Model training method, song labeling method, medium, device and computing device

Abstract

Embodiments of the present disclosure provide a model training method, a song labeling method, a medium, an apparatus and a computing device. The model training method comprises: a computing device obtaining a first training set, and performing multi-round model training on a first initial model through the first training set to obtain a song evaluation model. In any round of training process, the first label is obtained by inputting the evaluation data and the singing quality feature of the first sample song into the second initial model after the current round of training, the singing quality feature is an intermediate result obtained by inputting the audio signal of the first sample song into the first initial model after the last round of training, and the song evaluation model is used to determine the target label of the to-be-labeled song according to the audio signal of the to-be-labeled song, and the target label is used to indicate the singing quality of the to-be-labeled song. The technical solution effectively improves the accuracy of the song evaluation model.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the field of data processing, and more specifically, the embodiments of this disclosure relate to model training methods, song annotation methods, media, apparatus, and computing devices. Background Technology

[0002] This section is intended to provide background or context for embodiments of this disclosure. The description herein is not intended to imply that it is prior art simply because it is included in this section.

[0003] In the internet age, users can not only listen to songs on music websites, but also sing and upload their own songs, greatly satisfying people's music needs. Among these features, song rating is a technology that provides real-time feedback on users' individual performances. It can evaluate the quality of a singer's performance, bringing users a novel experience while also helping them understand and improve their singing skills.

[0004] Currently, existing technologies mainly obtain song ratings through models, but the accuracy of labeling sample songs during model training is low. Summary of the Invention

[0005] This disclosure provides a model training method, a song annotation method, a medium, an apparatus, and a computing device to solve the problem of low accuracy in the annotation of sample songs in the prior art.

[0006] In a first aspect of this disclosure, a model training method is provided, comprising:

[0007] Obtain a first training set, which includes multiple first sample songs, the audio signal of each first sample song, and a first tag;

[0008] The first initial model is trained multiple times using the first training set to obtain a singing evaluation model. In any training round, the first label is obtained by inputting the evaluation data and singing quality features of the first sample song into the second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled.

[0009] In one embodiment of this disclosure, the method further includes:

[0010] The second initial model is trained multiple times using the second training set to obtain the labeled model. In any training round, the second training set includes multiple second sample songs, evaluation data for each second sample song, a second label, and performance quality features. The second label is manually labeled to characterize the performance quality of the second sample song. The labeled model is used to determine the target label of the song to be labeled based on the performance quality features and evaluation data of the song to be labeled.

[0011] In one embodiment of this disclosure, the step of performing multiple rounds of model training on the second initial model using a second training set to obtain a labeled model includes:

[0012] In the first round of training, the second initial model is trained using multiple second sample songs, the evaluation data of each second sample song, and the second label to obtain the second initial model after the first round of training.

[0013] In any subsequent training round, the second initial model after the previous training round is trained using multiple second sample songs and the evaluation data, second labels, and singing quality features of each second sample song to obtain the labeled model.

[0014] In one embodiment of this disclosure, after training the second initial model using multiple second sample songs, evaluation data of each second sample song, and second labels in the first round of training to obtain the second initial model after the first round of training, the method further includes:

[0015] In the first round of training, multiple first sample songs and the evaluation data of each first sample song are input into the second initial model after the first round of training to obtain the first label of the first sample song;

[0016] In the first round of training, multiple first sample songs, along with the first label and audio signal of each first sample song, are determined as the first training set of the first initial model.

[0017] In one embodiment of this disclosure, the method further includes:

[0018] In any round of training, multiple first sample songs, the evaluation data of each first sample song, and the singing quality features are input into the second initial model after this round of training, and the first label of the first sample song is updated.

[0019] In one embodiment of this disclosure, the method further includes:

[0020] In any round of training, multiple first sample songs and the audio signal of each first sample song are input into the first initial model after the previous round of training to update the singing quality features of the first sample songs.

[0021] In any training round, multiple second sample songs and the audio signal of each second sample song are input into the first initial model after the previous training round to update the singing quality features of the second sample songs.

[0022] In a second aspect of this disclosure, a song annotation method is provided, comprising:

[0023] Obtain the audio signal of the song to be labeled;

[0024] The audio signal of the song to be labeled is input into the singing evaluation model to obtain the target label of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. The singing evaluation model is trained using the methods provided in the first aspect and various possible designs.

[0025] In one embodiment of this disclosure, the step of inputting the audio signal of the song to be labeled into a singing evaluation model to obtain the target label of the song to be labeled includes:

[0026] The audio signal of the song to be labeled is input into the singing evaluation model to obtain the singing quality characteristics of the song to be labeled.

[0027] The singing quality features and evaluation data of the song to be labeled are input into the labeling model to obtain the target label of the song to be labeled. The labeling model is trained using the methods provided in the first aspect and various possible designs.

[0028] In a third aspect of this disclosure, a model training apparatus is provided, comprising:

[0029] The acquisition module is used to acquire a first training set, which includes multiple first sample songs, the audio signal of each first sample song, and a first tag.

[0030] The training module is used to train the first initial model multiple times using the first training set to obtain a singing evaluation model. In any training round, the first label is obtained by inputting the evaluation data and singing quality features of the first sample song into the second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled.

[0031] In one embodiment of this disclosure, the training module is further configured to:

[0032] The second initial model is trained multiple times using the second training set to obtain the labeled model. In any training round, the second training set includes multiple second sample songs, evaluation data for each second sample song, a second label, and performance quality features. The second label is manually labeled to characterize the performance quality of the second sample song. The labeled model is used to determine the target label of the song to be labeled based on the performance quality features and evaluation data of the song to be labeled.

[0033] In one embodiment of this disclosure, the training module is specifically used for:

[0034] In the first round of training, the second initial model is trained using multiple second sample songs, the evaluation data of each second sample song, and the second label to obtain the second initial model after the first round of training.

