Song scoring model training method and song scoring method

Abstract

The invention relates to a song scoring model training method and a song scoring method. The method comprises the following steps: acquiring song audio and comments of a first sample song and comments of a second sample song; inputting the song audio into an audio feature extraction model to obtain song audio features, and inputting comments of the first sample song and comments of the second sample song into a comment feature extraction model to obtain comment features of the first sample song and comment features of the second sample song; and training a song scoring model according to the similarity between the song audio feature and the comment feature of the first sample song and the similarity between the song audio feature and the comment feature of the second sample song. By adopting the method, song scoring can be performed by analyzing the association degree between the comments and the song audios from the perspective of the comments, so that song scoring does not stay in pure audio analysis, more flexible and variable scoring is performed on the songs by fully utilizing the comment information, and the accuracy of song scoring is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a song scoring model training method and a song scoring method, device, computer equipment, storage medium and computer program product. Background Art

[0002] Singing scores can be obtained by analyzing the singing audio of the evaluation object in terms of timbre, pitch, rhythm, breath, vibrato, glissando, emotion and other dimensions, and a comprehensive score can be obtained. This comprehensive score can help the evaluation object understand the strengths and weaknesses of his or her singing and increase the fun of karaoke. For example, when different users sing the same song, the singing score of each user can be determined by analyzing their singing performance, so as to select the winner.

[0003] Traditional technology trains relevant models through song scores, and then determines the song scores with the help of the trained models. However, traditional technology only realizes song scoring by analyzing and learning the relationship between songs and corresponding specific scores. When the scores cannot completely cover each segment, traditional technology cannot learn the differences between different segments, and song scores obtained based on different scoring standards will also lead to inaccurate song scores in the final results, which is not conducive to improving the accuracy of song scoring. Summary of the invention

[0004] Based on this, it is necessary to provide a song scoring method, device, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy of song scoring in response to the above technical problems.

[0005] In a first aspect, the present application provides a song scoring model training method, wherein the song scoring model includes an audio feature extraction model and a comment feature extraction model, including:

[0006] Acquire song audio information of a first sample song and comment information of the first sample song, and acquire comment information of a second sample song different from the first sample song;

[0007] Inputting the song audio information into the audio feature extraction model to be trained to obtain the song audio features of the first sample song, and inputting the comment information of the first sample song and the comment information of the second sample song into the comment feature extraction model to be trained to obtain the comment features of the first sample song based on the comment information of the first sample song and obtain the comment features of the second sample song based on the comment information of the second sample song;

[0008] According to the similarity between the song audio features of the first sample song and the comment features of the first sample song, and the similarity between the song audio features of the first sample song and the comment features of the second sample song, the song scoring model to be trained is trained to obtain a trained song scoring model.

[0009] In one embodiment, the comment information of the first sample song includes comment information of the first type and comment information of the second type; the comment information of the second sample song includes comment information of the first type and comment information of the second type; the first type represents comment information obtained by user input through the music platform, and the second type represents comment information obtained based on expert experience;

[0010] The training of the song rating model to be trained according to the similarity between the song audio feature of the first sample song and the comment feature of the first sample song, and the similarity between the song audio feature of the first sample song and the comment feature of the second sample song, comprises:

[0011] Inputting the song audio information of the first sample song, the first type of comment information of the first sample song, and the first type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the first type of comment features of the first sample song based on the first type of comment information of the first sample song, and obtain the first type of comment features of the second sample song based on the first type of comment information of the second sample song;

[0012] Inputting the song audio information of the first sample song, the second type of comment information of the first sample song, and the second type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the second type of comment features of the first sample song based on the second type of comment information of the first sample song, and obtain the second type of comment features of the second sample song based on the second type of comment information of the second sample song;

[0013] The song rating model to be trained is trained based on at least the song audio features of the first sample song, the first type of comment features of the first sample song, the second type of comment features of the first sample song, the first type of comment features of the second sample song, and the second type of comment features of the second sample song.

[0014] In one embodiment, the training of the song rating model to be trained based on at least the song audio feature of the first sample song, the first type comment feature of the first sample song, the second type comment feature of the first sample song, the first type comment feature of the second sample song, and the second type comment feature of the second sample song includes:

[0015] The song rating model to be trained is trained based on a first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, a second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, a third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and a fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

[0016] In one embodiment, the training of the song rating model to be trained based on at least the song audio feature of the first sample song, the first type comment feature of the first sample song, the second type comment feature of the first sample song, the first type comment feature of the second sample song, and the second type comment feature of the second sample song includes:

[0017] According to a first similarity between the song audio feature of the first sample song and the first type of comment feature of the first sample song, and a second similarity between the song audio feature of the first sample song and the first type of comment feature of the second sample song, the song scoring model to be trained is trained to obtain a preliminarily trained song scoring model;

[0018] The preliminarily trained song rating model is trained based on a third degree of similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and a fourth degree of similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

[0019] In a second aspect, the present application also provides a song scoring method, comprising:

[0020] Inputting the song to be rated into a pre-trained song rating model to extract the song audio features of the song to be rated; wherein the pre-trained song rating model is trained by the song rating model training method;

[0021] In a preset review database, at least one review information whose review feature and the song audio feature of the song to be rated meet a preset similarity condition is determined as target review information; wherein the review feature of the review information in the review database is obtained by the pre-trained song rating model;

[0022] The song score of the song to be rated is determined according to the rating information corresponding to each target review information.

