Model training method, audio classification method, device, medium and program product

Abstract

The application provides a model training method, an audio classification method, a device, a medium and a program product, and mainly relates to machine learning technology in the field of artificial intelligence. The training method comprises the following steps: obtaining a first audio, actual classification results and actual position encoding results of the first audio in multiple classification dimensions; inputting the first audio into a target neural network model to obtain predicted classification results and predicted position encoding results of the first audio in the multiple classification dimensions; obtaining a classification loss according to the actual classification results and the predicted classification results; fusing the actual position encoding results of the first audio in the multiple classification dimensions to obtain actual fusion results, and fusing the predicted position encoding results of the first audio in the multiple classification dimensions to obtain predicted fusion results; obtaining a position encoding loss according to the actual fusion results and the predicted fusion results; and training the target neural network model according to the classification loss and the position encoding loss, so that the classification accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence (AI) technology, and in particular to a model training method, an audio classification method, an apparatus, a medium, and a program product. Background Technology

[0002] Effective audio classification typically leads to better audio recommendations. Current audio classification methods are based on neural network models. For example, Figure 1 A schematic diagram of an audio classification method provided for related technologies, such as... Figure 1 As shown, assuming a given audio clip is used as input to a neural network model, the desired output is a classification result of the audio clip across five dimensions: voice, emotion, language, style, and scene. The training process of this neural network model includes: first, extracting a basic feature vector 0 through the neural network layers; then, further extracting features from feature vector 0 using detectors corresponding to different classification dimensions, obtaining feature vectors 1 to 5 to better represent the features for each classification dimension; further, processing the corresponding feature vectors using classifiers corresponding to different classification dimensions to obtain the predicted classification results for the audio clip across each dimension; finally, calculating the losses between the predicted and actual classification results for each dimension, which are loss values ​​1 to 5 respectively, and optimizing the model parameters through gradient backpropagation.

[0003] However, current model training methods suffer from low training accuracy, which leads to inaccurate audio classification results, especially when there are many classification dimensions, which can easily cause confusion between different classification dimensions. Summary of the Invention

[0004] This application provides a model training method, an audio classification method, an apparatus, a medium, and a program product, which can improve audio classification results.

[0005] In a first aspect, an embodiment of the present application provides a model training method, comprising: obtaining a first audio, actual classification results of the first audio in a plurality of classification dimensions, and actual position encoding results of the first audio in the plurality of classification dimensions; inputting the first audio into a target neural network model to obtain predicted classification results of the first audio in the plurality of classification dimensions and predicted position encoding results of the first audio in the plurality of classification dimensions; obtaining a classification loss according to the actual classification results and the predicted classification results of the first audio in the plurality of classification dimensions; fusing the actual position encoding results of the first audio in the plurality of classification dimensions to obtain actual fusion results, and fusing the predicted position encoding results of the first audio in the plurality of classification dimensions to obtain predicted fusion results; obtaining a position encoding loss according to the actual fusion results and the predicted fusion results; and training the target neural network model according to the classification loss and the position encoding loss.

[0006] In a second aspect, an embodiment of the present application provides an audio classification method, comprising: obtaining a second audio; and inputting the second audio into a target neural network model trained by the method provided in the first aspect to obtain predicted classification results of the second audio in a plurality of classification dimensions.

[0007] In a third aspect, a model training apparatus is provided, comprising: an obtaining module, an inputting module, a first calculating module, a fusing module, a second calculating module, and a training module. The obtaining module is configured to obtain a first audio, actual classification results of the first audio in a plurality of classification dimensions, and actual position encoding results of the first audio in the plurality of classification dimensions. The inputting module is configured to input the first audio into a target neural network model to obtain predicted classification results of the first audio in the plurality of classification dimensions and predicted position encoding results of the first audio in the plurality of classification dimensions. The first calculating module is configured to obtain a classification loss according to the actual classification results and the predicted classification results of the first audio in the plurality of classification dimensions. The fusing module is configured to fuse the actual position encoding results of the first audio in the plurality of classification dimensions to obtain actual fusion results, and fuse the predicted position encoding results of the first audio in the plurality of classification dimensions to obtain predicted fusion results. The second calculating module is configured to obtain a position encoding loss according to the actual fusion results and the predicted fusion results. The training module is configured to train the target neural network model according to the classification loss and the position encoding loss.

[0008] In a fourth aspect, an audio classification apparatus is provided, comprising: an obtaining module and an inputting module. The obtaining module is configured to obtain a second audio. The inputting module is configured to input the second audio into a target neural network model trained by the method provided in the first aspect to obtain predicted classification results of the second audio in a plurality of classification dimensions.

[0009] In a fifth aspect, an electronic device is provided, and includes a processor and a memory. The memory is configured to store a computer program, and the processor is configured to invoke and run the computer program stored in the memory to execute the method in any one of the first aspect to the second aspect or the implementation manners thereof.

[0010] In a sixth aspect, a computer readable storage medium is provided, and is configured to store a computer program. When the computer program is run on a computer, the computer is caused to execute the method in any one of the first aspect to the second aspect or the implementation manners thereof.

[0011] In a seventh aspect, a computer program product is provided, and includes computer program instructions. When the computer program instructions are run on a computer, the computer is caused to execute the method in any one of the first aspect to the second aspect or the implementation manners thereof.