[0035] In any subsequent training round, the second initial model after the previous training round is trained using multiple second sample songs and the evaluation data, second labels, and singing quality features of each second sample song to obtain the labeled model.

[0036] In one embodiment of this disclosure, after the second initial model is trained using multiple second sample songs, evaluation data of each second sample song, and second labels in the first round of training to obtain the second initial model after the first round of training, the apparatus further includes:

[0037] The input module is used to input multiple first sample songs and the evaluation data of each first sample song into the second initial model after the first round of training in the first round of training, so as to obtain the first label of the first sample song;

[0038] The determination module is used to determine, in the first round of training, multiple first sample songs and the first label and audio signal of each first sample song as the first training set of the first initial model.

[0039] In one embodiment of this disclosure, the input module is further configured to:

[0040] In any round of training, multiple first sample songs, the evaluation data of each first sample song, and the singing quality features are input into the second initial model after this round of training, and the first label of the first sample song is updated.

[0041] In one embodiment of this disclosure, the input module is further configured to:

[0042] In any round of training, multiple first sample songs and the audio signal of each first sample song are input into the first initial model after the previous round of training to update the singing quality features of the first sample songs.

[0043] In any training round, multiple second sample songs and the audio signal of each second sample song are input into the first initial model after the previous training round to update the singing quality features of the second sample songs.

[0044] In a fourth aspect of this disclosure, a song annotation device is provided, comprising:

[0045] The acquisition module is used to acquire the audio signal of the song to be labeled;

[0046] The input module is used to input the audio signal of the song to be labeled into the singing evaluation model to obtain the target label of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. The singing evaluation model is trained using the methods provided in the first aspect and various possible designs.

[0047] In one embodiment of this disclosure, the input module is specifically used for:

[0048] The audio signal of the song to be labeled is input into the singing evaluation model to obtain the singing quality characteristics of the song to be labeled.

[0049] The singing quality features and evaluation data of the song to be labeled are input into the labeling model to obtain the target label of the song to be labeled. The labeling model is trained using the methods provided in the first aspect and various possible designs.

[0050] In a fifth aspect of the present disclosure, a storage medium is provided that stores computer program instructions, which, when executed, implement the methods provided in the first aspect, the second aspect, and various possible designs.

[0051] In a sixth aspect of this disclosure, a computing device is provided, comprising: a processor, and a memory communicatively connected to the processor;

[0052] The memory stores computer-executed instructions;

[0053] The processor executes computer execution instructions stored in the memory to implement the methods provided by the first aspect, the second aspect, and various possible designs.

[0054] According to the model training method, song annotation method, medium, apparatus, and computing device of this disclosure, the model training method includes: the computing device acquiring a first training set, and performing multiple rounds of model training on a first initial model using the first training set to obtain a singing evaluation model. In any round of training, the first label is obtained by inputting the evaluation data and singing quality features of a first sample song into a second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be annotated based on the audio signal of the song to be annotated. The target label is used to describe the singing quality of the song to be annotated. In this technical solution, since the first training set and the second training set come from completely different data, the process of iteratively training the first initial model and the second initial model continuously integrates these two different training sets, thereby significantly improving the accuracy of the singing evaluation model obtained through model training. Attached Figure Description

[0055] The above and other objects, features, and advantages of this disclosure will become readily apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. Several embodiments of this disclosure are illustrated in the drawings by way of example and not limitation, in which:

[0056] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of the present disclosure;

[0057] Figure 2 A flowchart of a model training method provided in an embodiment of this disclosure;

[0058] Figure 3 A structural diagram of a first initial model provided in an embodiment of this disclosure;

[0059] Figure 4 This is a structural diagram of a second initial model provided in an embodiment of the present disclosure;

[0060] Figure 5 A flowchart of a song annotation method provided in an embodiment of this disclosure;

[0061] Figure 6 A structural diagram of a storage medium provided in an embodiment of this disclosure;

[0062] Figure 7 This is a structural diagram of a model training apparatus provided in an embodiment of the present disclosure;

[0063] Figure 8 This is a structural diagram of a song annotation device provided in an embodiment of the present disclosure;

[0064] Figure 9This is a structural diagram of a computing device provided in an embodiment of the present disclosure.

[0065] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0066] The principles and spirit of this disclosure will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0067] Those skilled in the art will recognize that the embodiments of this disclosure can be implemented as a model training method, a song annotation method, a medium, an apparatus, and a computing device. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0068] According to embodiments of this disclosure, a model training method, a song annotation method, a medium, an apparatus, and a computing device are proposed.

[0069] In this article, it is important to understand the following terms:

[0070] Performance quality: An overall evaluation of the pleasing quality of a musical work, taking into account factors such as pitch, rhythm, singing technique, and overall listening experience.

[0071] Dry audio: Unaccompanied, unprocessed pure vocals.

[0072] Furthermore, the number of any elements in the accompanying drawings is for illustrative purposes only and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0073] In addition, the data involved in this disclosure may be data authorized by the user or fully authorized by all parties. The collection, dissemination and use of the data shall comply with the requirements of relevant national laws and regulations. The implementation methods / executives of this disclosure may be combined with each other.

[0074] The principles and spirit of this disclosure will be explained in detail below with reference to several representative embodiments. Invention Overview

[0076] The inventors have discovered that currently, scoring songs typically requires the use of a model. Specifically, the song to be scored is input into the model, which then outputs the score for that song. Before using this model, it is necessary to construct an audio dataset and train the model using this dataset. It should be understood that a high-quality audio dataset containing a large number of sample songs is an essential foundation for building and training models that process audio-related tasks; therefore, how to construct an audio dataset is extremely important.