[0023] In one embodiment, the step of inputting the song to be rated into a pre-trained song rating model to extract the song audio features of the song to be rated includes:

[0024] Mapping the song audio information of the song to be rated to a preset space through a feature mapping network in an audio feature extraction model in the song rating model to obtain mapping features corresponding to the song audio information;

[0025] The mapping features corresponding to the song audio information are converted into song audio features of the song to be rated through the feature conversion network in the audio feature extraction model.

[0026] In one embodiment, determining the song score of the song to be scored according to the score information corresponding to each target comment information includes:

[0027] According to the weight information corresponding to each of the target review information, the score information corresponding to each of the target review information is weighted and summed to obtain the song score of the song to be rated.

[0028] In a third aspect, the present application further provides a song scoring model training device, wherein the song scoring model includes an audio feature extraction model and a comment feature extraction model, including:

[0029] An acquisition module, used to acquire song audio information of a first sample song, comment information of the first sample song, and acquire comment information of a second sample song different from the first sample song;

[0030] an extraction module, configured to input the song audio information into an audio feature extraction model to be trained to obtain the song audio features of the first sample song, and input the comment information of the first sample song and the comment information of the second sample song into a comment feature extraction model to be trained to obtain the comment features of the first sample song based on the comment information of the first sample song and obtain the comment features of the second sample song based on the comment information of the second sample song;

[0031] The training module is used to train the song scoring model to be trained according to the similarity between the song audio features of the first sample song and the comment features of the first sample song, and the similarity between the song audio features of the first sample song and the comment features of the second sample song, so as to obtain a trained song scoring model.

[0032] In a fourth aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the above method are implemented.

[0033] In a fifth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps of the above method when executed by a processor.

[0034] In a sixth aspect, the present application further provides a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0035] The above-mentioned song scoring model training method, song scoring method, device, computer equipment, storage medium and computer program product use the audio information of sample songs and the comments corresponding to the sample songs to train the song scoring model, so that the song scoring model can learn the correlation between the song audio and the corresponding comments, thereby accurately screening out the target comments corresponding to the song to be scored based on the pre-trained song scoring model, thereby accurately screening out at least one target comment that meets the preset similarity condition with the song audio features of the song to be scored from the comment database, and combining the scoring information corresponding to each target comment to determine the song score of the song to be scored. It can be seen that the embodiment of the present application can analyze the correlation between the song comments and the song audio from the perspective of song reviews to score the song, so that the song scoring does not stop at pure audio analysis, and does not require a strict score quantization mechanism. By making full use of the comment information, the song can be scored more flexibly and variedly, thereby improving the accuracy of the song scoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A diagram of an application environment of a song scoring model training method in one embodiment;

[0038] Figure 2 A schematic diagram of a flow chart of a song scoring model training method in one embodiment;

[0039] Figure 3 A schematic diagram of a flow chart of a song scoring method based on a cross-modal pre-trained large model in one embodiment;

[0040] Figure 4 A structural block diagram of a song scoring model training device in one embodiment;

[0041] Figure 5 The figure is a diagram of the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0043] The song scoring method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The server 104 obtains the song audio information of the first sample song, and the comment information of the first sample song, and obtains the comment information of the second sample song that is different from the first sample song; the server 104 inputs the song audio information into the audio feature extraction model to be trained to obtain the song audio feature of the first sample song, and inputs the comment information of the first sample song and the comment information of the second sample song into the comment feature extraction model to be trained to obtain the comment feature of the first sample song based on the comment information of the first sample song and the comment feature of the second sample song based on the comment information of the second sample song; the server 104 trains the song scoring model to be trained according to the similarity between the song audio feature of the first sample song and the comment feature of the first sample song, and the similarity between the song audio feature of the first sample song and the comment feature of the second sample song, and obtains the trained song scoring model. The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

[0044] In an exemplary embodiment, Figure 2 As shown, a song scoring model training method is provided, the song scoring model includes an audio feature extraction model and a comment feature extraction model, and the method is applied to a server as an example for explanation. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0045] Step S202, obtaining song audio information of a first sample song and comment information of the first sample song, and obtaining comment information of a second sample song different from the first sample song.

[0046] The first sample song may refer to any song in the music platform. In actual applications, the first sample song may include songs with corresponding user comments and expert comments in the music platform.

[0047] The comment information of the first sample song may refer to information representing the evaluation of the singing performance of the first sample song. In practical applications, the comment information of the first sample song may include user comments input by users of the music platform for the first sample song and expert comments obtained by evaluating the first sample song based on expert experience. The user comments in the comment information of the first sample song may be obtained from the comment area of ​​the first sample song in the music platform, and the expert comments in the comment information of the first sample song may be obtained by music scorers (such as musicians, etc.). It is understandable that the expression form of the comment on the song may include but is not limited to text form and voice form.

[0048] The second sample song may refer to a song in the music platform that is different from the first sample song.

[0049] The comment information of the second sample song may refer to information representing an evaluation of the singing performance of the second sample song.

[0050] As an example, in order to make the song scores output by the song scoring model more objective and accurate, the server can train the song scoring model. Specifically, the training samples of the song scoring model to be trained may include a sample song data set in the music platform, and the sample song data set includes several sample songs. The server can use the song audio information of the first sample song, the comment information of the first sample song, and the comment information of the second sample song different from the first sample song as sample data to train the song scoring model.

[0051] Step S204, inputting the song audio information into the audio feature extraction model to be trained to obtain the song audio features of the first sample song, and inputting the comment information of the first sample song and the comment information of the second sample song into the comment feature extraction model to be trained to obtain the comment features of the first sample song based on the comment information of the first sample song and to obtain the comment features of the second sample song based on the comment information of the second sample song.