[0012] Through the technical solution provided in the present application, the target neural network model can learn the fusion result of the first audio in multiple classification dimensions, so that the classification result of the output audio in multiple classification dimensions in the prediction stage of the neural network model is no longer isolated from each other. In other words, compared with the related art, the target neural network model additionally learns the fusion result of the first audio in multiple classification dimensions, so that the classification accuracy is higher. Especially for the case of multiple classification dimensions, since the classification result of the audio in multiple classification dimensions is no longer isolated from each other, the probability of confusion between different classification dimensions can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0014] Figure 1 A schematic diagram of an audio classification method provided for the related art;

[0015] Figure 2 A schematic diagram of a system architecture related to an embodiment of the present application;

[0016] Figure 3 A flowchart of a model training method provided by an embodiment of the present application;

[0017] Figure 4 A schematic diagram of a target neural network model provided by an embodiment of the present application;

[0018] Figure 5A schematic diagram of a position encoding module provided by an embodiment of the present application is shown in FIG. 1.

[0019] Figure 6 A flowchart of an audio classification method provided by an embodiment of the present application is shown in FIG. 2.

[0020] Figure 7 A schematic diagram of a model training device 700 provided by an embodiment of the present application is shown in FIG. 3.

[0021] Figure 8 A schematic diagram of an audio classification device 800 provided by an embodiment of the present application is shown in FIG. 4.

[0022] Figure 9 A schematic block diagram of an electronic device 900 provided by an embodiment of the present application is shown in FIG. 5. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0024] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server including a series of steps or units does not necessarily have to include only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0025] AI is a theory, method, technology and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0026] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software level technology. Artificial intelligence basic technology generally includes, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning and other major directions.

[0027] The present application mainly relates to the field of machine learning (ML) technology in artificial intelligence, wherein machine learning is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a special study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0028] As described above, the current model training method has the problem of low model training accuracy, which further leads to inaccurate audio classification results, especially for the case of multiple classification dimensions, which is easy to cause confusion between different classification dimensions.

[0029] To solve the above technical problems, the present application encodes the positions of the audio in multiple classification dimensions during the model training process, and fuses these position encodings, and trains the neural network model based on the fused information, so that the neural network model can learn the fusion information of the audio in multiple classification dimensions, and then the classification results of the audio in multiple classification dimensions output by the neural network model in the prediction stage are no longer isolated, so that the classification accuracy is higher.

[0030] In some embodiments, the system architecture of the present application embodiment is as shown in Figure 2

[0031] Figure 2 For a system architecture diagram related to the present application embodiment, the user equipment 101, the data acquisition equipment 102, the training equipment 103, the execution equipment 104, the database 105 and the content library 106.

[0032] ​The data collection device 102 is configured to read training data from the content library 106 and store the read training data into the database 105. The training data according to the embodiments of the present application comprises a first audio, actual classification results of the first audio in multiple classification dimensions, and actual position encoding results of the first audio in the multiple classification dimensions. The training data according to the embodiments of the present application can be at least one. The first audio can be any audio, such as a certain music. The first audio is named as such to distinguish from the audio in the prediction stage.

[0033] The training device 103 is configured to train the target neural network model based on the training data maintained in the database 105, so that the trained target neural network model can output predicted classification results of the first audio in the multiple classification dimensions and predicted position encoding results of the first audio in the multiple classification dimensions. The trained target neural network model can be applied to different systems or devices.

[0034] In addition, referring to Figure 2 The execution device 104 is configured with an I / O interface 107 to interact with external devices. For example, the I / O interface receives the second audio sent by the user device 101. The second audio refers to any audio to be predicted, such as a certain music to be predicted. The second audio is named as such to distinguish from the audio in the training stage. The computing module 109 in the execution device 104 processes the input second audio using the trained target neural network model, outputs predicted classification results of the second audio in the multiple classification dimensions, and sends the corresponding results to the user device 101 through the I / O interface.

[0035] The user device 101 can include a mobile phone, a tablet computer, a notebook computer, a palm computer, a mobile internet device (MID), or other terminal devices with a browser function.

[0036] The execution device 104 can be a server.

[0037] For example, the server can be a rack server, a blade server, a tower server, or a cabinet server, etc. The server can be a standalone test server, or a test server cluster composed of multiple test servers.

[0038] In this embodiment, the execution device 104 is connected with the user device 101 through a network. The network can be an Intranet, the Internet, a Global System of Mobile communication (GSM), a Wideband Code Division Multiple Access (WCDMA), a 4G network, a 5G network, Bluetooth, Wi-Fi, a telecommunication network, or other wireless or wired networks.

[0039] It should be noted that, Figure 2 The positional relationship between the devices, elements, modules, etc. shown in the figure does not constitute any limitation. In some embodiments, the data acquisition device 102, the user device 101, the training device 103, and the execution device 104 can be the same device. The database 105 can be distributed on one server or multiple servers, and the content library 106 can be distributed on one server or multiple servers.

[0040] The technical solutions of the embodiments of the present application will be described in detail in some embodiments. The following embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0041] First, the model training process related to the embodiments of the present application will be described. Figure 3 The model training process related to the embodiments of the present application will be described.

[0042] Figure 3 The flowchart of the model training method provided by an embodiment of the present application. The method can be executed by a training device, such as a server, as shown in the figure, the method can include: Figure 3

[0043] S310: obtaining a first audio, an actual classification result of the first audio in multiple classification dimensions, and an actual position encoding result of the first audio in multiple classification dimensions;

[0044] S320: inputting the first audio into a target neural network model to obtain a predicted classification result of the first audio in multiple classification dimensions and a predicted position encoding result of the first audio in multiple classification dimensions;

[0045] S330: obtaining a classification loss according to the actual classification result and the predicted classification result of the first audio in multiple classification dimensions;

[0046] ​S340: Fuse the actual position encoding results of the first audio on the multiple classification dimensions to obtain an actual fusion result, and fuse the predicted position encoding results of the first audio on the multiple classification dimensions to obtain a predicted fusion result;

[0047] S350: Obtain a position encoding loss according to the actual fusion result and the predicted fusion result;

[0048] S360: Train the target neural network model according to the classification loss and the position encoding loss.