[0077] Currently, existing technologies mainly employ the following methods to construct audio datasets:

[0078] Option 1: Implemented through entirely manual evaluation. After collecting a certain number of sample songs, a group of music experts with singing knowledge or experience are organized as testers to listen to and score the sample songs. Furthermore, to ensure the accuracy of the scoring as much as possible, the listening environment conditions for each tester need to be identical, and the scoring standards should be standardized beforehand. In addition, the number of testers participating in the scoring needs to be sufficient; typically, the average score given by each tester for the sample song is used as the label for that sample song.

[0079] However, this approach has the following problems: it is costly, requiring a lot of manpower and time, making it difficult to annotate a large number of sample songs to obtain labels for each sample song; it is highly subjective, as the songs are scored entirely by humans, inevitably introducing subjective factors into the scoring process, such as giving low scores to disliked melodies; and the audio dataset has poor scalability. If the size of the audio dataset is to be expanded, it is necessary to reproduce the listening environment and testers, otherwise the scoring standards may be inconsistent each time, resulting in low annotation accuracy.

[0080] Option 2: For sample songs with reference signals (such as the original recording or excellent performance samples), the sample song can be labeled using machine learning, based on the reference signal, to obtain its tag. This option typically relies on objective features contained in the reference signal, such as pitch, rhythm, and loudness. The more similar the sample song is to the reference signal, the better the performance is considered. Simply quantifying the degree of similarity is sufficient to use it as a tag; this method is low-cost and allows for the labeling of a large number of sample songs in a short time.

[0081] However, this scheme has the following problems: since it can only rely on some objective features to simulate human auditory perception, it cannot fully consider various influencing factors, the evaluation dimension is singular, and there is inevitably a certain gap with manual annotation; defining the similarity with the reference signal as the quality of the song is not always valid, it is not applicable to all situations, and it is often only applicable to singers whose singing ability is not very good.

[0082] Option 3: Use karaoke platforms to collect user ratings of sample songs to replace the scoring by testers in Option 1. Based on a large-scale collection of user ratings, the average score for each sample song can be used as its tag. Alternatively, metrics such as likes, comments, and shares for each sample song can be converted into tags. This option does not require a dedicated team of testers for manual scoring, and the audio dataset is highly scalable.

[0083] However, this method has the following problems: the scores are only valuable when the sample songs are popular enough and have a sufficient number of users rating them. For songs with low popularity, it is difficult to guarantee the quality of the labels, limiting the coverage of the audio dataset and causing the data distribution in the audio dataset to not reflect reality. For example, it is often difficult to collect enough user ratings for sample songs with poor singing quality, thus failing to reflect the true data distribution and making it difficult to guarantee the accuracy of the subsequently determined labels. At the same time, different users have different listening environments and varying levels of musical literacy, and are easily influenced by factors unrelated to singing quality, such as the singer's popularity, making it difficult to guarantee the quality of the sample song labels.

[0084] In summary, existing technologies suffer from low accuracy in labeling sample songs.

[0085] To address the aforementioned issues, this disclosure provides a model training method, a song annotation method, a medium, a device, and a computing device. Based on a multi-model loop, multiple rounds of model training are performed on a first initial model and a second initial model to obtain a song evaluation model. In practical applications, the audio signal of the song to be annotated can be input into the song evaluation model to obtain the labels output by the model, replacing manual processing. By considering various influencing factors, the accuracy of song annotation is effectively improved.

[0086] After introducing the basic principles of this disclosure, various non-limiting embodiments of this disclosure will be described in detail below.

[0087] Application Scenarios Overview

[0088] First refer to Figure 1 Examples of application scenarios for the solutions provided in this disclosure are given. Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of the present disclosure, such as... Figure 1 As shown, in this application scenario, the scenario includes a first initial model and a second initial model. The first initial model is trained based on a first training set, and the second initial model is trained based on a second training set.

[0089] In each training round, the first label in the first training set is generated using the second initial model obtained in this round, while the singing quality features in the second training set are generated using the first initial model obtained in the previous round. In other words, multiple models need to be trained in a loop between the first and second initial models.

[0090] In any round of training in this embodiment, it is first necessary to obtain a second training set for this round of training. The second training set includes evaluation data of the second sample song, performance quality features (not present in the first round), and a second label. The performance quality features of the second sample song are generated by the first initial model from the previous round of training. The generation process is as follows: the audio signal of the second sample song is input into the first initial model obtained from the previous round of training, thereby obtaining its output performance quality features.

[0091] Furthermore, the second initial model for this round is trained using the second training set to obtain the second initial model after this round of training. Next, the evaluation data and performance quality features of the first sample song (not present in the first round) are input into the second initial model after this round of training to obtain the first label of the first sample song, thereby updating the first training set. The updated first training set is then used to train the first initial model for this round to obtain the first initial model after this round of training. Afterwards, the audio signal of the first sample song is input into the second initial model after this round of training to update the performance quality features of the first sample song, and the audio signal of the second sample song is input into the second initial model after this round of training to update the performance quality features of the second sample song, thus updating the second training set.

[0092] The first and second initial models are trained in multiple rounds using the above method until both models converge.

[0093] Exemplary methods

[0094] The following is combined Figure 1 Application scenarios, refer to Figure 1 This document describes a model training method and a song annotation method according to exemplary embodiments of the present disclosure. It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in any way. Rather, the embodiments of the present disclosure can be applied to any applicable scenario.

[0095] First, the model training process will be introduced through specific examples.

[0096] Figure 2This is a flowchart illustrating a model training method provided in one embodiment of this disclosure. The method of this embodiment can be applied to a computing device, which may be a server or server cluster, or a terminal device. Figure 2 As shown, the method in this embodiment includes:

[0097] S201. Obtain the first training set.

[0098] In this step, song annotation is achieved through a vocal evaluation model. Therefore, before practical application, the initial model needs to be trained to obtain the vocal evaluation model. Before training the model, a first training set needs to be obtained for model training.

[0099] The first training set includes multiple first sample songs, the audio signal of each first sample song, and a first label.