[0052] Among them, the audio feature extraction model may refer to a model used to determine the audio features of the song to be scored. In practical applications, the audio feature extraction model may encode the audio information of the song to be scored to obtain the audio features of the song to be scored. The audio feature extraction model may include an audio representation model. The audio representation model may also be spliced ​​with an MLP (Multilayer Perceptron) module or other neural networks used for data preprocessing, feature extraction and representation learning to form a comment feature extraction model. Among them, the audio representation model may include but is not limited to models such as the Mert model (a music analysis model) and the Huber model (a speech analysis model). The MLP module may adjust the representation dimension of the output result of the audio representation model to make the representation dimension of the output result of the comment feature extraction model and the audio feature extraction model consistent. The Mert model may refer to a large audio representation model that has been trained.

[0053] The song audio feature may refer to a feature that characterizes the audio content of the song to be rated. In practical applications, the song audio feature may include data obtained by encoding the audio of the song to be rated.

[0054] Among them, the comment feature extraction model may refer to a model used to determine the comment features of a comment. In practical applications, the comment feature extraction model may encode the text information corresponding to the comment to obtain the comment features corresponding to the comment. The comment feature extraction model may include a text representation model. The text representation model may also be spliced ​​with an MLP (Multilayer Perceptron) module or other neural networks used for data preprocessing, feature extraction and representation learning to form a comment feature extraction model. Among them, the text representation model may include but is not limited to models such as the bert model (a language model) and word2vector (another language model). The MLP module may adjust the representation dimension of the output result of the text representation model so that the representation dimension of the output result of the comment feature extraction model and the audio feature extraction model are consistent. It should be noted that if the target comment information is in voice form, it can be converted into text form and then the corresponding comment features can be extracted by the comment feature extraction model.

[0055] The comment feature may refer to information representing the content features of the comment.

[0056] As an example, the server can input the song audio information of the first sample song in the sample song data set into the audio feature extraction model to be trained, and the audio feature extraction model can extract the song audio features of the first sample song. Then the server can input the comment information of the first sample song and the comment information of the second sample song into the comment feature extraction model to be trained. The comment feature extraction model to be trained can obtain the comment features of the first sample song based on the comment information of the first sample song. The comment feature extraction model to be trained can obtain the comment features of the second sample song based on the comment information of the second sample song. In actual applications, the comment feature extraction model can include a BERT model and a first MLP module (a neural network). The BERT model can process the text information corresponding to the comment to obtain the mapping features corresponding to the comment. Then the first MLP module processes the mapping features corresponding to the comment to obtain the comment features corresponding to the comment.

[0057] Step S206, training the song scoring model to be trained according to the similarity between the song audio features and the comment features of the first sample song, and the similarity between the song audio features and the comment features of the second sample song, to obtain a trained song scoring model.

[0058] As an example, the server trains the song scoring model to be trained based on the similarity between the song audio features of the first sample song and the comment features of the first sample song, as well as the similarity between the song audio features of the first sample song and the comment features of the second sample song, and modifies the model parameters of the song scoring model until the model loss function value corresponding to the song scoring model is less than the preset loss function threshold, thereby obtaining a trained song scoring model.

[0059] In this embodiment, the audio information of sample songs and the comments corresponding to the sample songs are used to train a song scoring model, so that the song scoring model can learn the correlation between the song audio and the corresponding comments, thereby accurately screening out the target comments corresponding to the songs to be scored based on the pre-trained song scoring model, and analyzing the correlation between the comments and the song audio from the perspective of the comments to score the songs, so that the song scoring does not stop at pure audio analysis, and does not require a strict score quantization mechanism. By making full use of the comment information, the songs can be scored more flexibly and variedly, thereby improving the accuracy of the song scoring.

[0060] In some embodiments, the comment information of the first sample song includes the comment information of the first type and the comment information of the second sample song, the comment information of the second sample song includes the comment information of the first type and the comment information of the second type, the first type of representative comment information is obtained by the user through the music platform input, and the second type of representative comment information is obtained based on expert experience; according to the similarity between the song audio feature of the first sample song and the comment feature of the first sample song, and the similarity between the song audio feature of the first sample song and the comment feature of the second sample song, training the song scoring model to be trained includes:

[0061] Input the song audio information of the first sample song, the first type of comment information of the first sample song, and the first type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the first type of comment features of the first sample song based on the first type of comment information of the first sample song, and obtain the first type of comment features of the second sample song based on the first type of comment information of the second sample song; input the song audio information of the first sample song, the second type of comment information of the first sample song, and the second type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the second type of comment features of the first sample song based on the second type of comment information of the first sample song, and obtain the second type of comment features of the second sample song based on the second type of comment information of the second sample song; train the song scoring model to be trained at least according to the song audio features of the first sample song, the first type of comment features of the first sample song, the second type of comment features of the first sample song, the first type of comment features of the second sample song, and the second type of comment features of the second sample song.

[0062] Among them, the first type may refer to information representing the classification of the first comment information. In actual applications, the first type may represent that the rating information is obtained by the user inputting the song through the music platform.

[0063] The second type may refer to information characterizing the classification of the second comment information. In practical applications, the second type may characterize that the rating information is obtained based on expert experience (such as evaluation of a song by a musician).

[0064] The first type of comment feature refers to the one obtained by processing the first type of comment information by the song rating model, and the second type of comment feature refers to the one obtained by processing the second type of comment information by the song rating model.