[0049] Optionally, the training device can obtain the first audio, the actual classification results of the first audio on the multiple classification dimensions, and the actual position encoding results of the first audio on the multiple classification dimensions from a database. The database can be distributed on one server or multiple servers.

[0050] Optionally, the number of the first audios can be multiple. It should be understood that, for each first audio, the first audio, the actual classification results of the first audio on the multiple classification dimensions, and the actual position encoding results of the first audio on the multiple classification dimensions constitute a training sample. In the embodiments of the present application, the model training process is exemplarily illustrated by taking training the target neural network model by one training sample as an example.

[0051] Optionally, the multiple classification dimensions include at least two of the following: a vocal classification, an emotion classification, a language classification, a style classification, and a scene classification, but are not limited thereto.

[0052] Optionally, the vocal classification can include the following categories, but is not limited thereto: an audio category with a vocal and an audio category without a vocal, for example, music without a vocal, i.e., pure music belongs to the audio category without a vocal, and the like.

[0053] Optionally, the emotion classification can include the following categories, but is not limited thereto: a happy category, a sad category, and the like.

[0054] Optionally, the language classification can include the following categories, but is not limited thereto: a Chinese category, an English category, a German category, and the like.

[0055] Optionally, the style classification can include the following categories, but is not limited thereto: a rock category, a ballad category, a hip-hop category, and the like.

[0056] Optionally, the scene classification can include the following categories, but is not limited thereto: a concert category, a recording studio category, and the like.

[0057] Optionally, the actual classification result of the first audio on a certain classification dimension can be one-hot. Each element in the one-hot indicates whether the first audio belongs to a certain category.

[0058] For example, the actual classification result of the first audio in the human voice classification dimension is (1, 0), the first element 1 indicates that the first audio belongs to the audio category with human voice, and the second element 0 indicates that it does not belong to the pure music category.

[0059] For example, the actual classification result of the first audio in the emotion classification dimension is (1, 0), the first element 1 indicates that the first audio belongs to the happy category, and the second element 0 indicates that it does not belong to the sad category.

[0060] For example, the actual classification result of the first audio in the language classification dimension is (1, 0, 0), the first element 1 indicates that the first audio belongs to the Chinese category, the second element 0 indicates that it does not belong to the English category, and the third element 0 indicates that it does not belong to the German category.

[0061] For example, the actual classification result of the first audio in the style classification dimension is (1, 0, 0), the first element 1 indicates that the first audio belongs to the rock category, the second element 0 indicates that it does not belong to the ballad category, and the third element 0 indicates that it does not belong to the hip-hop category.

[0062] For example, the actual classification result of the first audio in the scene classification dimension is (0, 1), the first element 0 indicates that the first audio does not belong to the concert category, and the second element 1 indicates that it belongs to the recording studio category.

[0063] Optionally, the predicted classification result of the first audio in a certain classification dimension can be the classification result obtained by taking the softmax or sigmoid, that is, the predicted classification result of the first audio in a certain classification dimension can be an N-dimensional vector, each element of the N-dimensional vector represents the probability of the first audio belonging to the category corresponding to the element, and N represents the number of categories included in the classification dimension.

[0064] For example, the predicted classification result of the first audio in the human voice classification dimension is (0.6, 0.4), the first element 0.6 indicates that the probability of the first audio belonging to the audio category with human voice is 0.6, and the second element 0 indicates that the probability of belonging to the pure music category is 0.4.

[0065] For example, the predicted classification result of the first audio in the emotion classification dimension is (0.7, 0.3), the first element 0.7 indicates that the probability of the first audio belonging to the happy category is 0.7, and the second element 0.3 indicates that the probability of belonging to the sad category is 0.3.

[0066] For example, the predicted classification result of the first audio in the language classification dimension is (0.6, 0.2, 0.2), the first element 1 indicates that the probability of the first audio belonging to the Chinese category is 0.6, the second element 0.2 indicates that the probability of belonging to the English category is 0.2, and the third element 0.2 indicates that the probability of belonging to the German category is 0.2.

[0067] For example, the predicted classification result of the first audio in the style classification dimension is (0.7, 0.2, 0.1), the first element 1 indicates that the probability of the first audio belonging to the rock category is 0.7, the second element 0.2 indicates that the probability of belonging to the ballad category is 0.2, and the third element 0 indicates that the probability of belonging to the hip-hop category is 0.1.

[0068] For example, the predicted classification result of the first audio in the scene classification dimension is (0.2, 0.8), the first element 0.2 indicates that the probability of the first audio belonging to the concert category is 0.2, and the second element 0.8 indicates that the probability of belonging to the studio category is 0.8.

[0069] Optionally, the actual position encoding result of the first audio in each classification dimension in the plurality of classification dimensions can be one-hot encoding, and each one-hot encoding can be an M-dimensional vector, M representing the number of classification dimensions, wherein each element in the M-dimensional vector respectively indicates whether the classification dimension is in the position corresponding to the element for the first audio.

[0070] For example, the actual position encoding result of the first audio in the vocal classification dimension can be (1, 0, 0, 0, 0), the actual position encoding result in the emotion classification dimension can be (0, 1, 0, 0, 0), the actual position encoding result in the language classification dimension can be (0, 0, 1, 0, 0), the actual position encoding result in the style classification dimension can be (0, 0, 0, 1, 0), and the actual position encoding result in the scene classification dimension can be (0, 0, 0, 0, 1). From these position encoding results, it can be known that for the first audio, the vocal classification dimension is actually in the first position, the emotion classification dimension is in the second position, the language classification dimension is in the third position, the style classification dimension is in the fourth position, and the scene classification dimension is in the fifth position.