[0100] The audio signal of the first sample song can be the dry audio signal of that song. Since the dry audio signal does not contain accompaniment and has not undergone post-processing, it effectively avoids the influence of irrelevant factors on the training process.

[0101] Optionally, the first sample song and its audio signal may be obtained from the local storage space of the computing device, through interaction with the user's terminal device, or from other service platforms. For example, the other service platform may be the service platform corresponding to a music website, or the service platform corresponding to a video website, or other music-related service platforms that can provide evaluation services to users. This disclosure does not impose specific limitations on the method of obtaining the first sample song and its audio signal.

[0102] S202. Train the first initial model multiple times using the first training set to obtain the singing evaluation model.

[0103] In this step, since the first initial model and the second initial model in this disclosure are trained cyclically, meaning that the training process of the first initial model depends on the second initial model, it is necessary to generate a first training set for each round of training of the first initial model based on the second initial model, so that the first initial model can be trained multiple times based on the first training set to obtain the singing evaluation model.

[0104] In any training round, the first label is obtained by inputting the evaluation data and singing quality features of the first sample song into the second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled.

[0105] For example, the target label can be a score, such as 60, 70, 80, etc., or a level, such as A, B, C; or, pleasant to listen to, average, unpleasant to listen to, etc. This disclosure does not limit the specific form of the target label.

[0106] Furthermore, the output of the first initial model will also serve as the input to the second initial model, participating in its training process. Therefore, in any training round, multiple second sample songs and their audio signals can be input into the first initial model after the previous training round to update the performance quality features of the second sample songs. This updated performance quality features will then be used to update the second training set, which will then participate in the training of the second initial model in the next round.

[0107] Optionally, multiple first sample songs and the audio signal of each first sample song can be input into the first initial model after the previous round of training to update the performance quality features of the first sample songs, so that the multiple first sample songs, the performance quality features of each first sample song, and the evaluation data can be input into the second initial model after the current round of training to update the first label of each first sample song in the first training set.

[0108] In the first training set, there are a large number of first sample songs, while in the second training set, there are a small number of second sample songs. Only a small number of sample songs need to be manually labeled to achieve the purpose of training the first and second initial models.

[0109] The model training method provided in this embodiment involves a computing device acquiring a first training set and performing multiple rounds of model training on a first initial model using the first training set to obtain a singing evaluation model. In any round of training, the first label is obtained by inputting the evaluation data and singing quality features of a first sample song into a second initial model after that round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. In this technical solution, since the first and second training sets come from completely different data, the process of iteratively training the first and second initial models continuously integrates these two different training sets, significantly improving the accuracy of the singing evaluation model obtained through model training.

[0110] Based on the above embodiments, while training the first initial model, it is also necessary to train the second initial model.

[0111] The training of the second initial model mainly includes the following steps:

[0112] The second initial model is trained multiple times using a second training set to obtain a labeled model. In each training round, the second training set includes multiple second sample songs, evaluation data for each second sample song, second labels, and performance quality features. The second labels are manually annotated and represent the performance quality of the second sample songs. The labeling model is used to determine the target label for the song to be labeled based on its performance quality features and evaluation data.

[0113] The evaluation data for each second sample song can be obtained from the local storage space of the computing device, through communication and interaction with the user's terminal device, or from other service platforms. For example, these other service platforms could be service platforms corresponding to music websites, or service platforms corresponding to video websites, or other music-related service platforms that provide evaluation services to users. This embodiment of the disclosure does not impose specific limitations on the method of obtaining the evaluation data for the second sample songs.

[0114] For example, the evaluation data is used to represent the user's listening experience of the second sample song. It may include the performance score data of the second sample song, such as the user's score for the pitch, rhythm, and emotion of the second sample song. It may also include the user's behavioral data, such as the number of likes, comments, shares, and specific comment content, such as "good", "off-key", "divine voice", etc. The specific content included in the evaluation data can be determined according to the actual situation, and this embodiment of the disclosure does not impose specific limitations on it.

[0115] The second label for each second sample song is obtained through manual annotation and is used to characterize the singing quality of the second sample song. The second label can be a score, such as 60, 70, 80, etc., or a grade, such as A, B, C; or, pleasant, average, unpleasant, etc. This embodiment of the disclosure does not limit the specific form of the second label.

[0116] In one possible implementation, training the second initial model can be achieved through the following steps (1) and (2):

[0117] Step (1): In the first round of training, the second initial model is trained using multiple second sample songs, the evaluation data of each second sample song, and the second label to obtain the second initial model after the first round of training.

[0118] Since the first round of training is the initial stage of model training and there is no previous round of training, the second training set of the second initial model in this round of training only contains multiple second sample songs, evaluation data for each second sample song, and second labels.

[0119] Furthermore, in the first round of training, after obtaining the second initial model after the first round of training, multiple first sample songs and the evaluation data of each first sample song can be input into the second initial model after the first round of training to obtain the first label of the first sample song. Then, the multiple first sample songs, the first label of each first sample song, and the audio signal are determined as the first training set of the first initial model. It should be understood that the first training set here is the training set required by the first initial model in the first round of training.

[0120] Step (2) In any subsequent training round, the second initial model after the previous training round is trained using multiple second sample songs and the evaluation data, second labels, and singing quality features of each second sample song to obtain the labeled model.

[0121] Furthermore, similar to step (1), in each training round in step (2), after training the second initial model after the previous round of training to obtain the second initial model after this round of training, multiple first sample songs and the evaluation data of each first sample song can be input into the second initial model after this round of training to obtain the first label of the first sample song. Then, the multiple first sample songs and the first label and audio signal of each first sample song are updated to the first training set of the first initial model for this round.

