Model training method, multidimensional scoring method, medium and device

Abstract

The embodiments of the present disclosure provide a model training method, a multidimensional scoring method, a medium and a device, wherein the multidimensional scoring method includes: determining the multidimensional feature attribute data of the target audio work, wherein the multidimensional feature attribute data includes scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data, and then, according to the above multidimensional feature attribute data and the target model, obtaining the multidimensional score of the target audio work, and based on the target audio work, pushing the above multidimensional score. The method of the present disclosure implements the multidimensional scoring of audio works through a model for multidimensional scoring of audio works, and by pushing the multidimensional score, it is convenient for relevant personnel to understand the score of the audio work and improve the user experience. In addition, the embodiments of the present disclosure provide a computing device.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology. More specifically, the embodiments of the present disclosure relate to a model training method, a multidimensional scoring method, a medium and a device. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the present disclosure. No description herein is admitted to be prior art by virtue of its inclusion in this section.

[0003] With the continuous development of technology, users can now realize a variety of functions through applications, such as online karaoke entertainment activities, which has led to the need for users to rate the corresponding audio works.

[0004] In the related technology, when scoring a user's audio work, the scoring is usually based on the dry sound domain attribute data in the audio work, for example, scoring based on pitch data, range data, etc., such as judging whether the pitch data reaches the set pitch, whether the range data coverage reaches the set range, etc.

[0005] However, the above-mentioned scoring method takes fewer scoring factors into consideration, resulting in less professional scoring results and making it difficult to make a comprehensive and scientific distinction between audio works. Summary of the invention

[0006] The present disclosure provides a model training method, a multi-dimensional scoring method, a medium and a device to achieve multi-dimensional scoring of audio works.

[0007] In a first aspect of the embodiments of the present disclosure, a model training method is provided, comprising: determining training data, the training data comprising multidimensional feature attribute data of a plurality of audio works, the multidimensional feature attribute data comprising scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; performing model training based on the above training data to obtain a trained target model, wherein the target model is used to perform multidimensional scoring on the audio works.

[0008] In one embodiment of the present disclosure, before performing model training based on training data to obtain a trained target model, it also includes: performing data cleaning on multiple audio works according to preset cleaning rules based on multidimensional feature attribute data, as well as the duration and generation method of multiple audio works; determining the training data includes: determining the training data based on the audio works obtained after data cleaning, and the training data includes the multidimensional feature attribute data of the audio works obtained after data cleaning.

[0009] In one embodiment of the present disclosure, model training is performed based on training data, including: feature quantization and normalization of the multidimensional feature attribute data of the audio work obtained after data cleaning, feature quantization is used to convert non-numerical attribute data into numerical attribute data, and normalization is used to perform dimensionless processing on the data and convert data of different dimensions to the same dimension; model training is performed based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization.

[0010] In one embodiment of the present disclosure, the audio work is generated after the user interacts with the audio and video application; the preset cleaning rules include cleaning the audio work without dry sound, cleaning the audio work with a score lower than the preset score threshold, cleaning the audio work that is not generated after the user interacts with the audio and video application, and cleaning the audio work with a duration lower than the preset duration threshold.

[0011] In one embodiment of the present disclosure, model training is performed based on training data to obtain a trained target model, including: performing model training based on scoring domain attribute data to obtain a first scoring model; performing model training based on dry voice domain attribute data and the first scoring model to obtain a second scoring model, and performing model training based on accompaniment domain attribute data and the first scoring model to obtain a third scoring model; determining a target model based on the second scoring model and the third scoring model.

[0012] In one embodiment of the present disclosure, model training is performed based on the rating domain attribute data to obtain a first rating model, including: obtaining a preset linear relationship between the work rating variables of the multiple audio works and the rating domain attribute data, the linear relationship including the rating parameter to be solved; performing fitting processing to determine the value of the rating parameter to be solved based on the linear relationship between the work rating variables of the multiple audio works and the rating domain attribute data, and the normal distribution of the work ratings of the multiple audio works; based on the linear relationship between the work rating variables of the multiple audio works and the rating domain attribute data, and the value of the rating parameter to be solved, the first rating model is obtained.

[0013] In one embodiment of the present disclosure, model training is performed based on dry voice domain attribute data and a first scoring model to obtain a second scoring model, including: determining work scores of multiple audio works based on the first scoring model, performing principal component analysis on the work scores of the multiple audio works and the dry voice domain attribute data to obtain multiple principal component parameters; clustering the multiple principal component parameters to obtain clustering results, and obtaining a second scoring model based on the clustering results.

[0014] In one embodiment of the present disclosure, model training is performed based on the accompaniment domain attribute data and the first scoring model to obtain a third scoring model, including: obtaining a linear relationship between work scoring variables of a plurality of preset audio works and the accompaniment domain attribute data, the linear relationship including the accompaniment parameters to be solved; performing fitting processing based on the linear relationship between the work scoring variables of the plurality of audio works and the accompaniment domain attribute data, and the first scoring model, to determine the values ​​of the accompaniment parameters to be solved; obtaining the third scoring model based on the linear relationship between the work scoring variables of the plurality of audio works and the accompaniment domain attribute data, and the values ​​of the accompaniment parameters to be solved.

[0015] In one embodiment of the present disclosure, determining a target model based on the second scoring model and the third scoring model includes: obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; obtaining the target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model, and the target weight coefficient of the third scoring model.

[0016] In one embodiment of the present disclosure, obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model includes: obtaining preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models; determining multiple target models according to the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models; obtaining multiple multidimensional ratings of the multiple audio works based on the training data and the multiple target models; determining the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models according to the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works.

[0017] In a second aspect of the disclosed embodiments, a multidimensional scoring method is provided, comprising: determining multidimensional feature attribute data of a target audio work, wherein the multidimensional feature attribute data comprises scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; obtaining a multidimensional score of the target audio work based on the multidimensional feature attribute data and a target model, wherein the target model is trained according to the model training method provided in the first aspect; and pushing the multidimensional score based on the target audio work.

[0018] In one embodiment of the present disclosure, before obtaining a multidimensional score of a target audio work based on the multidimensional feature attribute data and the target model, it also includes: performing data cleaning on the target audio work according to preset cleaning rules based on the multidimensional feature attribute data, as well as the duration and generation method of the target audio work; determining the multidimensional feature attribute data of the target audio work, including: determining the multidimensional feature attribute data of the target audio work based on the audio work obtained after data cleaning, the multidimensional feature attribute data of the target audio work including the multidimensional feature attribute data of the audio work obtained after data cleaning.

[0019] In one embodiment of the present disclosure, a multidimensional score of a target audio work is obtained based on the multidimensional feature attribute data and the target model, including: performing feature quantization and normalization on the multidimensional feature attribute data of the audio work obtained after data cleaning, wherein feature quantization is used to convert non-numerical attribute data into numerical attribute data, and normalization is used to perform dimensionless processing on the data and convert data of different dimensions to the same dimension; and obtaining a multidimensional score of the target audio work based on the multidimensional feature attribute data and the target model of the audio work obtained after feature quantization and normalization.

[0020] In one embodiment of the present disclosure, the target audio work is generated after the user interacts with the audio and video application; the preset cleaning rules include cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

[0021] In one embodiment of the present disclosure, a multidimensional score of a target audio work is obtained based on multidimensional feature attribute data and a target model, including: obtaining a first score based on the score domain attribute data of the target audio work and a first scoring model, wherein the first scoring model is trained based on the score domain attribute data of multiple audio works; obtaining a second score based on the dry voice domain attribute data of the target audio work, a first score, and a second scoring model, and obtaining a third score based on the accompaniment domain attribute data of the target audio work, a first score, and a third scoring model, wherein the second scoring model is trained based on the dry voice domain attribute data of multiple audio works and the first scoring model, and the third scoring model is trained based on the accompaniment domain attribute data of multiple audio works and the first scoring model; determining the multidimensional score of the target audio work based on the second score, the third score, and the target model, wherein the target model is determined based on the second scoring model and the third scoring model.

[0022] In one embodiment of the present disclosure, a first score is obtained based on the score domain attribute data of a target audio work and a first score model, including: obtaining a linear relationship between a preset score variable of a plurality of audio works and the score domain attribute data, the linear relationship including a score parameter to be solved; performing fitting processing based on the linear relationship between the score variables of the plurality of audio works and the score domain attribute data, and the normal distribution of the score of the plurality of audio works, to determine the value of the score parameter to be solved; obtaining a first score model based on the linear relationship between the score variables of the plurality of audio works and the score domain attribute data, and the value of the score parameter to be solved, and inputting the score domain attribute data of the target audio work into the first score model to obtain a first score.

[0023] In one embodiment of the present disclosure, a second score is obtained based on the dry domain attribute data of the target audio work, the first score and the second score model, including: determining the work scores of multiple audio works based on the first score model, performing principal component analysis on the work scores and the dry domain attribute data of the multiple audio works to obtain multiple principal component parameters; clustering the multiple principal component parameters to obtain clustering results, obtaining a second score model based on the clustering results, and inputting the dry domain attribute data of the target audio work and the first score into the second score model to obtain a second score.

[0024] In one embodiment of the present disclosure, a third score is obtained according to the accompaniment domain attribute data of a target audio work, a first score and a third score model, including: obtaining a linear relationship between a preset plurality of audio work score variables and the accompaniment domain attribute data, the linear relationship including the accompaniment parameters to be solved; performing fitting processing according to the linear relationship between the work score variables and the accompaniment domain attribute data of the plurality of audio works, and the first score model, to determine the values ​​of the accompaniment parameters to be solved; obtaining a third score model based on the linear relationship between the work score variables and the accompaniment domain attribute data of the plurality of audio works, and the values ​​of the accompaniment parameters to be solved, and inputting the accompaniment domain attribute data of the target audio work and the first score into the third score model to obtain a third score.

[0025] In one embodiment of the present disclosure, a multidimensional score of a target audio work is determined based on a second score, a third score and a target model, including: obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; obtaining a target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model, inputting the second score and the third score into the target model to obtain a multidimensional score.

[0026] In one embodiment of the present disclosure, obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model includes: obtaining preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models; determining multiple target models according to the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models; obtaining multiple multidimensional ratings of the multiple audio works based on multidimensional feature attribute data of the multiple audio works and the multiple target models; determining the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model from the preset weight coefficients of the multiple second scoring models and the preset weight coefficients of the multiple third scoring models according to the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works.

[0027] In a third aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, in which computer execution instructions are stored. When a processor executes the computer execution instructions, the model training method provided in the first aspect is implemented, and / or the multidimensional scoring method provided in the second aspect is implemented.