[0071] Optionally, the predicted position encoding result of the first audio in each classification dimension in the plurality of classification dimensions is the classification result obtained by taking the softmax or sigmoid, that is, the predicted position encoding result of the first audio in a certain classification dimension can be an M-dimensional vector, and each element in the M-dimensional vector respectively indicates the probability of the classification dimension in the position corresponding to the element for the first audio.

[0072] For example, the prediction position encoding result of the first audio on the vocal classification dimension can be (0.6, 0.2, 0.1, 0.1, 0), the prediction position encoding result on the emotion classification dimension can be (0.2, 0.8, 0, 0, 0), the prediction position encoding result on the language classification dimension can be (0.1, 0.1, 0.7, 0.1, 0), the prediction position encoding result on the style classification dimension can be (0.4, 0.1, 0, 5, 0, 0), and the prediction position encoding result on the scene classification dimension can be (0, 0, 0, 0.1, 0.9), that is, it is predicted that the vocal classification dimension is in the first position, the emotion classification dimension is in the second position, the language classification dimension is in the third position, the style classification dimension is in the fourth position, and the scene classification dimension is in the fifth position for the first audio according to the position encoding results.

[0073] Optionally, Figure 4 A schematic diagram of a target neural network model provided by an embodiment of the present application is shown in FIG. 1, which can include a neural network layer, a plurality of classification detection heads corresponding to a plurality of classification dimensions respectively, a plurality of classifiers corresponding to the plurality of classification dimensions respectively, and a plurality of position encoding modules corresponding to the plurality of classification dimensions respectively. The neural network layer is configured to perform feature extraction on the first audio to obtain a first feature vector, denoted as feature vector 0. Each classification detection head is configured to perform feature extraction on the first feature vector to obtain a second feature vector on the corresponding classification dimension, denoted as feature vector 1 to feature vector 5 respectively. Each classifier is configured to process the corresponding second feature vector to obtain a prediction classification result on the corresponding classification dimension. Each position encoding module is configured to perform position encoding on the second feature vector on the corresponding classification dimension to obtain a prediction position encoding result on the corresponding classification dimension. Figure 4

[0074] Optionally, the neural network layer can include at least one convolutional neural network (CNN) layer, or the neural network layer can be a Transfomer, etc., which is not limited in the present application.

[0075] It should be understood that the plurality of classification detection heads corresponding to the plurality of classification dimensions respectively involve different parameter values, so that they can obtain the second feature vector on the corresponding classification dimension.

[0076] It should be understood that the classification results obtained by the plurality of classifiers corresponding to the plurality of classification dimensions respectively can also be obtained by taking the softmax or sigmoid to obtain the classification results in the interval of 0 to 1.

[0077] Optionally, Figure 5 A schematic diagram of a position encoding module provided by an embodiment of the present application is shown in FIG. 2, which can include a plurality of position encoding heads corresponding to a plurality of classification dimensions respectively. Each position encoding head is configured to perform position encoding on the second feature vector on the corresponding classification dimension to obtain a prediction position encoding result on the corresponding classification dimension.​Figure 5 As shown, each position encoding module sequentially comprises a residual layer, a first fully connected layer, and a second fully connected layer. It should be noted that the number of fully connected layers included in the position encoding module is not limited by the present application. For example, the position encoding module can include one fully connected layer, or can include two fully connected layers as shown, or can even include more fully connected layers. Figure 5 As shown, the position encoding module includes two fully connected layers, or can even include more fully connected layers.

[0078] Optionally, the training device can fuse the actual position encoding results of the first audio in the plurality of classification dimensions and fuse the predicted position encoding results of the first audio in the plurality of classification dimensions in any of the following implementable manners, but not limited thereto:

[0079] Implementable manner one, the training device splices the actual position encoding results of the first audio in the plurality of classification dimensions according to a preset order of the plurality of classification dimensions to obtain an actual fusion result, and splices the predicted position encoding results of the first audio in the plurality of classification dimensions according to the preset order of the plurality of classification dimensions to obtain a predicted fusion result.

[0080] Optionally, the preset order can be predefined or indicated by other devices to the training device, or can be a preset order negotiated by the training device and other devices, which is not limited by the present application.

[0081] It should be understood that, as described above, the actual position encoding result of the first audio in each classification dimension can be an M-dimensional vector, based on which the training device splicing the actual position encoding results of the first audio in the plurality of classification dimensions means splicing the M-dimensional vectors corresponding to the M classification dimensions according to the preset order to form an actual position encoding result including M*M elements. Similarly, the training device splicing the predicted position encoding results of the first audio in the plurality of classification dimensions means splicing the M-dimensional vectors corresponding to the M classification dimensions according to the preset order to form a predicted position encoding result including M*M elements.

[0082] For example, assume that the multiple classification dimensions include: vocal classification, emotion classification, language classification, style classification, and scene classification. Their preset order is vocal classification > emotion classification > language classification > style classification > scene classification, where > means preceding. Assume that the actual position encoding result of the first audio in the vocal classification dimension can be (1, 0, 0, 0, 0), the actual position encoding result in the emotion classification dimension can be (0, 1, 0, 0, 0), the actual position encoding result in the language classification dimension can be (0, 0, 1, 0, 0), the actual position encoding result in the style classification dimension can be (0, 0, 0, 1, 0), and the actual position encoding result in the scene classification dimension can be (0, 0, 0, 0, 1). Then the actual fusion result obtained according to the preset order is (1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1). Assume that the predicted position encoding result of the first audio in the vocal classification dimension can be (0.6, 0.2, 0.1, 0.1, 0), the predicted position encoding result in the emotion classification dimension can be (0.2, 0.8, 0, 0, 0), the predicted position encoding result in the language classification dimension can be (0.1, 0.1, 0.7, 0.1, 0), the predicted position encoding result in the style classification dimension can be (0.4, 0.1, 0, 5, 0, 0), and the predicted position encoding result in the scene classification dimension can be (0, 0, 0, 0.1, 0.9). Then the predicted fusion result obtained according to the preset order is (0.6, 0.2, 0.1, 0.1, 0, 0.2, 0.8, 0, 0, 0, 0.1, 0.1, 0.7, 0.1, 0, 0.4, 0.1, 0, 5, 0, 0, 0, 0, 0, 0.1, 0.9).