[0122] Furthermore, in any training round, multiple first sample songs and their audio signals can be input into the first initial model after the previous training round to update the performance quality features of the first sample songs. Simultaneously, in any training round, multiple second sample songs and their audio signals can be input into the first initial model after the previous training round to update the performance quality features of the second sample songs.

[0123] In this approach, since the first and second training sets originate from completely different datasets, the training processes continuously integrate these two different sources, leading to a continuous improvement in the labeling accuracy of the second initial model. During iterative training, as the labeling accuracy of the first label output by the second initial model increases, the first initial model extracts performance quality features more accurately. Simultaneously, the manually labeled first labels from the first training set serve as supervision during training, guiding both the first and second initial models to learn in the desired direction. This iterative process is then repeated until the accuracy of both the first and second initial models converges.

[0124] Based on the above embodiments, during the training of the first initial model, it is also necessary to train the second initial model in order to provide the singing quality features required during the training of the first initial model.

[0125] The training of the second initial model mainly includes the following steps:

[0126] The second initial model is trained multiple times using the second training set to obtain the labeled model. In any training round, the second training set includes multiple second sample songs, evaluation data for each second sample song, second labels, and performance quality features. The second labels are manually labeled and used to characterize the performance quality of the second sample songs. The labeling model is used to determine the target label of the song to be labeled based on the performance quality features and evaluation data of the song to be labeled.

[0127] It should be understood that the second tag has the same expression as the first tag.

[0128] Since the first and second initial models are trained cyclically, in each training round, the output of the second initial model serves as the input to the first initial model, participating in its training process. Specifically, in any given training round, multiple first sample songs, their evaluation data, and performance quality features are input into the second initial model after this round of training. This updates the first labels of the first sample songs, allowing the first training set to be updated based on these updated labels, thus participating in the training process of the first initial model in this round.

[0129] Next, the structure of the singing evaluation model and the annotation model will be introduced.

[0130] The model training process will be explained using a specific example.

[0131] Step 1: Obtain relevant data for the first sample song and relevant data for the second sample song.

[0132] For example, the relevant data for the first sample song can be found in Table 1.

[0133] Table 1

[0134] character data B1 audio signal B2 Evaluation data B3 Singing quality characteristics B5 First tag

[0135] Referring to Table 1, the first label B5 and the singing quality feature B3 will be updated with each round of training.

[0136] For example, the relevant data for the second sample song can be found in Table 2.

[0137] Table 2

[0138] character data A1 audio signal A2 Evaluation data A3 Singing quality characteristics A4 Second tag

[0139] Referring to Table 2, the singing quality characteristic A3 will be updated with each round of training.

[0140] Step 2: Train the second initial model based on the second training set (evaluation data and second labels of the second sample songs) to obtain the second initial model after the first round of training.

[0141] Step 3: Input the evaluation data of the first sample song into the second initial model after the first round of training, and obtain the first label of the first sample song output by it.

[0142] Step 4: Train the first initial model based on the first training set (the first label and audio signal of the first sample song) to obtain the first initial model after the first round of training.

[0143] Step 5: Input the audio signal of the second sample song into the first initial model after the first round of training, and update the singing quality features of the second sample song to the second training set.

[0144] Step 6: Input the audio signal of the first sample song into the first initial model after the first round of training, and update the singing quality features of the first sample song in Table 1.

[0145] It should be understood that steps 2 to 6 constitute the first round of model training, while steps 7 to 11 below constitute the training process for any round after the first round.

[0146] Step 7: Train the second initial model obtained in the previous training round based on the second training set (evaluation data of the second sample songs, the second label, and the singing quality features) to obtain the second initial model after this round of training.

[0147] Step 8: Input the evaluation data and singing quality features of the first sample song into the second initial model after this round of training, obtain the first label of the first sample song output by it, and update the first training set.

[0148] Step 9: Train the first initial model obtained in the previous training round based on the updated first training set to obtain the first initial model after this training round.

[0149] Step 10: Input the audio signal of the second sample song into the first initial model after this round of training, and update the second training set according to the singing quality features of the second sample song.

[0150] Step 11: Input the audio signal of the first sample song into the first initial model after this round of training, and update the singing quality features of the first sample song in Table 1.

[0151] Repeat steps 7 to 11 until the first and second initial models converge.

[0152] In the above process, the iterative process of the first label can be represented by the following formula:

[0153]

[0154] Where F(X|Y) represents the output obtained by inputting data X into model F trained using data Y, and F(X|Y,Z) represents the output obtained by inputting data X into model F trained using data Y and Z. (n) (B1|B1,B5 (n) ) = B3 (n) D (n) (A1|B1,B5 (n)) = A3 (n) , where n represents the number of iterations, (·) (n) This represents the model or data generated in the nth iteration of the loop. D refers to the first initial model, and C refers to the second initial model.

[0155] The structures of the first and second initial models will be explained in detail below.

[0156] First Initial Model

[0157] Figure 3 This is a structural diagram of a first initial model provided in an embodiment of this disclosure. (See diagram below.) Figure 3 As shown, during training, the first initial model is trained and validated using B1 and B5. During the inference phase, i.e., the update process of B3 and A3, the first initial model uses A1 to infer A3, and can also use B1 to infer B3. The first initial model can be represented by the following formula:

[0158]

[0159] For example, for the first initial model, assuming that the audio signal of the sample song (first sample song or second sample song) input to the first initial model is a dry audio signal with a length of 2:43, the singing quality feature output by the first initial model is [0.65, 0.18, ...], and this singing quality feature is an intermediate result of the first initial model.

[0160] Second initial model

[0161] Figure 4 This is a structural diagram of a second initial model provided in one embodiment of this disclosure. Figure 4 As shown, during training, the second initial model is trained and iterated using A2, A3 (excluding the first round), and A4. During the inference phase, i.e., the update process for B5, the second initial model uses B2 and B3 (excluding the first round) to infer B5. The second initial model can be represented by the following formula:

[0162]

[0163] For example, for the second initial model, assuming the evaluation data of the first sample song input to the second initial model is: pitch score 96, rhythm score 87, number of likes 60, number of comments 24, etc., and the singing quality features are [0.65, 0.18, ...] (assuming that the second initial model has been trained once or more), then the first label output is "pleasant to listen to".