[0028] In a fourth aspect of the embodiments of the present disclosure, a model training device is provided, comprising: a data determination module for determining training data, the training data comprising multi-dimensional feature attribute data of a plurality of audio works, the multi-dimensional feature attribute data comprising scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; a model training module for performing model training according to the above training data to obtain a trained target model, the target model being used for performing multi-dimensional scoring of the audio works.

[0029] In one embodiment of the present disclosure, the data determination module is specifically used to: perform data cleaning on multiple audio works according to multi-dimensional feature attribute data, as well as the duration and generation method of the multiple audio works in accordance with preset cleaning rules; determine training data based on the audio works obtained after data cleaning, and the training data includes the multi-dimensional feature attribute data of the audio works obtained after data cleaning.

[0030] In one embodiment of the present disclosure, the model training module is specifically used to: perform feature quantization and normalization processing on the multidimensional feature attribute data of the audio work obtained after data cleaning, feature quantization is used to convert non-numerical attribute data into numerical attribute data, and normalization is used to perform dimensionless processing on the data and convert data of different dimensions to the same dimension; perform model training based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization processing.

[0031] In one embodiment of the present disclosure, the audio work is generated after the user interacts with the audio and video application; the preset cleaning rules include cleaning the audio work without dry sound, cleaning the audio work with a score lower than the preset score threshold, cleaning the audio work that is not generated after the user interacts with the audio and video application, and cleaning the audio work with a duration lower than the preset duration threshold.

[0032] In one embodiment of the present disclosure, the model training module is specifically used to: perform model training based on scoring domain attribute data to obtain a first scoring model; perform model training based on dry voice domain attribute data and the first scoring model to obtain a second scoring model, and perform model training based on accompaniment domain attribute data and the first scoring model to obtain a third scoring model; determine a target model based on the second scoring model and the third scoring model.

[0033] In one embodiment of the present disclosure, the model training module is specifically used to: obtain a linear relationship between a preset plurality of audio works' work score variables and the score domain attribute data, wherein the linear relationship includes a score parameter to be solved; perform fitting processing based on the linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and the normal distribution of the work scores of the plurality of audio works, to determine the value of the score parameter to be solved; obtain a first score model based on the linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and the value of the score parameter to be solved.

[0034] In one embodiment of the present disclosure, the model training module is specifically used to: determine the work scores of multiple audio works according to the first scoring model, perform principal component analysis on the work scores and dry voice domain attribute data of the multiple audio works to obtain multiple principal component parameters; cluster the multiple principal component parameters to obtain clustering results, and obtain a second scoring model based on the clustering results.

[0035] In one embodiment of the present disclosure, the model training module is specifically used to: obtain a linear relationship between a preset plurality of audio works' score variables and accompaniment domain attribute data, wherein the linear relationship includes the accompaniment parameters to be solved; perform fitting processing based on the linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, and a first scoring model, to determine the values ​​of the accompaniment parameters to be solved; and obtain a third scoring model based on the linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, and the values ​​of the accompaniment parameters to be solved.

[0036] In one embodiment of the present disclosure, the model training module is specifically used to: obtain the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model; obtain the target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model.

[0037] In one embodiment of the present disclosure, the model training module is specifically used to: obtain preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models; determine multiple target models according to the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models; obtain multiple multidimensional ratings of multiple audio works based on training data and multiple target models; determine the target weight coefficients of the second scoring model and the target weight coefficients of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models according to the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works.

[0038] In a fifth aspect of the embodiments of the present disclosure, a multidimensional scoring device is provided, comprising: an attribute determination module, for determining multidimensional feature attribute data of a target audio work, wherein the multidimensional feature attribute data comprises scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; a multidimensional scoring module, for obtaining a multidimensional score of the target audio work according to the multidimensional feature attribute data and a target model, wherein the target model is trained according to the model training method provided in the first aspect; a score pushing module, for pushing the multidimensional score based on the target audio work.

[0039] In one embodiment of the present disclosure, the attribute determination module is specifically used to: perform data cleaning on the target audio work according to the multidimensional feature attribute data, as well as the duration and generation method of the target audio work according to preset cleaning rules; determine the multidimensional feature attribute data of the target audio work according to the audio work obtained after data cleaning, and the multidimensional feature attribute data of the target audio work includes the multidimensional feature attribute data of the audio work obtained after data cleaning.

[0040] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: perform feature quantization and normalization on the multidimensional feature attribute data of the audio work obtained after data cleaning, wherein feature quantization is used to convert non-numerical attribute data into numerical attribute data, and normalization is used to perform dimensionless processing on the data and convert data of different dimensions to the same dimension; obtain a multidimensional score of the target audio work based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization and the target model.

[0041] In one embodiment of the present disclosure, the target audio work is generated after the user interacts with the audio and video application; the preset cleaning rules include cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

[0042] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: obtain a first score based on the scoring domain attribute data of the target audio work and a first scoring model, wherein the first scoring model is trained based on the scoring domain attribute data of multiple audio works; obtain a second score based on the dry domain attribute data of the target audio work, the first score and the second scoring model, and obtain a third score based on the accompaniment domain attribute data of the target audio work, the first score and the third scoring model, wherein the second scoring model is trained based on the dry domain attribute data of multiple audio works and the first scoring model, and the third scoring model is trained based on the accompaniment domain attribute data of multiple audio works and the first scoring model; determine the multidimensional score of the target audio work based on the second score, the third score and the target model, wherein the target model is determined based on the second scoring model and the third scoring model.

[0043] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: obtain a linear relationship between a preset plurality of audio work score variables and score domain attribute data, wherein the linear relationship includes a score parameter to be solved; perform fitting processing based on the linear relationship between the score variables of the plurality of audio works and the score domain attribute data, and the normal distribution of the work scores of the plurality of audio works, to determine the value of the score parameter to be solved; obtain a first scoring model based on the linear relationship between the score variables of the plurality of audio works and the score domain attribute data, and the value of the score parameter to be solved, and input the score domain attribute data of the target audio work into the first scoring model to obtain a first score.

[0044] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: determine the work scores of multiple audio works according to a first scoring model, perform principal component analysis on the work scores and dry domain attribute data of the multiple audio works, and obtain multiple principal component parameters; cluster the multiple principal component parameters to obtain clustering results, obtain a second scoring model based on the clustering results, and input the dry domain attribute data of the target audio work and the first score into the second scoring model to obtain a second score.

[0045] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: obtain a linear relationship between the work scoring variables and the accompaniment domain attribute data of a plurality of preset audio works, wherein the linear relationship includes the accompaniment parameters to be solved; perform fitting processing based on the linear relationship between the work scoring variables and the accompaniment domain attribute data of the plurality of audio works, and a first scoring model, to determine the values ​​of the accompaniment parameters to be solved; obtain a third scoring model based on the linear relationship between the work scoring variables and the accompaniment domain attribute data of the plurality of audio works, and the values ​​of the accompaniment parameters to be solved, and input the accompaniment domain attribute data of the target audio work and the first score into the third scoring model to obtain a third score.

[0046] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: obtain the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model; obtain the target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model, input the second score and the third score into the target model, and obtain a multidimensional score.

[0047] In one embodiment of the present disclosure, the multidimensional scoring module is specifically used to: obtain preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models; determine multiple target models according to the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models; obtain multiple multidimensional scores of multiple audio works based on multidimensional feature attribute data of multiple audio works and multiple target models; determine the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models according to the multiple multidimensional scores and the reference multidimensional scores of the multiple audio works.

[0048] In a sixth aspect of an embodiment of the present disclosure, a computing device is provided, comprising: at least one processor and a memory; the memory stores computer-executable instructions; at least one processor executes the computer-executable instructions stored in the memory, so that at least one processor executes the model training method provided in the first aspect, and / or, so that at least one processor executes the multidimensional scoring method provided in the second aspect.

[0049] In an embodiment of the present disclosure, a target model for scoring audio works is trained based on training data, wherein the training data includes multidimensional feature attribute data of multiple audio works, and the multidimensional feature attribute data includes scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data. Thus, by utilizing the idea of ​​model training, a method for multidimensional analysis of audio works is provided, which not only realizes multidimensional scoring of audio works, but also makes the scoring results more professional through multidimensional scoring of audio works, and makes a comprehensive and scientific distinction between audio works. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0051] Figure 1 The following schematically shows an application scenario diagram provided according to an embodiment of the present disclosure;

[0052] Figure 2 The following schematically shows a flow chart of a model training method provided according to an embodiment of the present disclosure;

[0053] Figure 3 A schematic diagram of a flow chart of a model training method provided according to another embodiment of the present disclosure is schematically shown;

[0054] Figure 4 The following schematically shows a flow chart of a model training method provided according to another embodiment of the present disclosure;

[0055] Figure 5 A schematic diagram of a flow chart of a model training method provided according to another embodiment of the present disclosure is shown schematically;

[0056] Figure 6 The flowchart of the multidimensional scoring method provided according to an embodiment of the present disclosure is schematically shown;

[0057] Figure 7 The following schematically shows a flow chart of a multidimensional scoring method provided according to another embodiment of the present disclosure;

[0058] Figure 8 The following schematically shows a flow chart of a multidimensional scoring method provided according to another embodiment of the present disclosure;

[0059] Fig. 9 The structure diagram of a storage medium provided according to an embodiment of the present disclosure is schematically shown;

[0060] Fig.10 The structure diagram of the model training device provided according to an embodiment of the present disclosure is schematically shown;

[0061] Fig.11 The structure diagram of a multi-dimensional scoring device provided according to an embodiment of the present disclosure is schematically shown;

[0062] Fig.12 The schematic diagram shows a structure of a computing device provided according to an embodiment of the present disclosure.

[0063] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0064] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0065] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0066] According to the embodiments of the present disclosure, a model training method, a multidimensional scoring method, a medium and a device are proposed.

[0067] In this article, it is important to understand that the terms and their meanings are as follows:

[0068] (1) Accompaniment domain attribute data: An audio work can be decomposed into the accompaniment domain, the dry voice domain, and the score domain. The features and attributes related to the accompaniment are collectively referred to as the accompaniment domain attribute data.

[0069] (2) Dry voice domain attribute data: In audio works, taking user karaoke works as an example, the sound file generated by the user's singing is the dry voice, and the characteristics and attributes related to the dry voice are collectively referred to as dry voice domain attribute data.

[0070] (3) Rating domain attribute data: In some products, such as audio and video applications, any behavior that has evaluative significance for audio works, such as the number of likes, comments, shares, follows, and gifts, can be converted into features and attributes of ratings, collectively referred to as rating domain attribute data.