[0065] As an example, the server can train the song scoring model for general evaluation and expert evaluation based on a large number of user comments on the sample songs in the music platform (first type of comment information) and the second type of comment information obtained by evaluating the sample songs based on expert experience. Specifically, the server can train the song scoring model based on the first type of comment information and the second type of comment information at the same time. For example: the server inputs the song audio information of the first sample song into the audio feature extraction model in the song scoring model to be trained to obtain the song audio feature of the first sample song, and inputs the first type of comment information corresponding to the first sample song and the first type of comment information corresponding to the second sample song into the comment feature extraction model in the song scoring model to be trained to obtain the first type of comment feature of the first sample song and the first type of comment feature of the second sample song. The server can train the song scoring model based on at least the song audio feature of the first sample song, the first type of comment feature of the first sample song, and the first type of comment feature of the second sample song to achieve general evaluation training.

[0066] When the server trains the song scoring model for expert evaluation, the server can input the song audio information of the first sample song into the audio feature extraction model in the song scoring model to be trained to obtain the song audio feature of the first sample song, and input the second type of comment information corresponding to the first sample song and the second type of comment information corresponding to the second sample song into the comment feature extraction model in the song scoring model to be trained to obtain the second type of comment feature of the first sample song and the second type of comment feature of the second sample song. The server can train the song scoring model based on at least the song audio feature of the first sample song, the second type of comment feature of the first sample song, and the second type of comment feature of the second sample song to achieve expert evaluation training.

[0067] In this embodiment, the song scoring model is trained based on sample audio and different types of comments, so that the song scoring model can learn the correlation between different types of comments and songs, optimize the performance of the song scoring model, and thus improve the accuracy of the song scores output by the song scoring model.

[0068] In an exemplary embodiment, training a song rating model to be trained based on at least a song audio feature of a first sample song, a first type comment feature of the first sample song, a second type comment feature of the first sample song, a first type comment feature of the second sample song, and a second type comment feature of the second sample song includes:

[0069] The song rating model to be trained is trained according to a first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, a second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, a third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and a fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

[0070] As an example, the server can perform general evaluation training and expert evaluation training on the song scoring model based on a large number of user comments on the sample songs in the music platform (first type of comment information) and the second type of comment information obtained by evaluating the sample songs based on expert experience. Specifically, the server can first calculate the first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, the second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, the third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and the fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song, and then train the song scoring model to be trained in combination with the first similarity, the second similarity, the third similarity and the fourth similarity.

[0071] In this embodiment, the song scoring model to be trained is trained based on the song audio features of the first sample song, the first type of comment features of the first sample song, the second type of comment features, and the first type of comment features of the second sample song, and the second type of comment features of the second sample song. This can enhance the generalization ability of the song scoring model, so that the song scoring model can be used to screen out accurate target comments, and then the song is scored in combination with the scores of the target comments, thereby improving the accuracy of the song scoring.

[0072] In some embodiments, training a song rating model to be trained based on at least the song audio feature of the first sample song, the first type comment feature of the first sample song, the second type comment feature of the first sample song, the first type comment feature of the second sample song, and the second type comment feature of the second sample song includes:

[0073] According to the first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, and the second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, the song rating model to be trained is trained to obtain a preliminary trained song rating model; according to the third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and the fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song, the preliminary trained song rating model is trained.

[0074] As an example, the server can perform general evaluation training and expert evaluation training on the song scoring model, such as the server can perform phased training on the song scoring model based on the first type of comment information and the second type of comment information, respectively. For example: in order to further optimize the model performance of the song scoring model, the server can train the song scoring model using the first similarity between the song audio feature of the first sample song and the first type of comment feature of the first sample song, and the second similarity between the song audio feature of the first sample song and the first type of comment feature of the second sample song, so as to achieve the first stage training (i.e., general evaluation training) to obtain a preliminarily trained song scoring model. Then, based on the preliminarily trained song scoring model obtained in the first stage training, the server trains the preliminarily trained song scoring model and fine-tunes the model parameters of the song scoring model according to the third similarity between the song audio feature of the first sample song and the second type of comment feature of the first sample song, and the fourth similarity between the song audio feature of the first sample song and the second type of comment feature of the second sample song, so as to achieve the second stage training (i.e., expert evaluation training), and obtain a trained song scoring model, so that the song scoring model can achieve more professional song scoring.

[0075] In this embodiment, by training the song scoring model in stages, the model parameters of the song scoring model can be gradually adjusted, so that the song scoring model learns expert evaluations on the basis of learning general evaluations to fine-tune model parameters, thereby optimizing the performance of the song scoring model and further improving the accuracy of the song scores output by the song scoring model.

[0076] In an exemplary embodiment, a song scoring method is provided, and the method is described by taking the application of the method to a server as an example. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0077] The song to be rated is input into a pre-trained song rating model to extract the song audio features of the song to be rated; wherein the pre-trained song rating model is trained by the above-mentioned song rating model training method.

[0078] The songs to be rated may refer to songs that need to be rated. In practical applications, the songs to be rated may include songs or audios sung by users on a music platform (such as a music platform with a karaoke function). The song audio features may refer to features that characterize the audio content of the songs to be rated. In practical applications, the song audio features may include data obtained by encoding the audio of the songs to be rated.

[0079] As an example, when a user needs to rate a song sung by the user on a music platform, the user can upload the song audio of the song to be rated to the server. The server receives the song audio of the song to be rated and extracts the song audio features of the song to be rated through the audio feature extraction model in the pre-trained song rating model.

[0080] In a preset comment database, at least one comment information whose comment features and the song audio features of the song to be rated meet the preset similarity conditions is determined as the target comment information; wherein the comment features of the comment information in the comment database are obtained by a pre-trained song rating model.