[0164] After obtaining the above-mentioned singing evaluation model, it can be used to annotate the song to be annotated, thereby obtaining the tag for the song. The following detailed description, with reference to specific embodiments, illustrates the method for obtaining the tag for the song using the singing evaluation model. These specific embodiments can be combined with each other; similar or identical concepts or processes may not be elaborated upon in some embodiments.

[0165] In practice, the execution entity of this song annotation method can be a terminal device or a computing device with processing capabilities, such as a server. It should be understood that the computing device executing the song annotation method and the computing device executing the aforementioned model training method can be the same device or different devices.

[0166] Figure 5 This is a flowchart illustrating a song annotation method according to an embodiment of this disclosure. The method of this embodiment can be applied to a computing device, which may be a server or server cluster, or a terminal device. Figure 5 As shown:

[0167] S501. Obtain the audio signal of the song to be labeled.

[0168] In this step, since this embodiment of the disclosure analyzes the audio signal of the song to be labeled based on a singing evaluation model to obtain the label for the song, it is necessary to first obtain the audio signal of the song to be labeled.

[0169] The audio signal of the song to be labeled can be the dry audio signal of the song. Since the dry audio signal does not contain accompaniment and has not undergone post-processing, the influence of irrelevant factors on the processing is effectively avoided.

[0170] Optionally, the audio signal of the song to be labeled may be obtained from the local storage space of the computing device, may be obtained through interaction with the user's terminal device, or may be obtained from other service platforms. For example, the other service platform may be the service platform corresponding to a music website, or the service platform corresponding to a video website, or other music-related service platforms that can provide evaluation services to users. This disclosure does not impose specific restrictions on the method of obtaining the song to be labeled and the audio signal of the song to be labeled.

[0171] S502. Input the audio signal of the song to be labeled into the singing evaluation model to obtain the target label of the song to be labeled.

[0172] In this step, after obtaining the audio signal of the song to be labeled, the audio signal can be input into the singing evaluation model to obtain the target label of the song to be labeled output by the singing evaluation model.

[0173] The target label is used to describe the singing quality of the song to be labeled, and the singing evaluation model is trained using the model training method shown in any of the above embodiments.

[0174] Based on the data contained in the song to be labeled, different methods can be used to determine the target tag for the song.

[0175] Method 1: When only the audio signal of the song to be labeled is obtained, the target label of the song to be labeled can be determined by the singing evaluation model.

[0176] In this method, the target tags for the song to be tagged can be determined using the following formula:

[0177] E5=D * (E1|B1,B5 * )

[0178] Where E5 is the target label, D * For the singing evaluation model, E1 is the audio signal of the song to be labeled.

[0179] Method 2: When acquiring the audio signal and evaluation data of the song to be labeled, in order to further improve the accuracy of the target label, the target label of the song to be labeled can be determined by combining the singing evaluation model and the labeling model. The specific implementation process can be achieved through the following steps (3) and (4):

[0180] Step (3): Input the audio signal of the song to be labeled into the singing evaluation model to obtain the singing quality characteristics of the song to be labeled.

[0181] Step (4): Input the singing quality features and evaluation data of the song to be labeled into the labeling model to obtain the target label of the song to be labeled. The labeling model is trained using the method shown in any of the above embodiments.

[0182] In this method, the target tags for the song to be tagged can be determined using the following formula:

[0183] E5 = C * (E2,D * (E1|B1,B5 * )|A2,D * (A1|B1,B5 * ),A4)

[0184] Since the evaluation data of the songs to be labeled reflects the auditory experience of users who listen to them, it indirectly provides feedback on the performance quality of the songs. In other words, the evaluation data of the songs to be labeled can play an auxiliary role in the labeling process. This method utilizes rating data that can assist in labeling, fully explores the value of the data, and further improves the accuracy of labeling.

[0185] The song annotation method provided in this disclosure involves a computing device acquiring the audio of a song to be annotated, inputting the audio signal of the song to be annotated into a singing evaluation model, and obtaining the target label signal of the song to be annotated. The target label is used to describe the singing quality of the song to be annotated, and the singing evaluation model is trained using the model training method shown in any of the above embodiments. This technical solution uses a singing evaluation model to annotate the song to be annotated, replacing the manual annotation process in the prior art, effectively improving the accuracy and efficiency of annotation. This allows for the subsequent construction of a large-scale, high-quality training set based on the annotated songs for use in the training process of existing models for obtaining song scores.

[0186] Exemplary media

[0187] After introducing the methods of exemplary embodiments of this disclosure, the following references are made. Figure 6 The storage medium of the exemplary embodiments of this disclosure will be described.

[0188] Figure 6 This is a structural diagram of a storage medium provided in an embodiment of the present disclosure, with reference to... Figure 6 As shown, the storage medium 60 stores a program product for implementing the above-described methods according to embodiments of the present disclosure. This program product may be a portable compact disc read-only memory (CD-ROM) and includes computer-executable instructions for causing a computing device to execute the model training method and song annotation method provided in this disclosure. However, the program product of this disclosure is not limited thereto.

[0189] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0190] A readable signal medium may include data signals propagated in baseband or as part of a carrier wave, carrying computer-executed instructions. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium.

[0191] Computer-executable instructions for performing the operations disclosed herein can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The computer-executable instructions can be executed entirely on the user's computing device, partially on the user's computing device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN).