[0083] In a second implementation manner, the training device sums the actual position encoding results of the first audio in the multiple classification dimensions to obtain an actual fusion result, and sums the predicted position encoding results of the first audio in the multiple classification dimensions to obtain a predicted fusion result.

[0084] For example, assume that the multiple classification dimensions include: vocal classification, emotion classification, language classification, style classification, and scene classification. Their preset order is vocal classification > emotion classification > language classification > style classification > scene classification, where > means preceding. Assume that the actual position encoding result of the first audio in the vocal classification dimension can be (1, 0, 0, 0, 0), the actual position encoding result of the first audio in the emotion classification dimension can be (0, 1, 0, 0, 0), the actual position encoding result of the first audio in the language classification dimension can be (0, 0, 1, 0, 0), the actual position encoding result of the first audio in the style classification dimension can be (0, 0, 0, 1, 0), and the actual position encoding result of the first audio in the scene classification dimension can be (0, 0, 0, 0, 1). Then the actual fusion result obtained according to the preset order is (1, 1, 1, 1, 1). Assume that the predicted position encoding result of the first audio in the vocal classification dimension can be (0.6, 0.2, 0.1, 0.1, 0), the predicted position encoding result of the first audio in the emotion classification dimension can be (0.2, 0.8, 0, 0, 0), the predicted position encoding result of the first audio in the language classification dimension can be (0.1, 0.1, 0.7, 0.1, 0), the predicted position encoding result of the first audio in the style classification dimension can be (0.4, 0.1, 0, 5, 0, 0), and the predicted position encoding result of the first audio in the scene classification dimension can be (0, 0, 0, 0.1, 0.9). Then the predicted fusion result obtained according to the preset order is (1.3, 1.2, 1.4, 0.3, 0.9).

[0085] Optionally, after the training device obtains the actual classification result and the predicted classification result of the first audio in the multiple classification dimensions, the actual classification result and the predicted classification result of the first audio in each classification dimension can be substituted into the first loss function to obtain a classification loss in the classification dimension.

[0086] Optionally, the first loss function can be a cross-entropy loss function, but is not limited thereto.

[0087] The cross-entropy loss function can be as follows:

[0088]

[0089] For a certain classification dimension, x represents a certain category under the classification dimension, q(x) represents the predicted occurrence probability of the first audio in the x category, i.e., one of the elements in the predicted classification result of the first audio in the classification dimension, which corresponds to the x classification, and p(x) represents the actual occurrence probability of the first audio in the x category, i.e., one of the elements in the actual classification result of the first audio in the classification dimension, which corresponds to the x category.

[0090] For example, the actual classification result of the first audio in the human voice classification dimension is (1, 0), and the predicted classification result of the first audio in the human voice classification dimension is (0.6, 0.4), then the classification loss of the actual classification result and the predicted classification result of the first audio in the human voice classification dimension is H1 = -1*log 0.6 - 0*log 0.4 = -log 0.6.

[0091] For example, the actual classification result of the first audio in the emotion classification dimension is (1, 0), and the predicted classification result of the first audio in the emotion classification dimension is (0.7, 0.3), then the classification loss of the actual classification result and the predicted classification result of the first audio in the emotion classification dimension is H2 = -1*log 0.7 - 0*log 0.3 = -log 0.7.

[0092] For example, the actual classification result of the first audio in the language classification dimension is (1, 0, 0), and the predicted classification result of the first audio in the language classification dimension is (0.6, 0.2, 0.2), then the classification loss of the actual classification result and the predicted classification result of the first audio in the language classification dimension is H3 = -1*log 0.6 - 0*log 0.2 - 0*log 0.2 = -log 0.6.

[0093] For example, the actual classification result of the first audio in the style classification dimension is (1, 0, 0), and the predicted classification result of the first audio in the style classification dimension is (0.7, 0.2, 0.1), then the classification loss of the actual classification result and the predicted classification result of the first audio in the style classification dimension is H4 = -1*log 0.7 - 0*log 0.2 - 0*log 0.2 = -log 0.7.

[0094] For example, the actual classification result of the first audio in the scene classification dimension is (0, 1), and the predicted classification result of the first audio in the scene classification dimension is (0.2, 0.8), then the classification loss of the actual classification result and the predicted classification result of the first audio in the scene classification dimension is H5 = -0*log 0.2 - 1*log 0.8 = -log 0.8.

[0095] Optionally, after the training device obtains the actual fusion result and the predicted fusion result, the actual fusion result and the predicted fusion result can be substituted into the second loss function to obtain a position encoding loss.

[0096] Optionally, the second loss function can be a cross-entropy loss function, but is not limited thereto.

[0097] The cross-entropy loss function can be as follows:

[0098]

[0099] wherein i represents the serial number of the i-th element in the fusion result, q i represents the i-th element in the predicted fusion result, p i represents the i-th element in the actual fusion result.