[0081] Among them, the comment database may refer to a pre-set database containing a number of comments. In practical applications, the comment information in the comment database may include comments based on expert experience, that is, expert comments (such as text / voice comments made by experts on songs). Expert comments can be obtained by collecting comments made on songs in the music platform based on expert experience. In practical applications, the content of expert comments may include appreciation and / or ratings of songs. In specific implementations, the server can calculate the distance between the audio features of the song to be rated and the comment features corresponding to each comment, and in combination with a preset distance threshold, use the comments corresponding to the comment features whose distance is less than the distance threshold as target comment information.

[0082] The target comment information may refer to comments associated with the song to be rated. In practical applications, the comment features of the target comment information may meet a preset similarity condition with the audio features of the song to be rated. The target comment information may be in text form or in voice form.

[0083] As an example, the server may collect several expert reviews in advance and construct a review database. When it is necessary to rate a song to be rated, the server may obtain the review features of each review in the review database, and determine the degree of similarity between the song audio features of the song to be rated and the review features of each review information. The server then uses the review whose degree of similarity meets the preset similarity condition as the target review. In actual applications, the review database may contain review information and / or review features of review information. The server may first obtain the review features of each review information in the review database, and then determine the degree of similarity between the song audio features of the song to be rated and the review features of each review. The server may also directly determine the degree of similarity between the song audio features of the song to be rated and the review features of each review based on the review features of each review.

[0084] The song score of the song to be rated is determined according to the rating information corresponding to each target review information.

[0085] The rating information may refer to information that has a corresponding relationship with the target comment and represents a rating value. In practical applications, each comment in the comment database may have one-to-one corresponding rating information, and each rating information may represent a specific rating value.

[0086] Among them, the song score may refer to information representing the singing score of the song to be scored. In practical applications, the song score can be used to judge the singing performance of the song to be scored. The higher the song score, the better the singing performance of the song.

[0087] As an example, the server can obtain the rating information corresponding to each target comment of the song to be rated, and then calculate the song rating of the song to be rated based on the above rating information. Specifically, the server can calculate the song rating of the song to be rated based on the weight and rating value corresponding to each target comment.

[0088] In the above song scoring method, at least one target comment is determined in a preset comment database according to the song audio features of the song to be scored, so that by comparing the comment features of each comment in the comment database with the song audio features, the target comments that meet the preset similarity conditions with the song audio features of the song to be scored are accurately screened out from the comment database, and the song score of the song to be scored is determined in combination with the scoring information corresponding to each target comment. From the perspective of the comment, the correlation between the comment and the song audio can be analyzed to score the song, so that the song scoring does not stop at pure audio analysis, and does not require a strict score quantification mechanism. By making full use of the comment information, the song can be scored more flexibly and variedly, thereby improving the accuracy of the song scoring.

[0089] In some embodiments, the song to be scored is input into a pre-trained song scoring model to extract song audio features of the song to be scored, including: mapping the song audio information of the song to be scored to a preset space through a feature mapping network in an audio feature extraction model in the song scoring model to obtain mapping features corresponding to the song audio information; converting the mapping features corresponding to the song audio information into song audio features of the song to be scored through a feature conversion network in the audio feature extraction model.

[0090] Among them, the audio feature extraction model may refer to a model used to determine the audio features of the song to be scored. In practical applications, the audio feature extraction model may encode the audio information of the song to be scored to obtain the audio features of the song to be scored. The audio feature extraction model may include an audio representation model. The audio representation model may also be spliced ​​with an MLP (Multilayer Perceptron) module or other neural networks used for data preprocessing, feature extraction and representation learning to form a comment feature extraction model. Among them, the audio representation model may include but is not limited to models such as the Mert model (a music analysis model) and the Huber model (a speech analysis model). The MLP module may adjust the representation dimension of the output result of the audio representation model to make the representation dimension of the output result of the comment feature extraction model and the audio feature extraction model consistent. The Mert model may refer to a large audio representation model that has been trained.

[0091] The feature mapping network may refer to a model for mapping the audio information of a song to a preset space. In practical applications, the feature mapping network may include any of the audio representation models such as the Mert model (a music analysis model) and the Huber model (a speech analysis model). The preset space may refer to a vector space for representing the audio features of a song.

[0092] Among them, the feature conversion network may refer to a model for converting feature dimensions of mapping features corresponding to song audio information. In practical applications, the feature conversion network may include an MLP module (a type of neural network).

[0093] As an example, the server can extract the song audio features of the song through the audio feature extraction model in the song scoring model. Specifically, the server can input the song audio information of the song to be scored into the pre-trained audio feature extraction model. The feature mapping network in the audio feature extraction model can map the song audio information of the song to be scored to a preset space to obtain the mapping features corresponding to the song audio information. Then, the feature conversion network in the audio feature extraction model can convert the mapping features corresponding to the song audio information into the song audio features of the song to be scored. In a specific implementation, the feature dimensions of the mapping features corresponding to the song audio information are different from the feature dimensions of the comment features corresponding to each comment. The server can use the feature conversion network (such as an MLP module) to process the mapping features corresponding to the song audio information so that the feature dimensions of the song audio features corresponding to the song audio information are the same as the feature dimensions of the comment features corresponding to each comment information.

[0094] In this embodiment, based on the feature extraction network and feature conversion network in the audio feature extraction model, the song audio information of the song to be rated is gradually mapped to a preset space and converted into song audio features of the song to be rated. The song audio information can be analyzed to obtain accurate song audio features, which facilitates accurate calculation of the similarity between the song audio features and the comment features corresponding to each comment, thereby screening out accurate target comments based on the similarity, thereby improving the accuracy of song ratings.