[0071] In addition, any number of elements in the drawings is for illustration and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0072] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0074] The inventors have found that as the functions of application programs continue to increase and improve, users can now use the application program to achieve a variety of functions, such as online karaoke entertainment activities, which has led to the need for users to rate the corresponding audio works. When rating a user's audio work, the rating is usually based on the dry sound domain attribute data in the audio work, and fewer rating factors are considered, resulting in poor professionalism in the rating results, making it difficult to rate high and low, and making the user's audio works less distinguishable. Therefore, it is necessary to consider how to rate audio works in multiple dimensions.

[0075] With the idea of ​​performing multi-dimensional scoring on audio works, in an embodiment of the present invention, a target model capable of performing multi-dimensional scoring on audio works is trained based on training data including multi-dimensional feature attribute data of multiple audio works, wherein the multi-dimensional feature attribute data includes scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data. Furthermore, based on the target model, not only is multi-dimensional scoring of audio works achieved, making the scoring results more professional, but also a comprehensive and scientific distinction is made between audio works through multi-dimensional scoring of audio works.

[0076] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0077] Application Scenario Overview

[0078] The scenarios to which the embodiments of the present disclosure are applicable include model training scenarios and / or multi-dimensional scoring scenarios.

[0079] First reference Figure 1 , Figure 1 The application scenario diagram provided according to an embodiment of the present disclosure is schematically shown, and the devices involved in the application scenario include a server 101.

[0080] When the application scenario is a model training scenario: the server 101 stores pre-collected training data and deploys a model to be trained, and the model can be trained based on the training data on the server 101.

[0081] When the application scenario is a multidimensional scoring scenario: the server 101 stores relevant data of the audio work to be multidimensionally scored and deploys a trained model. The multidimensional score of the audio work can be predicted on the server 101 based on the relevant data of the audio work and the trained model.

[0082] Optionally, the device involved in the application scenario also includes a terminal 102, and the server 101 and the terminal 102 can communicate through a network. In this case, the server 101 can be a product server of an application related to audio and video, and the terminal 102 can be a client deployed with the application.

[0083] The user can open an audio and video related application on the terminal 102 and interact with the application, for example, singing karaoke on the application. With the user's authorization, the terminal 102 can record the user's karaoke works and send the recorded karaoke works to the server 101 for model training and / or multi-dimensional scoring of the works.

[0084] Exemplary Methods

[0085] Combine the following Figure 1 For application scenarios, refer to Figure 2-Figure 8 To describe the model training method and multidimensional scoring method provided according to the exemplary embodiments of the present disclosure. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this regard. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0086] It should be noted that the embodiments of the present disclosure can be applied to electronic devices, and the electronic devices can be terminals or servers, that is, the model training method and / or multidimensional scoring method provided by the exemplary embodiments of the present disclosure can be executed on the terminal or server.

[0087] Among them, the terminal can be a personal digital assistant (PDA) device, a handheld device with wireless communication function (such as a smart phone, a tablet computer), a computing device (such as a personal computer (PC)), a vehicle-mounted device, a wearable device (such as a smart watch, a smart bracelet), a smart home device (such as a smart display device), etc.

[0088] The server can be a monolithic server or a distributed server across multiple computers or computer data centers. The server can also be of various types, such as but not limited to, a network server, an application server, or a database server, or a proxy server.

[0089] Optionally, the server may include hardware, software, or an embedded logic component or a combination of two or more such components for performing appropriate functions supported or implemented by the server. For example, the server may be a blade server, a cloud server, etc., or may be a server group consisting of multiple servers, including one or more of the above-mentioned categories of servers, etc.

[0090] It should be noted that the model training method and the multidimensional scoring method provided according to the exemplary embodiments of the present disclosure can be executed on the same device or on different devices.

[0091] refer to Figure 2 , Figure 2 The flowchart of the model training method provided according to an embodiment of the present disclosure is schematically shown. Figure 1 Taking the server in the example, the specific execution subject can be determined according to the actual application scenario, and the embodiment of the present disclosure does not impose any special restrictions on this. Figure 2 As shown, the model training method includes:

[0092] S201. Determine training data, where the training data includes multi-dimensional feature attribute data of a plurality of audio works, where the multi-dimensional feature attribute data includes score domain attribute data, dry voice domain attribute data, and accompaniment domain attribute data.

[0093] Here, the score domain attribute data may include one or more of the number of likes, comments, shares, followers, and gifts of the work. The dry voice domain attribute data may include one or more of intonation, rhythm, breath, emotional intensity, pronunciation accuracy, range, number of vibratos, and number of glissandos. The accompaniment domain attribute data may include one or more of accompaniment difficulty, accompaniment type, accompaniment age, accompaniment popularity, and accompaniment style.

[0094] The server may obtain the score domain attribute data of the multiple audio works from the business database. Similarly, the server may obtain the accompaniment domain attribute data of the multiple audio works from the accompaniment library; and may obtain the dry domain attribute data of the multiple audio works from the music library. Taking the acquisition of dry domain attribute data as an example, the server may obtain the annotation information corresponding to the multiple audio works from the music library, and then input the annotation information into a preset learning model, such as a neural network model, to learn and obtain the dry domain attribute data of the multiple audio works.

[0095] For the above training data, the above server can obtain the above multiple audio works from historical audio works as the above training data. Here, the above historical audio works can be audio works generated after multiple users interact with audio and video applications over a period of time. For example, taking online karaoke as an example, the user opens an application related to karaoke on a terminal (such as a mobile phone) and selects a karaoke recording mode, so that the above application records songs in this mode to generate the user's karaoke works.

[0096] In the historical audio works, the multi-dimensional characteristic attribute data of the historical audio works may be recorded, and the user identification and / or the identification of the audio and video application may also be recorded. Among them, the above period of time may be set according to actual conditions, for example, within one month from the current time.

[0097] S202. Perform model training based on the above training data to obtain a trained target model, which is used to perform multi-dimensional scoring on the audio works.

[0098] Among them, the multidimensional scoring of audio works can be reflected as the multidimensional feature scoring of audio works. For example, the multidimensional scoring of audio works includes the scoring domain scoring, dry voice domain scoring and accompaniment domain scoring of audio works.

[0099] Here, when the server performs model training according to the training data, it can first input the multidimensional feature attribute data of multiple audio works in the training data into the target model to obtain the multidimensional prediction scores corresponding to the multiple audio works, and then adjust the model parameters of the target model based on the multidimensional prediction scores, the multidimensional reference scores corresponding to the multiple audio works and the loss function to obtain the trained target model. Among them, the server can determine the function value of the loss function according to the difference between the multidimensional prediction scores corresponding to the multiple audio works and the multidimensional reference scores corresponding to the multiple audio works, and then adjust the model parameters of the target model according to the function value of the loss function so that the function value of the loss function obtained after adjustment is less than the preset threshold, and finally obtain the trained target model. Here, the preset threshold can be determined according to the actual situation. For example, when the difference between the multidimensional prediction score corresponding to the audio work and the multidimensional reference score corresponding to the audio work is small, such as less than 5 points, the function value of the determined loss function is used as the preset threshold. In addition, the multidimensional reference scores corresponding to the multiple audio works can be obtained by relevant personnel after scoring based on the multidimensional attribute data of the multiple audio works, that is, the scoring domain attribute data, the dry voice domain attribute data and the accompaniment domain attribute data.

[0100] In addition, when the above-mentioned server performs model training according to the above-mentioned training data, it can also adjust the above-mentioned training data, such as deleting or adding the multi-dimensional feature attribute data of the above-mentioned multiple audio works, that is, deleting or adding the score domain attribute data (such as the number of likes, comments, shares, attention and gifts of the works, etc.), dry voice domain attribute data (such as pitch, rhythm, breath, emotional intensity, pronunciation accuracy, range, number of vibratos and number of glissando, etc.) and accompaniment domain attribute data (such as accompaniment difficulty, accompaniment type, accompaniment age, accompaniment popularity and accompaniment style, etc.) of the above-mentioned multiple audio works, so as to obtain a trained target model, so that the difference between the multi-dimensional predicted scores of the multiple audio works obtained based on the above-mentioned target model and the multi-dimensional reference scores corresponding to the multiple audio works is small, such as less than 5 points, thereby improving the prediction accuracy of the target model.

[0101] In this embodiment, in the training data, the above-mentioned server uses the multidimensional feature attribute data of multiple audio works as the input data of the target model, and then performs supervised training on the target model based on the multidimensional reference scores corresponding to the multiple audio works and the multidimensional predicted scores corresponding to the multiple audio works output by the target model, so that the multidimensional scores predicted by the target model are close to the multidimensional reference scores corresponding to the audio works, thereby improving the accuracy of the multidimensional score prediction of the target model.

[0102] In the disclosed embodiment, training data is determined based on multi-dimensional feature attribute data of multiple audio works, and then a target model is trained based on the training data, so that multi-dimensional scoring of the audio works can be achieved based on the trained target model, making the scoring results of the audio works more professional and making a comprehensive and scientific distinction between the audio works.

[0103] In addition, the embodiments of the present disclosure also consider data cleaning, feature quantization and / or normalization before performing model training based on the above training data. Among them, the above data cleaning can clean out some bad data to improve the accuracy of subsequent processing. The above feature quantization can convert non-numerical attribute data into numerical attribute data to facilitate subsequent data processing. The above normalization can perform dimensionless processing on the data and convert data of different dimensions to the same dimension to facilitate comparison of data of different units or magnitudes. Figure 3 This is a flow chart of a model training method provided by another embodiment of the present disclosure. The execution subject of this embodiment is also as follows Figure 1 For example, Figure 3 As shown, this embodiment performs data cleaning, feature quantization and normalization processing during model training, and the method includes:

[0104] S301. Perform data cleaning on the multiple audio works according to the multidimensional feature attribute data, as well as the duration and generation method of the multiple audio works and the preset cleaning rules, wherein the multidimensional feature attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data.

[0105] Here, the above-mentioned multiple audio works may be generated after the user interacts with the audio and video application. The above-mentioned preset cleaning rules can be set according to actual conditions, for example, including cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold. Among them, the above-mentioned audio works without dry sounds may be audio works that do not make any sound during the interaction between the user and the audio and video application or the sound ratio is lower than the preset ratio threshold. The above-mentioned audio works that are not generated after the user interacts with the audio and video application may be audio works that are directly uploaded by the user without interacting with the audio and video application.

[0106] In this embodiment, the server performs data cleaning on multiple audio works to remove audio works without dry sound, audio works with low scores, audio works directly uploaded without interacting with audio and video applications, and audio works with short duration, thereby improving the quality of the audio works and further improving the accuracy of subsequent processing.

