plot. The time nodes in the heterogeneous spectrum-time plot are averaged to obtain the fourth feature vector. The feature vector in the stack node is determined as the fifth feature vector.

9. An audio authentication method characterized by, The method includes: Obtain the audio to be judged; The audio to be judged is preprocessed to obtain the processed audio; The processed audio is input into the target audio discrimination model to obtain the audio discrimination result.

10. A training device for an audio discrimination model, characterized in that, The device includes: The preprocessing module is used to acquire an audio sample dataset and preprocess the audio in the sample dataset to obtain a training dataset. The feature extraction module is used to input the audio from the training dataset into a preset feature extraction network to obtain a fused feature vector; An audio identification module is used to input the fused feature vector into a preset heterogeneous graph attention network to obtain a target feature vector. The processing module is used to obtain the corresponding loss value based on the target feature vector by learning a loss function; The update module is used to update the network parameters in the preset feature extraction network and the preset heterogeneous graph attention network based on the loss value, and to determine the target feature extraction network and target heterogeneous graph attention network obtained through training as the target audio discrimination model.

11. An audio authentication device, characterized by The device includes: The acquisition module is used to acquire the audio to be judged; The preprocessing module is used to preprocess the audio to be judged to obtain the processed audio; The audio discrimination module is used to input the processed audio into the target audio discrimination model to obtain the audio discrimination result.

12. A non-transitory computer-readable storage medium storing computer-executable instructions; wherein the computer-executable instructions, when executed by a processor, are capable of implementing the method of any one of claims 1-8 or 9.

13. A computer program product comprising a computer program that, when executed by a processor, implements the method of any one of claims 1-8 or 9.

Images