[0095] In some embodiments, the song score of the song to be rated is determined according to the rating information corresponding to each target comment information, including: according to the weight information corresponding to each target comment information, the rating information corresponding to each target comment information is weighted summed to obtain the song score of the song to be rated.

[0096] The weight information may refer to information representing the importance of each target comment.

[0097] As an example, each target comment may have corresponding rating information, and the server may calculate the song rating of the song to be rated based on the weight information pre-set for each target comment. For example: the weight information corresponding to the first target comment A of the song to be rated is f1, and the rating information corresponding to the first target comment A is x1. The weight information corresponding to the first target comment B of the song to be rated is f2, and the rating information corresponding to the first target comment A is x2. Then the song rating R of the song to be rated can be expressed as R=f1*x1+f2*x2. It can be understood that the sum of the weights corresponding to the target comments is 1. Then, when the weights corresponding to the target comments are equal, the song rating of the song to be rated can be expressed as the mean of the rating information corresponding to the target comments.

[0098] In this embodiment, the score information of each target comment is weighted and summed in combination with the weight information of each target comment, so that the song score of the song to be scored can be accurately calculated based on the weights of different comments, thereby improving the accuracy of the song score.

[0099] In some embodiments, the traditional song scoring scheme mainly uses the expert scoring method to train the model. The only prior information obtained by the model is the score. How to get this score depends entirely on the model's autonomous learning. The learning difficulty of the model is huge. The expert scoring considers a comprehensive and profound perspective. The model may only learn the surface. When using traditional methods to score songs, the requirements for data and models are very high. The expert scoring scores need to cover all segments, otherwise the model cannot learn the differences between different segments. Secondly, the strictness of the expert scoring needs to be consistent, otherwise the standard cannot be unified. However, the above two problems are difficult to take into account at the same time in practice. In order to solve the problem of low accuracy of song scoring obtained by traditional methods, Figure 3 As shown, a flow chart of a song scoring method based on a cross-modal pre-trained large model is provided. First, the server obtains the training samples of the song scoring model in the first training stage. The first training stage can be based on a large amount of data, with the purpose of allowing the song scoring model to learn a universal and comprehensive scoring method. In practical applications, the evaluation of the first training stage can be used as a general evaluation, which can be implemented based on the sample songs in the training samples and the comments corresponding to the sample songs.

[0100] It is understandable that the sample songs have corresponding comments in the music platform (such as user comments on the sample songs in the comment area corresponding to the sample songs, etc.). Although not all the comments in the music platform are objective and accurate, the comments in the music platform can roughly reflect the singing level of the sample songs. It is worth noting that each sample song and the corresponding comments of the sample song in the music platform (such as user comments) are one-to-one corresponding, such as each sample song and the corresponding comments of the sample song in the music platform can form a pair of data pairs with an associated relationship.

[0101] In the first training stage, the server can input the song audio information of each sample song and the first type of comment information (such as user comments) corresponding to each sample song into the song scoring model to be trained. The Mert model of the first branch in the song scoring model processes the song audio information of each sample song to obtain the mapping features (audio embedding) corresponding to the song audio information of each sample song. Then, the MLP module 1 of the first branch in the song scoring model processes the mapping features (audio embedding) corresponding to the song audio information of each sample song to obtain the song audio features (audio embedding1) of each sample song. The BERT model of the second branch in the song scoring model processes the first type of comment information corresponding to each sample song to obtain the mapping features (text embedding) corresponding to the first type of comment information corresponding to each sample song. Then, the MLP module 2 of the second branch in the song scoring model processes the mapping features (text embedding) corresponding to the first type of comment information corresponding to each sample song to obtain the first type of comment features (text embedding1) of each sample song. The feature dimensions of the mapping features output by the Mert model and the Bert model are inconsistent. The song audio features and comment features output by the module have the same feature dimension. Then the server calculates the contrast loss between audio embedding1 and text embedding2. The contrast loss function can be expressed as:

[0102] .

[0103] Among them, the numerator is the dot product of q with k+, which represents the distance between vector q and vector k+, and the denominator is the sum of the dot products of all positive examples and all negative examples. The parameter τ can be used to control the range of the loss function. Among them, q can be understood as the song audio feature of audio a corresponding to the sample song, k+ is the first type of comment information corresponding to audio a (such as user comments and other general evaluations), audio a (song audio features) and k+ constitute a pair of positive examples, a and other sample songs The first type of comment information constitutes multiple negative examples. Similarly, other sample songs and a The corresponding first type of comment information also constitutes negative examples. The learning goal of the song scoring model is to make the representation of audio embedding1 closer to the corresponding text embedding2 and farther away from the text embedding2 of other audios, so that the song scoring model can learn the correlation between the sample song and the comments of the sample song (such as user comments and other general evaluations).

[0104] The training of the song scoring model may also include a second training stage. After the song scoring model is trained in the first training stage until the model converges, it enters the second training stage. The second training stage may introduce expert evaluation. The data of expert evaluation is often more accurate than user comments, but the number is also less than user comments. Therefore, it is mainly used to fine-tune the model on the basis of the first training stage, so that the song scoring model can learn to score more professionally. Specifically, the server can input the song audio information of each sample song and the second type of comment information (such as expert comments) corresponding to each sample song into the song scoring model to be trained. The process of the song scoring model processing the song audio information of each sample song and the second type of comment information (such as expert comments) corresponding to each sample song can refer to the above-mentioned first training stage, so that the song scoring model completes the training of the second training stage and obtains a pre-trained song scoring model. It can be understood that the server may not use two training stages to train the song scoring model, that is, the server only uses general evaluations or expert comments to train the song scoring model.