[0107] S302. Determine training data based on the audio work obtained after data cleaning. The training data includes multi-dimensional feature attribute data of the audio work obtained after data cleaning.

[0108] S303, performing feature quantification and normalization processing on the multi-dimensional feature attribute data of the audio work obtained after the above data cleaning.

[0109] Among them, for some multidimensional feature attribute data, such as the accompaniment style and accompaniment type of the accompaniment domain, which are non-numerical attribute data, they need to be converted into numerical attribute data. Therefore, the above-mentioned server performs feature quantification on the multidimensional feature attribute data of the audio work obtained after data cleaning, so as to convert the non-numerical attribute data therein into numerical attribute data, which is convenient for subsequent data processing. Among them, the above-mentioned server can also preset the conversion rules for converting non-numerical attribute data into numerical attribute data, and then, based on the conversion rules, the multidimensional feature attribute data of the audio work obtained after data cleaning is quantified. Exemplarily, the above-mentioned server can obtain the numerical attribute data corresponding to multiple non-numerical attribute data that have been converted, and then construct the conversion rules for converting the above-mentioned non-numerical attribute data into numerical attribute data based on the acquired data.

[0110] In addition, for data of different units and / or different dimensions, in order to facilitate subsequent data comparison, the above server normalizes the data, that is, performs dimensionless processing on the data and converts the data of different dimensions into the same dimension.

[0111] S304. Perform model training based on the multi-dimensional feature attribute data of the audio work obtained after feature quantization and normalization processing to obtain a trained target model, which is used to perform multi-dimensional scoring on the audio work.

[0112] In the disclosed embodiment, before model training, data cleaning is performed to remove some bad data, and the data is further feature quantified to convert non-numerical attribute data into numerical attribute data, and the data is normalized so that data of different units and / or different dimensions can be compared, thereby enabling the training data to be better processed by the target model, thereby improving the accuracy of the target model's multi-dimensional scoring of audio works.

[0113] In addition, when performing model training based on the above-mentioned training data, the embodiment of the present disclosure also considers performing model training based on the above-mentioned scoring domain attribute data to obtain a first scoring model, and then, performing model training based on the above-mentioned dry voice domain attribute data and the above-mentioned first scoring model to obtain a second scoring model, and performing model training based on the above-mentioned accompaniment domain attribute data and the above-mentioned first scoring model to obtain a third scoring model, and finally, based on the above-mentioned second scoring model and the third scoring model, determining the above-mentioned target model to achieve multi-dimensional scoring of the audio work. Figure 4 A flow chart of a model training method provided in another embodiment of the present disclosure. The execution subject of this embodiment is also as follows Figure 1 For example, Figure 4 As shown, the method includes:

[0114] S401. Determine training data, where the training data includes multi-dimensional feature attribute data of a plurality of audio works, where the multi-dimensional feature attribute data includes score domain attribute data, dry voice domain attribute data, and accompaniment domain attribute data.

[0115] The implementation of step S401 is shown in Figure 2 and Figure 3 The relevant description in the embodiments will not be repeated here.

[0116] S402: Based on the above-mentioned scoring domain attribute data, model training is performed to obtain a first scoring model.

[0117] Here, the server may obtain a linear relationship between the preset score variables of multiple audio works and the score domain attribute data, wherein the linear relationship includes the score parameters to be solved, and then, based on the linear relationship and the normal distribution of the score of multiple audio works, a fitting process is performed to determine the values ​​of the score parameters to be solved, thereby obtaining the first score model based on the linear relationship and the values ​​of the score parameters to be solved.

[0118] Among them, if the above scoring domain attribute data is represented by {x1,x2,…,x m}, (m≥2), the linear relationship obtained by the server can be expressed as:

[0119] y1=θ0+θ1x1+θ2x2+…+θ m x m +ε

[0120] Here, y1 represents the score variables of the above multiple audio works, θ0, θ1, …, θ m , and ε represents the scoring parameter to be solved.

[0121] In this embodiment, the normal distribution of the work scores of multiple audio works is:

[0122]

[0123] Among them, σ and ν are known parameters.

[0124] The server brings the above-mentioned scoring domain attribute data into the above two equations for fitting, and can solve the scoring parameters to be solved θ0, θ1, …, θ m , and ε, thereby obtaining a linear relationship between the work score variables of multiple audio works and the above-mentioned score domain attribute data, and the linear relationship can be used as the above-mentioned first scoring model.

[0125] S403: Perform model training based on the dry voice domain attribute data and the first scoring model to obtain a second scoring model, and perform model training based on the accompaniment domain attribute data and the first scoring model to obtain a third scoring model.

[0126] In this embodiment, after obtaining the first scoring model, the server may further determine the scores of the multiple audio works according to the first scoring model, and then perform principal component analysis on the scores of the multiple audio works and the dry sound domain attribute data to obtain multiple principal component parameters, thereby clustering the multiple principal component parameters to obtain clustering results, and obtain a second scoring model according to the clustering results. The second scoring model can divide the audio works into several categories, so that the score distribution of the works is more uniform, and avoid a large number of works from being clustered in a certain segment.

[0127] Here, the server performs principal component analysis on the scores of the multiple audio works and the dry sound domain attribute data, maps high-dimensional features into low-dimensional space, and retains all key information of the high-dimensional features as much as possible, while removing noise and multicollinearity in some high-dimensional features, thereby improving the subsequent processing speed while ensuring the normal progress of subsequent processing.

[0128] Before the principal component analysis is performed, the server may perform some data processing on the scores of the multiple audio works and the dry domain attribute data, such as data standardization. For example, assuming that there are n works and m feature parameters, the m feature parameters include the scores of the multiple audio works and the dry domain attribute data, thus, the matrix of the original data is:

[0129]

[0130] Among them, x ij is the j-th data of the i-th audio work.

[0131] The server performs some data processing on the data in the matrix, such as data standardization, to obtain:

[0132]

[0133] Among them, y ij represents the jth data of the i-th audio work after normalization, represents the mean of the jth data, s j Represents the standard deviation of the j-th data.

[0134] Further calculate the correlation coefficient matrix of the standardized data:

[0135]

[0136] Then, solve the eigenvalues ​​and corresponding eigenvectors of the above correlation coefficient matrix:

[0137] |λI-R|=0

[0138] Then calculate the cumulative contribution and select the principal component containing the main information, that is, complete the principal component analysis of the work scores of the above multiple audio works and the above dry sound domain attribute data, and obtain:

[0139]

[0140] Assume that after principal component analysis, the server obtains multiple principal component parameters: CP1, CP2, …, CP a , and obtain a principal component parameters. Then, the server can cluster the a principal component parameters, for example, analyze the data, connect with the business scenario, determine the number of categories, randomly determine the initial cluster center in each category, calculate the distance (Euclidean distance) between each sample and the cluster center, classify the sample into the category closest to the cluster center, reselect a new cluster center for each category, reclassify according to the new cluster center, repeat the above operation until the specified number of iterations is reached or the new cluster center does not change compared with the previous cluster center, that is, stop clustering.

[0141] Assume that the score of the initial cluster center of the Kth class is S K , for the work rating and dry sound domain attribute data of the audio work, through the above principal component analysis and clustering, they are clustered into m categories with a Euclidean distance of d, thus obtaining the second scoring model:

[0142] y2=S K +λd

[0143] Among them, λ is the weight coefficient of Euclidean distance.

[0144] In addition, for obtaining the third scoring model, the above-mentioned server can first obtain a linear relationship between the preset work scoring variables of the above-mentioned multiple audio works and the above-mentioned accompaniment domain attribute data, wherein the linear relationship includes the accompaniment parameters to be solved, and then, according to the linear relationship and the above-mentioned first scoring model, perform fitting processing to determine the values ​​of the above-mentioned accompaniment parameters to be solved, and finally, based on the above-mentioned linear relationship and the values ​​of the above-mentioned accompaniment parameters to be solved, obtain the above-mentioned third scoring model.

[0145] For example, if the above accompaniment domain attribute data is represented by {z1, z2, ..., z m}, (m≥2), the linear relationship obtained by the server can be expressed as:

[0146] y3=θ′0+θ′1z1+θ′2z2+…+θ′ m z m +ε′

[0147] Among them, y3 represents the work score variables of the above multiple audio works, θ′0,θ′1,…,θ′ mm , and ε′ represents the accompaniment parameters to be solved.

[0148] In this embodiment, in order to find the optimal solution for the multivariate linear regression equation to ensure that the predicted value of the model is closest to the actual value, the following equation is solved:

[0149]

[0150] The server uses the accompaniment domain attribute data and the first scoring model as the input and performs fitting processing to solve the scoring parameters θ′0, θ′1, …, θ′ m , and ε′, thereby obtaining a linear relationship between the work score variables of multiple audio works and the above-mentioned accompaniment domain attribute data, and this linear relationship can be used as the above-mentioned third scoring model.

[0151] S404: Based on the second scoring model and the third scoring model, determine a target model, where the target model is used to perform multi-dimensional scoring on the audio work.

[0152] Here, after obtaining the second scoring model and the third scoring model, the server may further obtain the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model, thereby obtaining the target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model. For example, the target model is expressed as:

[0153] y=αy2+βy3

[0154] Among them, y represents the above-mentioned target model, α represents the target weight coefficient of the second scoring model, and β represents the target weight coefficient of the third scoring model.

[0155] Among them, when obtaining the above-mentioned coefficients α and β, the above-mentioned server can obtain the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models, and then determine multiple target models according to the above-mentioned second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models, and then obtain multiple multidimensional scores of multiple audio works based on the above-mentioned training data and the multiple target models, and finally, according to the multiple multidimensional scores and the reference multidimensional scores of the above-mentioned multiple audio works, determine the target weight coefficients of the above-mentioned second scoring model and the target weight coefficients of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models.

[0156] In this embodiment, it is considered that if α is too large, the model score will lack accompaniment attributes, resulting in insufficient professionalism, and if β is too large, the model score will lack dry sound attributes, resulting in serious distortion of the score. Therefore, the above-mentioned server adjusts the correlation coefficient of the model through testing and tuning, considering the generalization effect of the model on the training data, so as to obtain the model with the best generalization ability as the target model, thereby improving the accuracy of subsequent multi-dimensional scoring of audio works based on the target model.

[0157] In the disclosed embodiment, based on the above-mentioned scoring domain attribute data, model training is performed to obtain a first scoring model, and then, based on the above-mentioned dry voice domain attribute data and the above-mentioned first scoring model, model training is performed to obtain a second scoring model, and based on the above-mentioned accompaniment domain attribute data and the above-mentioned first scoring model, model training is performed to obtain a third scoring model, thereby, based on the above-mentioned second scoring model and the third scoring model, the above-mentioned target model is determined, and the audio work is scored for multi-dimensional attributes, and the scoring is more professional; moreover, the disclosed embodiment effectively solves the problem of scoring discrimination through the above-mentioned classification scoring and fitting scoring, and makes the work scoring tend to be normally distributed, which is more in line with business scenarios.