[0192] Exemplary device

[0193] Having introduced the medium of exemplary embodiments of this disclosure, the following references are made to... Figure 6 The model training apparatus of exemplary embodiments of this disclosure will be described to implement the model training method in any of the above method embodiments. Meanwhile, reference will be made to... Figure 7 The song annotation apparatus according to exemplary embodiments of this disclosure is described below, which is used to implement the song annotation method in any of the above method embodiments. Its implementation principle and technical effects are similar, and will not be repeated here.

[0194] Figure 7 This is a structural diagram of a model training apparatus provided in one embodiment of the present disclosure. Figure 7 As shown, the model training device 700 includes:

[0195] The acquisition module 701 is used to acquire the first training set, which includes multiple first sample songs, the audio signal of each first sample song, and a first tag.

[0196] Training module 702 is used to train the first initial model multiple times using the first training set to obtain a singing evaluation model. In any training round, the first label is obtained by inputting the evaluation data and singing quality features of the first sample song into the second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled.

[0197] In one possible design of this embodiment, the training module 702 is further configured to:

[0198] The second initial model is trained multiple times using a second training set to obtain a labeled model. In each training round, the second training set includes multiple second sample songs, evaluation data for each second sample song, second labels, and performance quality features. The second labels are manually annotated and represent the performance quality of the second sample songs. The labeling model is used to determine the target label for the song to be labeled based on its performance quality features and evaluation data.

[0199] In one possible design of this embodiment, the training module 702 is specifically used for:

[0200] In the first round of training, the second initial model is trained using multiple second sample songs, the evaluation data of each second sample song, and the second label, to obtain the second initial model after the first round of training.

[0201] In any subsequent training round, the second initial model after the previous training round is trained using multiple second sample songs and the evaluation data, second labels, and singing quality features of each second sample song to obtain the labeled model.

[0202] In one possible design of this disclosure embodiment, after training the second initial model using multiple second sample songs, evaluation data of each second sample song, and second labels in the first round of training to obtain the second initial model after the first round of training, the model training device 700 further includes:

[0203] The input module is used to input multiple first sample songs and the evaluation data of each first sample song into the second initial model after the first round of training in the first round of training, so as to obtain the first label of the first sample song.

[0204] The determination module is used to determine, in the first round of training, multiple first sample songs and the first label and audio signal of each first sample song as the first training set of the first initial model.

[0205] In one possible design of this disclosure embodiment, the input module is further configured to:

[0206] In any round of training, multiple first sample songs, the evaluation data of each first sample song, and the singing quality features are input into the second initial model after this round of training, and the first label of the first sample song is updated.

[0207] In one possible design of this disclosure embodiment, the input module is further configured to:

[0208] In any training round, multiple first sample songs and the audio signal of each first sample song are input into the first initial model after the previous training round to update the singing quality features of the first sample songs.

[0209] In any training round, multiple second sample songs and the audio signal of each second sample song are input into the first initial model after the previous training round to update the singing quality features of the second sample songs.

[0210] The model training apparatus provided in this disclosure can be used to execute the model training method in any of the above embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0211] Figure 8 This is a structural diagram of a song annotation device provided in one embodiment of the present disclosure. Figure 8 As shown, the song annotation device 800 includes:

[0212] The acquisition module 801 is used to acquire the audio signal of the song to be labeled.

[0213] The input module 802 is used to input the audio signal of the song to be labeled into the singing evaluation model to obtain the target label of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. The singing evaluation model is trained using the model training method in any of the above embodiments.

[0214] In one possible design of this embodiment, the input module 802 is specifically used for:

[0215] The audio signal of the song to be labeled is input into the singing evaluation model to obtain the singing quality features of the song to be labeled.

[0216] Input the singing quality features and evaluation data of the song to be labeled into the labeling model to obtain the target label of the song to be labeled. The labeling model is trained using the model training method in any of the above embodiments.

[0217] The song annotation device provided in this disclosure can be used to execute the song annotation method in any of the above embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0218] Exemplary computing device

[0219] Having described the methods, media, and apparatus of exemplary embodiments of this disclosure, the following references... Figure 9 A computing device according to an exemplary embodiment of the present disclosure will be described.

[0220] Figure 9The computing device 90 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0221] like Figure 9 As shown, the computing device 90 is presented in the form of a general-purpose computing device. The components of the computing device 90 may include, but are not limited to: at least one processing unit 901, at least one storage unit 902, and a bus 903 connecting different system components (including the processing unit 901 and the storage unit 902). The at least one storage unit 902 stores computer-executable instructions; the at least one processing unit 901 includes a processor that executes the computer-executable instructions to implement the methods described above.

[0222] The 903 bus includes a data bus, a control bus, and an address bus.

[0223] Storage unit 902 may include readable media in the form of volatile memory, such as random access memory (RAM) 9021 and / or cache memory 9022, and may further include readable media in the form of non-volatile memory, such as read-only memory (ROM) 9023.

[0224] Storage unit 902 may also include a program / utility 9025 having a set (at least one) program model 9024, such program model 9024 including but not limited to: operating system, one or more application programs, other program models and program data, each of these examples or some combination of these examples may include an implementation of a network environment.

[0225] The computing device 90 can also communicate with one or more external devices 904 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 905. Furthermore, the computing device 90 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 906. Figure 9 As shown, network adapter 906 communicates with other models of computing device 90 via bus 903. It should be understood that, although not shown in the figure, other hardware and / or software models can be used in conjunction with computing device 90, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0226] It should be noted that although several units / models or sub-units / models of the model training device and song annotation device are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / models described above can be embodied in one unit / model. Conversely, the features and functions of one unit / model described above can be further divided and embodied by multiple units / models.