[0105] After the song rating model training is completed, in order to obtain a quantitative score, the server can set a corresponding evaluation score (such as rating information) for each expert comment in the preset comment database. The server can determine the comment features of each expert comment in the preset comment database based on the pre-trained song rating model, and build a database based on the evaluation scores corresponding to each expert comment in the preset comment database. When it is necessary to score the song to be scored, the server can input the song audio information of the song to be scored and the expert comments in the preset comment database into the pre-trained song scoring model. The first branch of the pre-trained song scoring model (such as the song feature extraction model) uses the Mert model and the MLP module 1 to determine the song audio features of the song to be scored. The second branch of the pre-trained song scoring model (such as the comment feature extraction model) uses the Bert model and the MLP module 2 to determine the comment features of the expert comments in the preset comment database. The pre-trained song scoring model calculates the distance between the song audio features of the song to be scored and the comment features of the expert comments, and determines n comment features whose distance to the song audio features of the song to be scored is less than a preset distance threshold (or the n comment features whose distance to the song audio features of the song to be scored is closest), where n can be a positive integer greater than or equal to 1, and n is a preset parameter that can be flexibly adjusted based on actual needs. The server calculates the song scoring information of the song to be scored according to the average of the evaluation scores of the expert comments corresponding to the above n comment features.

[0106] In this embodiment, the song scoring model is trained based on two stages, so that the song scoring model can learn general scoring and evaluation rules through a large amount of non-expert data, and learn more professional scoring through a small amount of expert scoring data. The overall architecture of the model is a double-tower input structure, one is the singing audio input, and the other is the text input about the singing of the song. The two inputs are mapped to the embedding space of the same dimension after passing through the audio encoder and the text encoder respectively. Through comparative learning, the network learns the correlation between the singing audio and the comments. In the prediction stage, the similarity of the audio to be predicted and each text in the comment database is calculated, and the text with the closest similarity is taken as the evaluation result. After quantifying the evaluation result, the evaluation score is obtained, which can make full use of the text information of the comment, so that the song scoring no longer stays at the level of pure audio. The introduction of text information can more effectively guide the model to evaluate, and the score is more flexible and changeable, and no longer requires a strict score quantization mechanism to deal with the shortcomings of pure audio evaluation, thereby improving the accuracy of song scoring.

[0107] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0108] Based on the same inventive concept, the embodiment of the present application also provides a song scoring model training device for implementing the song scoring model training method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more song scoring model training device embodiments provided below can refer to the limitations of the song scoring model training method above, and will not be repeated here.

[0109] In an exemplary embodiment, Figure 4 As shown, a song scoring model training device is provided, wherein the song scoring model includes an audio feature extraction model and a comment feature extraction model, including: an acquisition module 402, an extraction module 404 and a training module 406, wherein:

[0110] The acquisition module 402 is used to acquire song audio information of a first sample song, comment information of the first sample song, and to acquire comment information of a second sample song different from the first sample song.

[0111] The extraction module 404 is used to input the song audio information into the audio feature extraction model to be trained to obtain the song audio features of the first sample song, and input the comment information of the first sample song and the comment information of the second sample song into the comment feature extraction model to be trained to obtain the comment features of the first sample song based on the comment information of the first sample song and obtain the comment features of the second sample song based on the comment information of the second sample song.

[0112] The training module 406 is used to train the song scoring model to be trained according to the similarity between the song audio features of the first sample song and the comment features of the first sample song, and the similarity between the song audio features of the first sample song and the comment features of the second sample song, so as to obtain a trained song scoring model.

[0113] In an exemplary embodiment, the comment information of the first sample song includes a first type of comment information and a second type of comment information; the comment information of the second sample song includes the first type of comment information and the second type of comment information; the first type of comment information is obtained by user input through a music platform, and the second type of comment information is obtained based on expert experience; the training module 406 is also specifically used to input the song audio information of the first sample song, the first type of comment information of the first sample song, and the first type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the first type of comment features of the first sample song based on the first type of comment information of the first sample song, and obtain the first type of comment features of the first sample song based on the first type of comment information of the second sample song. The first type of comment features of the second sample song are obtained based on the song audio information of the first sample song, the second type of comment features of the first sample song and the second type of comment information of the second sample song are input into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the second type of comment features of the first sample song based on the second type of comment information of the first sample song, and obtain the second type of comment features of the second sample song based on the second type of comment information of the second sample song; the song scoring model to be trained is trained according to at least the song audio features of the first sample song, the first type of comment features of the first sample song, the second type of comment features of the first sample song, the first type of comment features of the second sample song and the second type of comment features of the second sample song.

[0114] In an exemplary embodiment, the training module 406 is specifically used to train the song rating model to be trained based on a first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, a second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, a third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and a fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

[0115] In an exemplary embodiment, the training module 406 is specifically used to train the song scoring model to be trained according to a first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, as well as a second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, to obtain a preliminarily trained song scoring model; and to train the preliminarily trained song scoring model according to a third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, as well as a fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

[0116] Each module in the above song rating model training device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0117] In an exemplary embodiment, a song scoring device is provided, comprising: a feature extraction module, a comment determination module and a song scoring module, wherein:

[0118] The feature extraction module is used to input the song to be rated into a pre-trained song rating model to extract the song audio features of the song to be rated; wherein the pre-trained song rating model is trained by the above-mentioned song rating model training method.

[0119] A comment determination module is used to determine at least one comment information whose comment features and the song audio features of the song to be rated meet the preset similarity conditions in a preset comment database as target comment information; wherein the comment features of the comment information in the comment database are obtained by the pre-trained song rating model.