[0158] In addition, if Figure 5 As shown, a flow chart of a model training method provided by another embodiment of the present disclosure is shown, and the execution subject is Figure 1The server in is taken as an example. The server first obtains multidimensional feature attribute data of multiple audio works, and the multidimensional feature attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data. Then, according to the preset cleaning rules, the multiple audio works are cleaned, and training data is determined based on the audio works obtained after data cleaning. The training data includes the multidimensional feature attribute data of the audio works obtained after data cleaning. Then, the multidimensional feature attribute data of the audio works obtained after data cleaning is subjected to feature quantization and normalization processing, so as to perform model training based on the multidimensional feature attribute data of the audio works obtained after feature quantization and normalization processing. For example, the server performs model training based on the scoring domain attribute data of the audio work obtained after feature quantization and normalization processing to obtain a first scoring model, and performs model training based on the dry voice domain attribute data of the audio work obtained after feature quantization and normalization processing and the first scoring model to obtain a second scoring model, and performs model training based on the accompaniment domain attribute data of the audio work obtained after feature quantization and normalization processing and the first scoring model to obtain a third scoring model, thereby, based on the second scoring model and the third scoring model, determining a target model, and the target model is used to perform multi-dimensional scoring on the audio work.

[0159] Among them, the embodiment of the present disclosure considers the multi-dimensional characteristic attributes of the audio work, provides a method for multi-dimensional analysis of the audio work, and realizes multi-dimensional scoring of the audio work. Moreover, before the model training, the embodiment of the present disclosure performs data cleaning, feature quantification and normalization processing, so that the training data can be better processed by the target model, and the accuracy of the multi-dimensional scoring of the audio work by the target model is improved. In addition, by multi-dimensional scoring of the audio work, the embodiment of the present disclosure makes the scoring result more professional and makes a comprehensive and scientific distinction between the audio works.

[0160] Subsequently, multiple embodiments of applying the target model to multi-dimensional scoring are provided.

[0161] refer to Figure 6 , Figure 6 The flowchart of the multi-dimensional scoring method provided according to an embodiment of the present disclosure is schematically shown. The execution subject of this embodiment is also as follows Figure 1 For example, Figure 6 As shown, the multidimensional scoring method includes:

[0162] S601. Determine multi-dimensional characteristic attribute data of a target audio work, wherein the multi-dimensional characteristic attribute data includes score domain attribute data, dry voice domain attribute data, and accompaniment domain attribute data.

[0163] Among them, the above-mentioned scoring domain attribute data, dry voice domain attribute data and accompaniment domain attribute data can refer to the description of the previous embodiment and will not be repeated here.

[0164] The server may obtain the target audio work from historical audio works. The historical audio work may be an audio work generated by multiple users interacting with an audio and video application over a period of time. The historical audio work may contain the multi-dimensional feature attribute data of the historical audio work, and may also contain a user ID and / or an ID of the audio and video application.

[0165] Optionally, the server may determine the multi-dimensional characteristic attribute data of the target audio work at preset time intervals, thereby periodically performing multi-dimensional scoring analysis on the target audio work to improve the accuracy of the diverse scoring.

[0166] Optionally, upon receiving a request for multi-dimensional scoring of the target audio work, the server may determine the multi-dimensional characteristic attribute data of the target audio work, thereby meeting the demand for multi-dimensional scoring of the target audio work and providing users with a better service experience.

[0167] S602. Obtain a multidimensional score for the target audio work based on the multidimensional feature attribute data and the target model.

[0168] Among them, the target model is trained according to the model training method provided in any of the aforementioned embodiments. For details, please refer to the aforementioned embodiments and will not be repeated here.

[0169] In this embodiment, the above-mentioned multidimensional feature attribute data can be input into the target model, and the target model can perform a multidimensional score on the target audio work based on the above-mentioned multidimensional feature attribute data to obtain a multidimensional score predicted by the target model. This multidimensional score is called the multidimensional score of the target audio work.

[0170] Among them, the multidimensional score of the target audio work can be reflected as the multidimensional feature score of the target audio work, such as the multidimensional score of the target audio work includes the score domain score, the dry voice domain score and the accompaniment domain score of the target audio work.

[0171] S603. Based on the target audio work, the multi-dimensional rating is pushed.

[0172] In this embodiment, the diversity prediction score can be displayed by pushing a reminder message. For example, the target user is determined based on the above-mentioned target audio work, and a reminder message such as "User XX, after evaluation and analysis, the multi-dimensional score of your audio work is..." is sent to the target user; and / or, the multi-dimensional score can be displayed to the target user by displaying the diversity prediction score of the target user on the display screen.

[0173] Exemplarily, the server may send the multi-dimensional rating of the target audio work to the terminal of the target user for display. In addition, if the current execution subject is a terminal, the terminal may display the multi-dimensional rating of the target audio work on its own display screen.

[0174] Optionally, the above-mentioned server can respond to the target user's access request to the user center page, and display the multi-dimensional rating of the above-mentioned target audio work to the target user on the user center page, thereby facilitating the user to view the multi-dimensional rating of his own audio work in the personal center, improving the convenience of viewing the multi-dimensional rating, and providing the user with a better experience.

[0175] Optionally, the server may display the multi-dimensional rating of the target audio work to the target user by combining one or more preset patterns, texts, and audios.

[0176] Optionally, the server may determine the multidimensional grade corresponding to the multidimensional grade of the target audio work of the target user based on the correspondence between the multidimensional grade and the multidimensional grade, and display the multidimensional grade to the target user. For example, the server may display to the target user that the target user has achieved a certain achievement related to the multidimensional grade of the audio work.

[0177] In the disclosed embodiment, a multi-dimensional score is performed on the target audio work based on the multi-dimensional feature attribute data of the target audio work and a pre-trained target model, and the multi-dimensional score is pushed based on the target audio work. Thus, the score result of the audio work is more professional, which is conducive to assigning high and low scores, improving the differentiation of audio works, and also helps relevant users to understand the situation of their own audio works more accurately.

[0178] In addition, before obtaining the multi-dimensional score of the target audio work based on the above-mentioned multi-dimensional feature attribute data and the target model, the embodiment of the present disclosure also considers data cleaning, feature quantification and normalization processing, so that the target audio work can be better processed by the target model, thereby improving the accuracy of the target model's multi-dimensional scoring of the audio work. Figure 7 This is a flow chart of a multidimensional scoring method provided by another embodiment of the present disclosure. The execution subject of this embodiment is also as follows Figure 1 For example, Figure 7 As shown, the method includes:

[0179] S701. Based on the multidimensional feature attribute data, as well as the duration and generation method of the target audio work, the target audio work is cleaned according to preset cleaning rules, wherein the multidimensional feature attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data.

[0180] Here, the target audio work may be generated after the user interacts with the audio and video application. The preset cleaning rules may be set according to actual conditions, for example, including cleaning audio works without dry sound, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

[0181] In this embodiment, the server performs data cleaning on the target audio works to remove audio works without dry sound, audio works with low scores, audio works directly uploaded without interacting with audio and video applications, and audio works with short duration, thereby improving the quality of the audio works and further improving the accuracy of subsequent processing.

[0182] S702. Determine multidimensional feature attribute data of a target audio work based on the audio work obtained after data cleaning. The multidimensional feature attribute data of the target audio work includes the multidimensional feature attribute data of the audio work obtained after data cleaning.

[0183] S703, performing feature quantification and normalization processing on the multi-dimensional feature attribute data of the audio work obtained after the above data cleaning.

[0184] The feature quantization is used to convert non-numeric attribute data into numeric attribute data, and the normalization is used to perform dimensionless processing on the data and convert data of different dimensions into the same dimension.

[0185] In this embodiment, considering that some multidimensional feature attribute data, such as the accompaniment style and accompaniment type of the accompaniment domain, are non-numerical attribute data and need to be converted into numerical attribute data, the multidimensional feature attribute data of the audio work obtained after data cleaning is subjected to feature quantification to convert the non-numerical attribute data therein into numerical attribute data, so as to facilitate subsequent data processing. Among them, the above-mentioned server can also preset a conversion rule for converting non-numerical attribute data into numerical attribute data, and then, based on the conversion rule, the multidimensional feature attribute data of the audio work obtained after data cleaning is subjected to feature quantification. Exemplarily, the above-mentioned server can obtain the numerical attribute data corresponding to a plurality of non-numerical attribute data that have been converted, and then, based on the obtained data, construct a conversion rule for converting the above-mentioned non-numerical attribute data into numerical attribute data.

[0186] In addition, this embodiment takes into account that subsequent data comparison is inconvenient for data with different units and / or different dimensions, and therefore further normalizes the data, i.e., performs dimensionless processing and converts data with different dimensions to the same dimension to facilitate subsequent data comparison.

[0187] S704. Obtain a multidimensional score for the target audio work based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization processing and the target model.

[0188] S705. Based on the target audio work, the multi-dimensional rating is pushed.

[0189] In the disclosed embodiment, before obtaining the multidimensional score of the target audio work based on the above-mentioned multidimensional feature attribute data and the target model, data cleaning is performed to remove some bad data, and the data is further feature quantified to convert non-numerical attribute data into numerical attribute data, and the data is normalized so that data of different units and / or different dimensions can be compared, so that the target audio work can be better processed by the target model, thereby improving the accuracy of the target model's multidimensional scoring of the audio work.

[0190] In addition, when obtaining the multidimensional score of the above-mentioned target audio work based on the above-mentioned multidimensional feature attribute data and the target model, the embodiment of the present disclosure also considers obtaining a first score based on the score domain attribute data and the first scoring model of the target audio work, and then obtains a second score based on the dry voice domain attribute data, the first score and the second scoring model of the target audio work, and obtains a third score based on the accompaniment domain attribute data, the first score and the third scoring model of the target audio work, and finally, based on the above-mentioned second score, third score and target model, realizes multidimensional scoring of the target audio work. Figure 8 A flowchart of a multidimensional scoring method provided by another embodiment of the present disclosure is shown in FIG. Figure 1 For example, Figure 8 As shown, the method includes:

[0191] S801. Determine multi-dimensional characteristic attribute data of a target audio work, wherein the multi-dimensional characteristic attribute data includes score domain attribute data, dry voice domain attribute data, and accompaniment domain attribute data.