[0227] Furthermore, although the operations of the methods disclosed herein are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0228] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A model training method, comprising: Obtain a first training set, which includes multiple first sample songs, the audio signal of each first sample song, and a first tag; The first initial model is trained multiple times using the first training set to obtain a singing evaluation model. In any training round, the first label is obtained by inputting the evaluation data and singing quality features of the first sample song into the second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. While the first initial model is being trained, the second initial model is also being trained. The training of the second initial model includes the following steps: training the second initial model multiple times using a second training set to obtain a labeled model; wherein, in any round of training, the second training set includes multiple second sample songs, evaluation data for each second sample song, a second label, and performance quality features, wherein the second label is manually labeled and used to characterize the performance quality of the second sample song, and the labeled model is used to determine the target label of the song to be labeled based on the performance quality features and evaluation data of the song to be labeled.

2. The method according to claim 1, wherein the step of performing multiple rounds of model training on the second initial model using the second training set to obtain the labeled model includes: In the first round of training, the second initial model is trained using multiple second sample songs, the evaluation data of each second sample song, and the second label to obtain the second initial model after the first round of training. In any subsequent training round, the second initial model after the previous training round is trained using multiple second sample songs and the evaluation data, second labels, and singing quality features of each second sample song to obtain the labeled model.

3. The method according to claim 2, wherein after training the second initial model using multiple second sample songs, evaluation data of each second sample song, and second labels in the first round of training to obtain the second initial model after the first round of training, the method further includes: In the first round of training, multiple first sample songs and the evaluation data of each first sample song are input into the second initial model after the first round of training to obtain the first label of the first sample song; In the first round of training, multiple first sample songs, along with the first label and audio signal of each first sample song, are determined as the first training set of the first initial model.

4. The method according to claim 2 or 3, further comprising: In any round of training, multiple first sample songs, the evaluation data of each first sample song, and the singing quality features are input into the second initial model after this round of training, and the first label of the first sample song is updated.

5. The method according to any one of claims 1 to 3, wherein the method further comprises: In any round of training, multiple first sample songs and the audio signal of each first sample song are input into the first initial model after the previous round of training to update the singing quality features of the first sample songs. In any training round, multiple second sample songs and the audio signal of each second sample song are input into the first initial model after the previous training round to update the singing quality features of the second sample songs.

6. A song annotation method, comprising: Obtain the audio signal of the song to be labeled; The audio signal of the song to be labeled is input into the singing evaluation model to obtain the target label of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. The singing evaluation model is trained using the method described in any one of claims 1-5.

7. The method according to claim 6, wherein inputting the audio signal of the song to be labeled into a singing evaluation model to obtain the target label of the song to be labeled includes: The audio signal of the song to be labeled is input into the singing evaluation model to obtain the singing quality characteristics of the song to be labeled. The singing quality features and evaluation data of the song to be labeled are input into the labeling model to obtain the target label of the song to be labeled. The labeling model is trained using the method described in any one of claims 1-6.

8. A model training device, comprising: The acquisition module is used to acquire a first training set, which includes multiple first sample songs, the audio signal of each first sample song, and a first tag. The training module is used to train the first initial model multiple times using the first training set to obtain a singing evaluation model. In any training round, the first label is obtained by inputting the evaluation data and singing quality features of the first sample song into the second initial model after this round of training. The singing quality features are intermediate results obtained by inputting the audio signal of the first sample song into the first initial model after the previous round of training. The singing evaluation model is used to determine the target label of the song to be labeled based on the audio signal of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. While training the first initial model, a second initial model is also trained. The training module is further configured to: perform multiple rounds of model training on the second initial model using a second training set to obtain a labeled model. In any round of training, the second training set includes multiple second sample songs, evaluation data for each second sample song, a second label, and performance quality features. The second label is manually labeled and is used to characterize the performance quality of the second sample song. The labeled model is used to determine the target label of the song to be labeled based on the performance quality features and evaluation data of the song to be labeled.

9. The apparatus according to claim 8, wherein the training module is specifically used for: In the first round of training, the second initial model is trained using multiple second sample songs, the evaluation data of each second sample song, and the second label to obtain the second initial model after the first round of training. In any subsequent training round, the second initial model after the previous training round is trained using multiple second sample songs and the evaluation data, second labels, and singing quality features of each second sample song to obtain the labeled model.

10. The apparatus according to claim 9, wherein after training the second initial model using multiple second sample songs, evaluation data of each second sample song, and second labels in the first round of training to obtain the second initial model after the first round of training, the apparatus further comprises: The input module is used to input multiple first sample songs and the evaluation data of each first sample song into the second initial model after the first round of training in the first round of training, so as to obtain the first label of the first sample song; The determination module is used to determine, in the first round of training, multiple first sample songs and the first label and audio signal of each first sample song as the first training set of the first initial model.

11. The apparatus according to claim 10, wherein the input module is further configured to: In any round of training, multiple first sample songs, the evaluation data of each first sample song, and the singing quality features are input into the second initial model after this round of training, and the first label of the first sample song is updated.

12. The apparatus according to claim 10, wherein the input module is further configured to: In any round of training, multiple first sample songs and the audio signal of each first sample song are input into the first initial model after the previous round of training to update the singing quality features of the first sample songs. In any training round, multiple second sample songs and the audio signal of each second sample song are input into the first initial model after the previous training round to update the singing quality features of the second sample songs.

13. A song annotation device, comprising: The acquisition module is used to acquire the audio signal of the song to be labeled; The input module is used to input the audio signal of the song to be labeled into the singing evaluation model to obtain the target label of the song to be labeled. The target label is used to describe the singing quality of the song to be labeled. The singing evaluation model is trained using the method described in any one of claims 1-5.

14. The apparatus according to claim 13, wherein the input module is specifically used for: The audio signal of the song to be labeled is input into the singing evaluation model to obtain the singing quality characteristics of the song to be labeled. The singing quality features and evaluation data of the song to be labeled are input into the labeling model to obtain the target label of the song to be labeled. The labeling model is trained using the method described in any one of claims 1-5.

15. A storage medium storing computer program instructions that, when executed, implement the method as described in any one of claims 1 to 7.

16. A computing device, comprising: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

Images