[0100] For example, the actual fusion result of the first audio is (1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1), and the predicted fusion result is (0.6, 0.2, 0.1, 0.1, 0, 0.2, 0.8, 0, 0, 0, 0.1, 0.1, 0.7, 0.1, 0, 0.4, 0.1, 0, 5, 0, 0, 0, 0, 0, 0.1, 0.9). Then the position encoding loss obtained by the training device can be:

[0101] loss = -1*log 0.6 -1*log 0.8 -1*log 0.7 -1*log 0.9

[0102] Optionally, after the training device obtains the position encoding loss of the first audio and the classification loss of the first audio in the classification dimension, the target neural network model can be trained by any of the following implementation manners, but is not limited thereto:

[0103] In one implementation manner, the training device calculates the sum of the classification loss and the position encoding loss to obtain a target loss; and trains the target neural network model according to the target loss.

[0104] For example, the target loss H = H1 + H2 + H3 + H4 + H5 + loss, wherein H1 represents the classification loss of the first audio in the vocal classification dimension, H2 represents the classification loss of the first audio in the emotion classification dimension, H3 represents the classification loss of the first audio in the language classification dimension, H4 represents the classification loss of the first audio in the style classification dimension, H5 represents the classification loss of the first audio in the scene classification dimension, and loss represents the position encoding loss.

[0105] After the training device obtains the target loss, the parameters of the target neural network model can be optimized by gradient backpropagation, so as to realize the training of the target neural network model.

[0106] In another implementation manner, the training device can weight and sum the classification loss and the position encoding loss to obtain a target loss; and train the target neural network model according to the target loss.

[0107] For example, the target loss H = ω1H1+ ω2H2+ ω3H3+ ω4H4+ ω5H5+ ω6loss, wherein H1 represents the classification loss of the first audio in the human voice classification dimension, H2 represents the classification loss of the first audio in the emotion classification dimension, H3 represents the classification loss of the first audio in the language classification dimension, H4 represents the classification loss of the first audio in the style classification dimension, H5 represents the classification loss of the first audio in the scene classification dimension, loss represents the position encoding loss, and ω1, ω2, ω3, ω4, ω5 and ω6 represent weight coefficients, wherein the values of the weight coefficients are in the range of 【0, 1】, and the sum of the weight coefficients is 1.

[0108] After the training device obtains the target loss, the parameters of the target neural network model can be optimized through gradient backpropagation, so as to realize the training of the target neural network model.

[0109] Through the technical solutions provided in the present application, the training device can input the first audio into the target neural network model to obtain the predicted classification results of the first audio in multiple classification dimensions and the predicted position encoding results of the first audio in multiple classification dimensions; the classification loss is obtained according to the actual classification results and the predicted classification results of the first audio in multiple classification dimensions; the actual fusion results are obtained by fusing the actual position encoding results of the first audio in multiple classification dimensions, and the predicted fusion results are obtained by fusing the predicted position encoding results of the first audio in multiple classification dimensions; the position encoding loss is obtained according to the actual fusion results and the predicted fusion results; and the target neural network model is trained according to the classification loss and the position encoding loss. Thus, the target neural network model can learn the fusion results of the first audio in multiple classification dimensions, so that the classification results of the audio output by the neural network model in the prediction stage are no longer isolated from each other, in other words, for the target neural network model, it additionally learns the fusion results of the first audio in multiple classification dimensions, so that the classification accuracy is higher. Especially for the case of multiple classification dimensions, since the classification results of the audio in multiple classification dimensions are no longer isolated from each other, the probability of confusion between different classification dimensions can be reduced.

[0110] Figure 6 A flowchart of an audio classification method provided by an embodiment of the present application, which can be executed by an execution device, is shown in Figure 6 The method can include the features that it comprises:

[0111] S610: obtaining a second audio;

[0112] S620: inputting the second audio into the target neural network model trained by the above-mentioned model training method to obtain the predicted classification results of the second audio in multiple classification dimensions.

[0113] It should be understood that the second audio refers to any audio to be predicted, such as certain music to be predicted, and is referred to as the second audio in order to distinguish from the audio in the training stage.

[0114] Optionally, the plurality of classification dimensions include at least two of the following: a vocal classification, an emotion classification, a language classification, a style classification, and a scene classification, but are not limited thereto.

[0115] It should be understood that the explanations of the various classification dimensions and the target neural network model can refer to the above, and the present application will not repeat them.

[0116] Optionally, the predicted classification result of the second audio in a certain classification dimension can be a classification result obtained by taking a softmax or a sigmoid, that is, the predicted classification result of the second audio in a certain classification dimension can be an N-dimensional vector, each element of the N-dimensional vector representing a probability that the second audio belongs to a category corresponding to the element, and N representing a number of categories included in the classification dimension.

[0117] According to the technical solutions provided in the present application, since the neural network model can learn the fusion result of the audio in multiple classification dimensions, the classification results of the second audio in multiple classification dimensions output by the neural network model in the prediction stage are no longer isolated from each other, so that the classification accuracy is higher. Especially for the case of a large number of classification dimensions, since the classification results of the audio in multiple classification dimensions are no longer isolated from each other, the probability of confusion between different classification dimensions can be reduced.

[0118] Figure 7 A schematic diagram of a model training device 700 provided in an embodiment of the present application is shown in FIG. 7. Figure 7As shown, the apparatus includes 700: an acquisition module 710, an input module 720, a first calculation module 730, a fusion module 740, a second calculation module 750, and a training module 760. The acquisition module 710 is configured to acquire the first audio, the actual classification result of the first audio in the plurality of classification dimensions, and the actual position encoding result of the first audio in the plurality of classification dimensions; the input module 720 is configured to input the first audio into the target neural network model to obtain the predicted classification result of the first audio in the plurality of classification dimensions and the predicted position encoding result of the first audio in the plurality of classification dimensions; the first calculation module 730 is configured to obtain the classification loss according to the actual classification result and the predicted classification result of the first audio in the plurality of classification dimensions; the fusion module 750 is configured to fuse the actual position encoding result of the first audio in the plurality of classification dimensions to obtain an actual fusion result, and fuse the predicted position encoding result of the first audio in the plurality of classification dimensions to obtain a predicted fusion result; the second calculation module 740 is configured to obtain the position encoding loss according to the actual fusion result and the predicted fusion result; and the training module 760 is configured to train the target neural network model according to the classification loss and the position encoding loss.