[0192] The implementation of step S801 is shown in Figure 6 and Figure 7 The relevant description in the embodiments will not be repeated here.

[0193] S802. Obtain a first score based on the score domain attribute data of the target audio work and a first score model, wherein the first score model is trained based on the score domain attribute data of multiple audio works.

[0194] Here, the server may first obtain a linear relationship between the score variables of a preset plurality of audio works and the score domain attribute data, wherein the linear relationship includes the score parameters to be solved, and then, based on the linear relationship and the normal distribution of the score of the plurality of audio works, perform fitting processing to determine the values ​​of the score parameters to be solved, thereby obtaining a first score model based on the linear relationship and the values ​​of the score parameters to be solved, and inputting the score domain attribute data of the target audio work into the first score model to obtain a first score.

[0195] S803. Obtain a second score based on the dry domain attribute data of the target audio work, the first score and the second score model, and obtain a third score based on the accompaniment domain attribute data of the target audio work, the first score and the third score model, wherein the second score model is trained based on the dry domain attribute data of multiple audio works and the first score model, and the third score model is trained based on the accompaniment domain attribute data of multiple audio works and the first score model.

[0196] In this embodiment, the above-mentioned server can determine the work scores of multiple audio works based on the first scoring model, and then perform principal component analysis on the work scores and dry voice domain attribute data of the multiple audio works to obtain multiple principal component parameters, thereby clustering the multiple principal component parameters to obtain clustering results, and obtain a second scoring model based on the clustering results, and input the dry voice domain attribute data and the first score of the above-mentioned target audio work into the second scoring model to obtain a second score.

[0197] In addition, the above-mentioned server can also obtain a linear relationship between the work score variables and the accompaniment domain attribute data of a plurality of preset audio works, wherein the linear relationship includes the accompaniment parameters to be solved, and then, according to the linear relationship and the first scoring model, a fitting process is performed to determine the values ​​of the accompaniment parameters to be solved, thereby, based on the above-mentioned linear relationship and the values ​​of the accompaniment parameters to be solved, a third scoring model is obtained, and the accompaniment domain attribute data and the first score of the above-mentioned target audio work are input into the third scoring model to obtain a third score.

[0198] S804. Determine a multi-dimensional score of the target audio work based on the second score, the third score and the target model, wherein the target model is determined based on the second score model and the third score model.

[0199] Here, the above-mentioned server can obtain the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model, and obtain the target model based on the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model, and input the above-mentioned second score and third score into the target model to obtain a multidimensional score.

[0200] Among them, the above-mentioned server can obtain preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models, and then determine multiple target models according to the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models, thereby obtaining multiple multidimensional scores of multiple audio works based on the multidimensional feature attribute data of multiple audio works and multiple target models, and finally, according to the multiple multidimensional scores and the reference multidimensional scores of the multiple audio works, determine the target weight coefficients of the second scoring model and the target weight coefficient of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models.

[0201] S805. Based on the target audio work, the multi-dimensional rating is pushed.

[0202] In the disclosed embodiment, a first score is obtained based on the score domain attribute data of the target audio work and a first scoring model, and then a second score is obtained based on the dry voice domain attribute data of the target audio work, the first score and the second scoring model, and a third score is obtained based on the accompaniment domain attribute data of the target audio work, the first score and the third scoring model. Finally, based on the above-mentioned second score, third score and target model, the audio work is scored for multi-dimensional attributes, the scoring is more professional and effectively solves the problem of scoring discrimination.

[0203] Exemplary Media

[0204] After introducing the method of the exemplary embodiment of the present disclosure, next, refer to Fig. 9 A storage medium according to an exemplary embodiment of the present disclosure is described.

[0205] refer to Fig. 9 As shown, the storage medium 90 stores a program product for implementing the above method according to an embodiment of the present disclosure, which can be a portable compact disk read-only memory (CD-ROM) and includes program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0206] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0207] The readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, wherein the readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable signal medium may also be any readable medium other than a readable storage medium.

[0208] Program code for performing the disclosed operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0209] Exemplary Devices

[0210] After introducing the medium of the exemplary embodiment of the present disclosure, next, reference is made to Figure 10-11 The model training device and the multidimensional scoring device of the exemplary embodiment of the present disclosure are explained, wherein the model training device is used to implement the model training method provided by any of the above method embodiments, and the multidimensional scoring device is used to implement the multidimensional scoring method provided by any of the above method embodiments. The implementation principles and technical effects are similar and will not be repeated here.

[0211] refer to Fig.10 , Fig.10 This is a schematic diagram of the structure of a model training device provided by an embodiment of the present disclosure. Fig.10 As shown, the model training device includes:

[0212] The data determination module 1001 is used to determine training data, which includes multi-dimensional feature attribute data of multiple audio works, and the multi-dimensional feature attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data.

[0213] The model training module 1002 is used to perform model training based on the above training data to obtain a trained target model, and the target model is used to perform multi-dimensional scoring on the audio works.

[0214] In one embodiment of the present disclosure, the data determination module 1001 is specifically configured to:

[0215] According to the multidimensional feature attribute data, as well as the duration and generation method of the multiple audio works, data cleaning is performed on the multiple audio works according to preset cleaning rules; based on the audio works obtained after data cleaning, training data is determined, and the training data includes the multidimensional feature attribute data of the audio works obtained after the data cleaning.

[0216] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0217] The multi-dimensional feature attribute data of the audio works obtained after data cleaning are subjected to feature quantization and normalization. Feature quantization is used to convert non-numerical attribute data into numerical attribute data, and normalization is used to perform dimensionless processing on the data and convert data of different dimensions to the same dimension. Model training is performed based on the multi-dimensional feature attribute data of the audio works obtained after feature quantization and normalization.

[0218] In another embodiment of the present disclosure, the audio work is generated after the user interacts with the audio and video application. The preset cleaning rules include cleaning the audio work without dry sound, cleaning the audio work with a score lower than a preset score threshold, cleaning the audio work that is not generated after the user interacts with the audio and video application, and cleaning the audio work with a duration lower than a preset duration threshold.

[0219] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0220] Based on the scoring domain attribute data, model training is performed to obtain a first scoring model; based on the dry voice domain attribute data and the first scoring model, model training is performed to obtain a second scoring model, and based on the accompaniment domain attribute data and the first scoring model, model training is performed to obtain a third scoring model; based on the second scoring model and the third scoring model, a target model is determined.

[0221] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0222] A linear relationship between work score variables and score domain attribute data of a plurality of preset audio works is obtained, wherein the linear relationship includes a score parameter to be solved; a fitting process is performed based on the linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and the normal distribution of the work scores of the plurality of audio works to determine the value of the score parameter to be solved; a first score model is obtained based on the linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and the value of the score parameter to be solved.

[0223] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0224] According to the first scoring model, the work scores of the multiple audio works are determined, and principal component analysis is performed on the work scores and dry voice domain attribute data of the multiple audio works to obtain multiple principal component parameters; the multiple principal component parameters are clustered to obtain clustering results, and based on the clustering results, a second scoring model is obtained.

[0225] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0226] Obtain a linear relationship between the work scoring variables and the accompaniment domain attribute data of a plurality of preset audio works, wherein the linear relationship includes the accompaniment parameters to be solved; perform fitting processing based on the linear relationship between the work scoring variables and the accompaniment domain attribute data of the plurality of audio works and the first scoring model to determine the values ​​of the accompaniment parameters to be solved; obtain a third scoring model based on the linear relationship between the work scoring variables and the accompaniment domain attribute data of the plurality of audio works and the values ​​of the accompaniment parameters to be solved.

[0227] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0228] Obtain a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; obtain a target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model.

[0229] In yet another embodiment of the present disclosure, the model training module 1002 is specifically used to:

[0230] Obtain preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models; determine multiple target models according to the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models; obtain multiple multidimensional scores of multiple audio works based on the training data and the multiple target models; determine the target weight coefficients of the second scoring model and the target weight coefficients of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models according to the multiple multidimensional scores and the reference multidimensional scores of the multiple audio works.

[0231] refer to Fig.11 , Fig.11 This is a schematic diagram of the structure of a multi-dimensional scoring device provided by an embodiment of the present disclosure. Fig.11 As shown, the multidimensional scoring device includes:

[0232] The attribute determination module 1101 is used to determine multi-dimensional characteristic attribute data of the target audio work, wherein the multi-dimensional characteristic attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data.

[0233] The multidimensional scoring module 1102 is used to obtain a multidimensional score of the target audio work based on the multidimensional feature attribute data and the target model, wherein the target model is trained according to the model training method provided in any of the aforementioned embodiments.

[0234] The score push module 1103 is used to push multi-dimensional scores based on the target audio work.

[0235] In one embodiment of the present disclosure, the attribute determination module 1101 is specifically configured to:

[0236] According to the multidimensional characteristic attribute data, as well as the duration and generation method of the target audio work, data cleaning is performed on the target audio work in accordance with preset cleaning rules; based on the audio work obtained after data cleaning, the multidimensional characteristic attribute data of the target audio work is determined, and the multidimensional characteristic attribute data of the target audio work includes the multidimensional characteristic attribute data of the audio work obtained after data cleaning.

[0237] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0238] The multidimensional feature attribute data of the audio work obtained after data cleaning are subjected to feature quantization and normalization processing. Feature quantization is used to convert non-numerical attribute data into numerical attribute data, and normalization is used to perform dimensionless processing on the data and convert data of different dimensions into the same dimension. Based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization processing and the target model, a multidimensional score of the target audio work is obtained.

[0239] In another embodiment of the present disclosure, the target audio work is generated after the user interacts with the audio and video application. The preset cleaning rules include cleaning the audio work without dry sound, cleaning the audio work with a score lower than a preset score threshold, cleaning the audio work that is not generated after the user interacts with the audio and video application, and cleaning the audio work with a duration lower than a preset duration threshold.

[0240] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0241] A first score is obtained based on the score domain attribute data of the target audio work and a first scoring model, wherein the first scoring model is trained based on the score domain attribute data of multiple audio works; a second score is obtained based on the dry voice domain attribute data of the target audio work, the first score and the second scoring model, and a third score is obtained based on the accompaniment domain attribute data of the target audio work, the first score and the third scoring model, wherein the second scoring model is trained based on the dry voice domain attribute data of multiple audio works and the first scoring model, and the third scoring model is trained based on the accompaniment domain attribute data of multiple audio works and the first scoring model; a multi-dimensional score of the target audio work is determined based on the second score, the third score and the target model, wherein the target model is determined based on the second scoring model and the third scoring model.