[0120] The song scoring module is used to determine the song score of the song to be scored according to the scoring information corresponding to each target comment information.

[0121] In an exemplary embodiment, the feature extraction module is also specifically used to map the song audio information of the song to be scored to a preset space through a feature mapping network in an audio feature extraction model in the song scoring model to obtain mapping features corresponding to the song audio information; and convert the mapping features corresponding to the song audio information into song audio features of the song to be scored through a feature conversion network in the audio feature extraction model.

[0122] In an exemplary embodiment, the song rating module is further used to perform weighted summation on the rating information corresponding to each target comment information according to the weight information corresponding to each target comment information, so as to obtain the song rating of the song to be rated.

[0123] Each module in the above song scoring device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0124] The embodiment of the present application also provides a song scoring model training device for implementing the song scoring model training method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more song scoring model training device embodiments provided below can refer to the limitations of the song scoring model training method above, and will not be repeated here.

[0125] In an exemplary embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 5 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a song scoring model training method is implemented. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0126] Those skilled in the art will understand that Figure 5The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0127] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0128] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0129] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0130] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0131] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0132] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A song scoring model training method, characterized in that: The song scoring model includes an audio feature extraction model and a comment feature extraction model, and the method includes: Acquire song audio information of a first sample song and comment information of the first sample song, and acquire comment information of a second sample song different from the first sample song; Inputting the song audio information into the audio feature extraction model to be trained to obtain the song audio features of the first sample song, and inputting the comment information of the first sample song and the comment information of the second sample song into the comment feature extraction model to be trained to obtain the comment features of the first sample song based on the comment information of the first sample song and obtain the comment features of the second sample song based on the comment information of the second sample song; According to the similarity between the song audio features of the first sample song and the comment features of the first sample song, and the similarity between the song audio features of the first sample song and the comment features of the second sample song, the song scoring model to be trained is trained to obtain a trained song scoring model.

2. The method according to claim 1, characterized in that The comment information of the first sample song includes comment information of the first type and comment information of the second type; the comment information of the second sample song includes comment information of the first type and comment information of the second type; the first type represents comment information obtained by user input through the music platform, and the second type represents comment information obtained based on expert experience; The training of the song rating model to be trained according to the similarity between the song audio feature of the first sample song and the comment feature of the first sample song, and the similarity between the song audio feature of the first sample song and the comment feature of the second sample song, comprises: Inputting the song audio information of the first sample song, the first type of comment information of the first sample song, and the first type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the first type of comment features of the first sample song based on the first type of comment information of the first sample song, and obtain the first type of comment features of the second sample song based on the first type of comment information of the second sample song; Inputting the song audio information of the first sample song, the second type of comment information of the first sample song, and the second type of comment information of the second sample song into the song scoring model to be trained, so as to obtain the song audio features of the first sample song based on the song audio information of the first sample song, obtain the second type of comment features of the first sample song based on the second type of comment information of the first sample song, and obtain the second type of comment features of the second sample song based on the second type of comment information of the second sample song; The song rating model to be trained is trained based on at least the song audio features of the first sample song, the first type of comment features of the first sample song, the second type of comment features of the first sample song, the first type of comment features of the second sample song, and the second type of comment features of the second sample song.

3. The method according to claim 2, characterized in that The step of training the song rating model to be trained based on at least the song audio feature of the first sample song, the first type comment feature of the first sample song, the second type comment feature of the first sample song, the first type comment feature of the second sample song, and the second type comment feature of the second sample song comprises: The song rating model to be trained is trained based on a first similarity between the song audio features of the first sample song and the first type of comment features of the first sample song, a second similarity between the song audio features of the first sample song and the first type of comment features of the second sample song, a third similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and a fourth similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

4. The method according to claim 2, characterized in that: The step of training the song rating model to be trained based on at least the song audio feature of the first sample song, the first type comment feature of the first sample song, the second type comment feature of the first sample song, the first type comment feature of the second sample song, and the second type comment feature of the second sample song comprises: According to a first similarity between the song audio feature of the first sample song and the first type of comment feature of the first sample song, and a second similarity between the song audio feature of the first sample song and the first type of comment feature of the second sample song, the song scoring model to be trained is trained to obtain a preliminarily trained song scoring model; The preliminarily trained song rating model is trained based on a third degree of similarity between the song audio features of the first sample song and the second type of comment features of the first sample song, and a fourth degree of similarity between the song audio features of the first sample song and the second type of comment features of the second sample song.

5. A song scoring method, characterized in that: The method comprises: Inputting the song to be rated into a pre-trained song rating model to extract the song audio features of the song to be rated; wherein the pre-trained song rating model is trained by the method described in any one of claims 1 to 4; In a preset review database, at least one review information whose review feature and the song audio feature of the song to be rated meet a preset similarity condition is determined as target review information; wherein the review feature of the review information in the review database is obtained by the pre-trained song rating model; The song score of the song to be rated is determined according to the rating information corresponding to each target review information.

6. The method according to claim 5, characterized in that The step of inputting the song to be rated into a pre-trained song rating model to extract the song audio features of the song to be rated includes: Mapping the song audio information of the song to be rated to a preset space through a feature mapping network in an audio feature extraction model in the song rating model to obtain mapping features corresponding to the song audio information; The mapping features corresponding to the song audio information are converted into song audio features of the song to be rated through the feature conversion network in the audio feature extraction model.

7. The method according to claim 5, characterized in that Determining the song score of the song to be rated according to the rating information corresponding to each target comment information includes: According to the weight information corresponding to each of the target review information, the score information corresponding to each of the target review information is weighted and summed to obtain the song score of the song to be rated.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images