[0119] Optionally, the fusion module 750 is specifically configured to: splice the actual position encoding result of the first audio in the plurality of classification dimensions according to a preset order of the plurality of classification dimensions to obtain the actual fusion result; and splice the predicted position encoding result of the first audio in the plurality of classification dimensions according to the preset order of the plurality of classification dimensions to obtain the predicted fusion result.

[0120] Optionally, the target neural network model includes: a neural network layer, a classification detection head corresponding to each classification dimension, a classifier corresponding to each classification dimension, and a position encoding module corresponding to each classification dimension; the neural network layer is configured to perform feature extraction on the first audio to obtain a first feature vector; each classification detection head is configured to perform feature extraction on the first feature vector to obtain a second feature vector in the corresponding classification dimension; each classifier is configured to process the corresponding second feature vector to obtain a predicted classification result in the corresponding classification dimension; and each position encoding module is configured to perform position encoding on the second feature vector in the corresponding classification dimension to obtain a predicted position encoding result in the corresponding classification dimension.

[0121] Optionally, each position encoding module sequentially includes: a residual layer, a first full connection layer, and a second full connection layer.

[0122] Optionally, the training module 760 is specifically configured to: calculate the sum of the classification loss and the position encoding loss to obtain a target loss; and train the target neural network model according to the target loss.

[0123] Optionally, the multiple classification dimensions include at least two of the following: a vocal classification, an emotion classification, a language classification, a style classification, and a scene classification.

[0124] It should be understood that the device embodiments and the model training method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, details are not described here. Specifically, Figure 7 The device 700 shown can perform the model training method embodiments, and the foregoing and other operations and / or functions of each module in the device 700 are respectively to realize the corresponding processes in the model training method. For the sake of brevity, details are not described here.

[0125] The device 700 of the embodiments of the present application is described above in conjunction with the functional modules. It should be understood that the functional modules can be implemented in the form of hardware, or in the form of instructions of software, or in the form of a combination of hardware and software modules. Specifically, each step of the model training method embodiments in the embodiments of the present application can be completed by the integrated logic circuit of hardware and / or software instructions in the processor. The steps of the model training method disclosed in the embodiments of the present application can be directly embodied as hardware code processor execution completion, or executed by a combination of hardware and software modules in the code processor. Optionally, the software module can be located in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register, etc. The storage medium is located in the memory, and the processor reads the information in the memory, and combines the hardware to complete the steps in the above model method embodiments.

[0126] Figure 8 A schematic diagram of an audio classification device 800 provided by the embodiments of the present application is shown as Figure 8 The device 800 includes an acquisition module 810 and an input module 820, wherein the acquisition module 810 is configured to acquire a second audio; and the input module 820 is configured to input the second audio into a target neural network model trained by the above model training method to obtain a predicted classification result of the second audio in multiple classification dimensions.

[0127] It should be understood that the device embodiments and the audio classification method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, details are not described here. Specifically, Figure 8 The device 800 shown can perform the audio classification method embodiments, and the foregoing and other operations and / or functions of each module in the device 800 are respectively to realize the corresponding processes in the audio classification method. For the sake of brevity, details are not described here.

[0128] The apparatus 800 of the embodiments of the present application is described above from the perspective of functional modules in combination with the drawings. It should be understood that the functional modules can be implemented in the form of hardware, or in the form of instructions of software, or in the form of a combination of hardware and software modules. Specifically, each step of the audio classification method embodiments in the embodiments of the present application can be completed by integrated logic circuits of hardware in a processor and / or instructions of software. The steps of the audio classification method disclosed in the embodiments of the present application can be directly embodied as hardware decoding processor for execution, or can be executed by a combination of hardware and software modules in the decoding processor. Alternatively, the software module can be located in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps in the above model method embodiments in combination with the hardware thereof.

[0129] Figure 9 FIG. 9 is a schematic block diagram of an electronic device 900 provided by the embodiments of the present application. The electronic device 900 can be a training device or an execution device.

[0130] As shown in FIG. 9, the electronic device 900 can include: Figure 9

[0131] a memory 910 and a processor 920, the memory 910 being configured to store a computer program and transmit the program code to the processor 920. In other words, the processor 920 can call and run the computer program from the memory 910 to implement the method in the embodiments of the present application.

[0132] For example, the processor 920 can be configured to execute the above method embodiments according to the instructions in the computer program.

[0133] In some embodiments of the present application, the processor 920 can include but is not limited to:

[0134] a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, etc.

[0135] In some embodiments of the present application, the memory 910 includes but is not limited to:

[0136] ​The non-volatile memory can be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a Random Access Memory (RAM) used as an external cache. By way of example, and not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0137] In some embodiments of the present application, the computer program can be divided into one or more modules, which are stored in the memory 910 and executed by the processor 920 to complete the method provided by the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the electronic device.

[0138] As shown in Figure 9 The electronic device can further include:

[0139] The transceiver 930 can be connected to the processor 920 or the memory 910.

[0140] The processor 920 can control the transceiver 930 to communicate with other devices, specifically, can send information or data to other devices, or receive information or data sent by other devices. The transceiver 930 can include a transmitter and a receiver. The transceiver 930 can further include an antenna, and the number of antennas can be one or more.

[0141] It should be understood that the various components within the electronic device are connected by a bus system, which includes, in addition to a data bus, a power bus, a control bus, and a state signal bus.