[0242] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0243] A linear relationship between work score variables and score domain attribute data of a plurality of preset audio works is obtained, wherein the linear relationship includes score parameters to be solved; a fitting process is performed based on the linear relationship between the work score variables and score domain attribute data of the plurality of audio works and the normal distribution of the work scores of the plurality of audio works to determine the values ​​of the score parameters to be solved; a first score model is obtained based on the linear relationship between the work score variables and score domain attribute data of the plurality of audio works and the values ​​of the score parameters to be solved, and the score domain attribute data of the target audio work is input into the first score model to obtain a first score.

[0244] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0245] According to the first scoring model, the work scores of multiple audio works are determined, and principal component analysis is performed on the work scores and dry voice domain attribute data of the multiple audio works to obtain multiple principal component parameters; the multiple principal component parameters are clustered to obtain clustering results, and according to the clustering results, a second scoring model is obtained, and the dry voice domain attribute data of the target audio work and the first score are input into the second scoring model to obtain a second score.

[0246] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0247] Obtain a linear relationship between the work scoring variables and the accompaniment domain attribute data of a plurality of preset audio works, wherein the linear relationship includes the accompaniment parameters to be solved; perform fitting processing based on the linear relationship between the work scoring variables and the accompaniment domain attribute data of the plurality of audio works and a first scoring model to determine the values ​​of the accompaniment parameters to be solved; obtain a third scoring model based on the linear relationship between the work scoring variables and the accompaniment domain attribute data of the plurality of audio works and the values ​​of the accompaniment parameters to be solved, input the accompaniment domain attribute data of the target audio work and the first score into the third scoring model to obtain a third score.

[0248] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0249] Obtain the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model; obtain the target model according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model, input the second score and the third score into the target model, and obtain a multidimensional score.

[0250] In yet another embodiment of the present disclosure, the multidimensional scoring module 1102 is specifically configured to:

[0251] Obtain preset weight coefficients of multiple second scoring models and preset weight coefficients of multiple third scoring models; determine multiple target models based on the second scoring model, the third scoring model, the preset weight coefficients of multiple second scoring models, and the preset weight coefficients of multiple third scoring models; obtain multiple multidimensional scores of multiple audio works based on the multidimensional feature attribute data of multiple audio works and the multiple target models; determine the target weight coefficients of the second scoring model and the target weight coefficient of the third scoring model from the preset weight coefficients of multiple second scoring models and the preset weight coefficients of multiple third scoring models based on the multiple multidimensional score and the reference multidimensional score of the multiple audio works.

[0252] Exemplary Computing Devices

[0253] After introducing the method, medium and apparatus of the exemplary embodiments of the present disclosure, next, reference is made to Fig.12 A computing device according to an exemplary embodiment of the present disclosure is described.

[0254] Fig.12 The computing device 120 shown is merely an example and should not bring any limitation to the functionality and scope of use of the embodiments of the present disclosure.

[0255] like Fig.12As shown, the computing device 120 is in the form of a general computing device. The components of the computing device 120 may include but are not limited to: at least one processing unit 1201, at least one storage unit 1202, and a bus 1203 connecting different system components (including the processing unit 1201 and the storage unit 1202).

[0256] The bus 1203 includes a data bus, a control bus, and an address bus.

[0257] The storage unit 1202 may include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 12021 and / or a cache memory 12022 , and may further include a readable medium in the form of a non-volatile memory, such as a read-only memory (ROM) 12023 .

[0258] The storage unit 1202 may also include a program / utility 12025 having a set (at least one) of program modules 12024, such program modules 12024 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0259] The computing device 120 may also communicate with one or more external devices 1204 (e.g., keyboards, pointing devices, etc.). Such communication may be performed via an input / output (I / O) interface 1205. In addition, the computing device 120 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 1206. Fig.12 As shown, the network adapter 1206 communicates with other modules of the computing device 120 via the bus 1203. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the computing device 120, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0260] It should be noted that although several units / modules or sub-units / modules of the model training device and the multidimensional scoring device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to an embodiment of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0261] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0262] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A model training method, comprising: Determine training data, the training data comprising multi-dimensional feature attribute data of a plurality of audio works, the multi-dimensional feature attribute data comprising score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; Based on the scoring domain attribute data, model training is performed to obtain a first scoring model; Determine, according to the first scoring model, the scores of the plurality of audio works, perform principal component analysis on the scores of the plurality of audio works and the dry voice domain attribute data, and obtain a plurality of principal component parameters; Clustering the plurality of principal component parameters to obtain a clustering result, and obtaining a second scoring model according to the clustering result; Obtaining a linear relationship between the preset score variables of the plurality of audio works and the accompaniment domain attribute data, wherein the linear relationship includes the accompaniment parameter to be solved; Perform fitting processing according to the linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, and the first scoring model, to determine the value of the accompaniment parameter to be solved; Obtaining a third scoring model based on a linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, and the value of the accompaniment parameter to be solved; Based on the second scoring model and the third scoring model, a target model is determined; the target model is used to perform multi-dimensional scoring on the audio work.

2. The model training method according to claim 1, before performing model training according to the training data to obtain a trained target model, further comprises: According to the multi-dimensional feature attribute data, as well as the duration and generation method of the multiple audio works, and in accordance with preset cleaning rules, data cleaning is performed on the multiple audio works; The determining of training data comprises: The training data is determined based on the audio work obtained after data cleaning, and the training data includes the multi-dimensional feature attribute data of the audio work obtained after data cleaning.

3. The model training method according to claim 2, wherein the model training is performed according to the training data, comprising: The multidimensional characteristic attribute data of the audio work obtained after the data cleaning is subjected to feature quantization and normalization processing, wherein the feature quantization is used to convert non-numerical attribute data into numerical attribute data, and the normalization is used to perform dimensionless processing on the data and convert data of different dimensions into the same dimension; Model training is performed based on the multi-dimensional feature attribute data of the audio work obtained after feature quantization and normalization processing.

4. The model training method according to claim 2, wherein the audio work is generated by a user interacting with an audio and video application; The preset cleaning rules include cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

5. The model training method according to claim 1, wherein the step of performing model training based on the scoring domain attribute data to obtain a first scoring model comprises: Obtaining a linear relationship between the preset work score variables of the plurality of audio works and the score domain attribute data, wherein the linear relationship includes a score parameter to be solved; Perform fitting processing according to a linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and a normal distribution of the work scores of the plurality of audio works, to determine the value of the score parameter to be solved; The first scoring model is obtained based on the linear relationship between the work scoring variables of the multiple audio works and the scoring domain attribute data, and the value of the scoring parameter to be solved.

6. The model training method according to claim 1, wherein determining the target model based on the second scoring model and the third scoring model comprises: Obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; The target model is obtained according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model, and the target weight coefficient of the third scoring model.

7. The model training method according to claim 6, wherein obtaining the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model comprises: Obtaining a plurality of preset weight coefficients of the second scoring model and a plurality of preset weight coefficients of the third scoring model; Determining a plurality of target models according to the second scoring model, the third scoring model, a plurality of preset weight coefficients of the second scoring models, and a plurality of preset weight coefficients of the third scoring models; Based on the training data and the multiple target models, obtaining multiple multi-dimensional scores for the multiple audio works; Based on the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works, the target weight coefficient of the second rating model and the target weight coefficient of the third rating model are determined from the multiple preset weight coefficients of the second rating models and the multiple preset weight coefficients of the third rating models.

8. A multidimensional scoring method, comprising: Determining multidimensional characteristic attribute data of the target audio work, wherein the multidimensional characteristic attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; A first score is obtained based on the score domain attribute data of the target audio work and a first score model, wherein the first score model is trained based on the score domain attribute data of multiple audio works; a second score is obtained based on the dry domain attribute data of the target audio work, the first score, and the second score model, and a third score is obtained based on the accompaniment domain attribute data of the target audio work, the first score, and a third score is obtained based on the accompaniment domain attribute data of the target audio work, the first score, and a third score model, wherein the second score model is trained based on the dry domain attribute data of the multiple audio works and the first score model, and the third score model is trained based on the accompaniment domain attribute data of the multiple audio works and the first score model; a multidimensional score of the target audio work is determined based on the second score, the third score, and a target model, wherein the target model is trained according to the model training method according to any one of claims 1 to 7; Based on the target audio work, the multi-dimensional rating is pushed.

9. The multidimensional scoring method according to claim 8, before obtaining the multidimensional score of the target audio work according to the multidimensional feature attribute data and the target model, further comprising: According to the multi-dimensional feature attribute data, as well as the duration and generation method of the target audio work, and in accordance with preset cleaning rules, data cleaning is performed on the target audio work; The step of determining the multi-dimensional characteristic attribute data of the target audio work includes: Based on the audio work obtained after data cleaning, the multidimensional feature attribute data of the target audio work is determined, and the multidimensional feature attribute data of the target audio work includes the multidimensional feature attribute data of the audio work obtained after data cleaning.

10. The multidimensional scoring method according to claim 9, wherein obtaining the multidimensional score of the target audio work according to the multidimensional feature attribute data and the target model comprises: The multidimensional characteristic attribute data of the audio work obtained after the data cleaning is subjected to feature quantization and normalization processing, wherein the feature quantization is used to convert non-numerical attribute data into numerical attribute data, and the normalization is used to perform dimensionless processing on the data and convert data of different dimensions into the same dimension; Based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization processing and the target model, a multidimensional score of the target audio work is obtained.

11. The multi-dimensional scoring method according to claim 9, wherein the target audio work is generated by a user interacting with an audio and video application; The preset cleaning rules include cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

12. The multidimensional scoring method according to claim 8, wherein obtaining a first score based on the scoring domain attribute data of the target audio work and a first scoring model comprises: Obtaining a preset linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, wherein the linear relationship includes a score parameter to be solved; Perform fitting processing according to the linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and the normal distribution of the work scores of the plurality of audio works, to determine the value of the score parameter to be solved; Based on the linear relationship between the score variables of the multiple audio works and the score domain attribute data, and the value of the score parameter to be solved, the first score model is obtained, and the score domain attribute data of the target audio work is input into the first score model to obtain the first score.

13. The multidimensional scoring method according to claim 8, wherein obtaining a second score based on the dry domain attribute data of the target audio work, the first score and the second scoring model comprises: Determine the scores of the plurality of audio works according to the first scoring model, perform principal component analysis on the scores of the plurality of audio works and the dry voice domain attribute data, and obtain a plurality of principal component parameters; Cluster the multiple principal component parameters to obtain a clustering result, obtain the second scoring model based on the clustering result, input the dry domain attribute data of the target audio work and the first score into the second scoring model to obtain the second score.