[0142] The application also provides a computer storage medium, which stores a computer program, and the computer program enables a computer to execute the method of the method embodiment when executed by the computer. Alternatively, the application embodiment also provides a computer program product containing instructions, and the instructions enable the computer to execute the method of the method embodiment when executed by the computer.

[0143] When implemented by using software, the computer program product can be implemented in the form of a computer program product in whole or in part. The computer program product includes one or more computer instructions. When loaded and executed by a computer, the computer program instructions produce the flow or function of the embodiment of the application in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available medium can be a magnetic medium (for example, floppy disk, hard disk, magnetic tape), an optical medium (for example, digital video disc (DVD)), or a semiconductor medium (for example, solid state disk (SSD)) and the like.

[0144] Those skilled in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0145] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the division of the above-described device embodiments is only a logical function division, and there can be another division manner for the actual implementation, for example, multiple devices or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different parts can be indirect couplings or communication connections through some interfaces, devices or modules, and can be electrical, mechanical or in other forms.

[0146] The modules explained as separated components can or can not be physically separated, and the components shown as modules can or can not be physical modules, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments. For example, the functional modules in the embodiments of the present application can be integrated into a processing module, or each module can be physically present separately, or two or more modules can be integrated into one module.

[0147] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A model training method, characterized in that, The method comprises the following steps: acquiring a first audio, actual classification results of the first audio in a plurality of classification dimensions, and actual position encoding results of the first audio in the plurality of classification dimensions; inputting the first audio into a target neural network model to obtain predicted classification results of the first audio in the plurality of classification dimensions and predicted position encoding results of the first audio in the plurality of classification dimensions; obtaining a classification loss according to the actual classification results and the predicted classification results of the first audio in the plurality of classification dimensions; fusing the actual position encoding results of the first audio in the plurality of classification dimensions to obtain actual fusion results, and fusing the predicted position encoding results of the first audio in the plurality of classification dimensions to obtain predicted fusion results; obtaining a position encoding loss according to the actual fusion results and the predicted fusion results; training the target neural network model according to the classification loss and the position encoding loss.

2. The method of claim 1, wherein, The fusing the actual position encoding results of the first audio in the plurality of classification dimensions to obtain actual fusion results, and the fusing the predicted position encoding results of the first audio in the plurality of classification dimensions to obtain predicted fusion results comprises: splicing the actual position encoding results of the first audio in the plurality of classification dimensions in a preset order of the plurality of classification dimensions to obtain the actual fusion results; splicing the predicted position encoding results of the first audio in the plurality of classification dimensions in the preset order of the plurality of classification dimensions to obtain the predicted fusion results.

3. The method according to claim 1 or 2, characterized in that, The target neural network model comprises: a neural network layer, a classification detection head corresponding to each of the plurality of classification dimensions, a classifier corresponding to each of the plurality of classification dimensions, and a position encoding module corresponding to each of the plurality of classification dimensions; the neural network layer is configured to extract features of the first audio to obtain a first feature vector; each classification detection head is configured to extract features of the first feature vector to obtain a second feature vector in a corresponding classification dimension; each classifier is configured to process the corresponding second feature vector to obtain a predicted classification result in the corresponding classification dimension; each position encoding module is configured to perform position encoding on the second feature vector in the corresponding classification dimension to obtain a predicted position encoding result in the corresponding classification dimension.

4. The method of claim 3, wherein, Each position encoding module comprises, in sequence: a residual layer, a first full connection layer, and a second full connection layer.

5. The method according to claim 1 or 2, characterized in that, The training the target neural network model according to the classification loss and the position encoding loss comprises: calculating a sum of the classification loss and the position encoding loss to obtain a target loss; training the target neural network model according to the target loss.

6. The method of claim 1 or 2, wherein, The plurality of classification dimensions comprise at least two of the following: a vocal classification, an emotion classification, a language classification, a style classification, and a scene classification.

7. An audio classification method characterized by, The method comprises the following steps: acquiring a second audio; inputting the second audio into the target neural network model trained by the method of any one of claims 1 to 6 to obtain predicted classification results of the second audio in the plurality of classification dimensions.

8. A model training apparatus characterized by comprising: The method comprises the following steps: The acquisition module is configured to acquire a first audio, actual classification results of the first audio in a plurality of classification dimensions, and actual position encoding results of the first audio in the plurality of classification dimensions. The input module is configured to input the first audio into a target neural network model to obtain predicted classification results of the first audio in the plurality of classification dimensions and predicted position encoding results of the first audio in the plurality of classification dimensions. The first calculation module is configured to obtain a classification loss according to the actual classification results and the predicted classification results of the first audio in the plurality of classification dimensions. The fusion module is configured to fuse the actual position encoding results of the first audio in the plurality of classification dimensions to obtain actual fusion results, and fuse the predicted position encoding results of the first audio in the plurality of classification dimensions to obtain predicted fusion results. The second calculation module is configured to obtain a position encoding loss according to the actual fusion results and the predicted fusion results. The training module is configured to train the target neural network model according to the classification loss and the position encoding loss.

9. An audio classification apparatus characterized by comprising: The acquisition module is configured to acquire a second audio. The input module is configured to input the second audio into a target neural network model trained by the method in any one of claims 1 to 6 to obtain predicted classification results of the second audio in the plurality of classification dimensions. The processor and the memory are included.

10. An electronic device, comprising: The memory is configured to store a computer program. The processor is configured to invoke and run the computer program stored in the memory to execute the method in any one of claims 1 to 7.

11. A computer readable storage medium, characterized in that, The computer program is configured to enable a computer to execute the method in any one of claims 1 to 7.

12. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to implement the method in any one of claims 1 to 7.

Images