14. The multidimensional scoring method according to claim 8, wherein obtaining a third score based on the accompaniment domain attribute data of the target audio work, the first score and the third scoring model comprises: Obtaining a preset linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, wherein the linear relationship includes the accompaniment parameter to be solved; Perform fitting processing according to the linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, and the first scoring model, to determine the value of the accompaniment parameter to be solved; Based on the linear relationship between the work score variables and the accompaniment domain attribute data of the multiple audio works, and the value of the accompaniment parameter to be solved, the third scoring model is obtained, and the accompaniment domain attribute data of the target audio work and the first score are input into the third scoring model to obtain the third score.

15. The multidimensional scoring method according to claim 8, wherein determining the multidimensional score of the target audio work based on the second score, the third score and the target model comprises: Obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; The target model is obtained according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model, and the second score and the third score are input into the target model to obtain the multidimensional score.

16. The multidimensional scoring method according to claim 15, wherein obtaining the target weight coefficient of the second scoring model and the target weight coefficient of the third scoring model comprises: Obtaining a plurality of preset weight coefficients of the second scoring model and a plurality of preset weight coefficients of the third scoring model; Determining a plurality of target models according to the second scoring model, the third scoring model, a plurality of preset weight coefficients of the second scoring models, and a plurality of preset weight coefficients of the third scoring models; Based on the multidimensional feature attribute data of the plurality of audio works and the plurality of target models, obtaining a plurality of multidimensional scores of the plurality of audio works; Based on the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works, the target weight coefficient of the second rating model and the target weight coefficient of the third rating model are determined from the multiple preset weight coefficients of the second rating models and the multiple preset weight coefficients of the third rating models.

17. A computer-readable storage medium, wherein computer-executable instructions are stored in the computer-readable storage medium. When a processor executes the computer-executable instructions, the model training method as described in any one of claims 1 to 7 is implemented, and / or the multidimensional scoring method as described in any one of claims 8 to 16 is implemented.

18. A model training device, comprising: A data determination module, used to determine training data, wherein the training data includes multi-dimensional feature attribute data of a plurality of audio works, wherein the multi-dimensional feature attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; A model training module, used for performing model training based on the scoring domain attribute data to obtain a first scoring model; Determine, according to the first scoring model, the scores of the plurality of audio works, perform principal component analysis on the scores of the plurality of audio works and the dry voice domain attribute data, and obtain a plurality of principal component parameters; Cluster the multiple principal component parameters to obtain clustering results, and obtain a second scoring model based on the clustering results; obtain a linear relationship between the preset score variables of the multiple audio works and the accompaniment domain attribute data, wherein the linear relationship includes the accompaniment parameters to be solved; perform fitting processing based on the linear relationship between the score variables of the multiple audio works and the accompaniment domain attribute data, and the first scoring model, to determine the values ​​of the accompaniment parameters to be solved; obtain a third scoring model based on the linear relationship between the score variables of the multiple audio works and the accompaniment domain attribute data, and the values ​​of the accompaniment parameters to be solved; determine a target model based on the second scoring model and the third scoring model, and the target model is used to perform multi-dimensional scoring on the audio works.

19. The model training device according to claim 18, wherein the data determination module is specifically used to: According to the multi-dimensional feature attribute data, as well as the duration and generation method of the multiple audio works, and in accordance with preset cleaning rules, data cleaning is performed on the multiple audio works; The training data is determined based on the audio work obtained after data cleaning, and the training data includes the multi-dimensional feature attribute data of the audio work obtained after data cleaning.

20. The model training device according to claim 19, wherein the model training module is specifically used for: The multidimensional characteristic attribute data of the audio work obtained after the data cleaning is subjected to feature quantization and normalization processing, wherein the feature quantization is used to convert non-numerical attribute data into numerical attribute data, and the normalization is used to perform dimensionless processing on the data and convert data of different dimensions into the same dimension; Model training is performed based on the multi-dimensional feature attribute data of the audio work obtained after feature quantization and normalization processing.

21. The model training device according to claim 19, wherein the audio work is generated by a user interacting with an audio and video application; The preset cleaning rules include cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

22. The model training device according to claim 18, wherein the model training module is specifically used for: Obtaining a linear relationship between the preset work score variables of the plurality of audio works and the score domain attribute data, wherein the linear relationship includes a score parameter to be solved; Perform fitting processing according to a linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and a normal distribution of the work scores of the plurality of audio works, to determine the value of the score parameter to be solved; The first scoring model is obtained based on the linear relationship between the work scoring variables of the multiple audio works and the scoring domain attribute data, and the value of the scoring parameter to be solved.

23. The model training device according to claim 18, wherein the model training module is specifically used for: Obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; The target model is obtained according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model, and the target weight coefficient of the third scoring model.

24. The model training device according to claim 23, wherein the model training module is specifically used for: Obtaining a plurality of preset weight coefficients of the second scoring model and a plurality of preset weight coefficients of the third scoring model; Determining a plurality of target models according to the second scoring model, the third scoring model, a plurality of preset weight coefficients of the second scoring models, and a plurality of preset weight coefficients of the third scoring models; Based on the training data and the multiple target models, obtaining multiple multi-dimensional scores for the multiple audio works; Based on the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works, the target weight coefficient of the second rating model and the target weight coefficient of the third rating model are determined from the multiple preset weight coefficients of the second rating models and the multiple preset weight coefficients of the third rating models.

25. A multidimensional scoring device, comprising: An attribute determination module, used to determine multi-dimensional characteristic attribute data of a target audio work, wherein the multi-dimensional characteristic attribute data includes score domain attribute data, dry voice domain attribute data and accompaniment domain attribute data; A multidimensional scoring module, for obtaining a first score based on the score domain attribute data of the target audio work and a first scoring model, wherein the first scoring model is trained based on the score domain attribute data of multiple audio works; obtaining a second score based on the dry domain attribute data of the target audio work, the first score and the second scoring model, and obtaining a third score based on the accompaniment domain attribute data of the target audio work, the first score and a third scoring model, wherein the second scoring model is trained based on the dry domain attribute data of the multiple audio works and the first scoring model, and the third scoring model is trained based on the accompaniment domain attribute data of the multiple audio works and the first scoring model; determining a multidimensional score of the target audio work based on the second score, the third score and a target model, wherein the target model is trained according to the model training method according to any one of claims 1 to 7; The score push module is used to push the multi-dimensional score based on the target audio work.

26. The multi-dimensional scoring device according to claim 25, wherein the attribute determination module is specifically used to: According to the multi-dimensional feature attribute data, as well as the duration and generation method of the target audio work, and in accordance with preset cleaning rules, data cleaning is performed on the target audio work; Based on the audio work obtained after data cleaning, the multidimensional feature attribute data of the target audio work is determined, and the multidimensional feature attribute data of the target audio work includes the multidimensional feature attribute data of the audio work obtained after data cleaning.

27. The multidimensional scoring device according to claim 26, wherein the multidimensional scoring module is specifically used for: The multidimensional characteristic attribute data of the audio work obtained after the data cleaning is subjected to feature quantization and normalization processing, wherein the feature quantization is used to convert non-numerical attribute data into numerical attribute data, and the normalization is used to perform dimensionless processing on the data and convert data of different dimensions into the same dimension; Based on the multidimensional feature attribute data of the audio work obtained after feature quantization and normalization processing and the target model, a multidimensional score of the target audio work is obtained.

28. The multi-dimensional scoring device according to claim 27, wherein the target audio work is generated by a user interacting with an audio and video application; The preset cleaning rules include cleaning audio works without dry sounds, cleaning audio works with scores lower than a preset score threshold, cleaning audio works that are not generated after the user interacts with the audio and video application, and cleaning audio works with a duration lower than a preset duration threshold.

29. The multidimensional scoring device according to claim 25, wherein the multidimensional scoring module is specifically used for: Obtaining a preset linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, wherein the linear relationship includes a score parameter to be solved; Perform fitting processing according to the linear relationship between the work score variables of the plurality of audio works and the score domain attribute data, and the normal distribution of the work scores of the plurality of audio works, to determine the value of the score parameter to be solved; Based on the linear relationship between the score variables of the multiple audio works and the score domain attribute data, and the value of the score parameter to be solved, the first score model is obtained, and the score domain attribute data of the target audio work is input into the first score model to obtain the first score.

30. The multidimensional scoring device according to claim 25, wherein the multidimensional scoring module is specifically used for: Determine the scores of the plurality of audio works according to the first scoring model, perform principal component analysis on the scores of the plurality of audio works and the dry voice domain attribute data, and obtain a plurality of principal component parameters; Cluster the multiple principal component parameters to obtain a clustering result, obtain the second scoring model based on the clustering result, input the dry domain attribute data of the target audio work and the first score into the second scoring model to obtain the second score.

31. The multidimensional scoring device according to claim 25, wherein the multidimensional scoring module is specifically used for: Obtaining a preset linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, wherein the linear relationship includes the accompaniment parameter to be solved; Perform fitting processing according to the linear relationship between the score variables of the plurality of audio works and the accompaniment domain attribute data, and the first scoring model, to determine the value of the accompaniment parameter to be solved; Based on the linear relationship between the work score variables and the accompaniment domain attribute data of the multiple audio works, and the value of the accompaniment parameter to be solved, the third scoring model is obtained, and the accompaniment domain attribute data of the target audio work and the first score are input into the third scoring model to obtain the third score.

32. The multidimensional scoring device according to claim 25, wherein the multidimensional scoring module is specifically used for: Obtaining a target weight coefficient of the second scoring model and a target weight coefficient of the third scoring model; The target model is obtained according to the second scoring model, the target weight coefficient of the second scoring model, the third scoring model and the target weight coefficient of the third scoring model, and the second score and the third score are input into the target model to obtain the multidimensional score.

33. The multidimensional scoring device according to claim 32, wherein the multidimensional scoring module is specifically used for: Obtaining a plurality of preset weight coefficients of the second scoring model and a plurality of preset weight coefficients of the third scoring model; Determining a plurality of target models according to the second scoring model, the third scoring model, a plurality of preset weight coefficients of the second scoring models, and a plurality of preset weight coefficients of the third scoring models; Based on the multidimensional feature attribute data of the plurality of audio works and the plurality of target models, obtaining a plurality of multidimensional scores of the plurality of audio works; Based on the multiple multidimensional ratings and the reference multidimensional ratings of the multiple audio works, the target weight coefficient of the second rating model and the target weight coefficient of the third rating model are determined from the multiple preset weight coefficients of the second rating models and the multiple preset weight coefficients of the third rating models.

34. A computing device comprising: at least one processor and memory; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the model training method as described in any one of claims 1 to 7, and / or, so that the at least one processor executes the multidimensional scoring method as described in any one of claims 8 to 16.

Images