Song evaluation information acquisition method, computer device, and storage medium

Abstract

The application relates to a singing evaluation information acquisition method, computer equipment and a storage medium, relates to the singing evaluation technical field, and can improve the accuracy of singing evaluation information. The method comprises the following steps: acquiring singing characteristics corresponding to a to-be-evaluated singing; comparing the singing characteristics with pre-acquired positive sample singing characteristics corresponding to a positive sample singing and negative sample singing characteristics corresponding to a negative sample singing, obtaining a first similarity between the singing characteristics and the positive sample singing characteristics, and a second similarity between the singing characteristics and the negative sample singing characteristics; wherein the singing quality of the positive sample singing is higher than that of the negative sample singing; and according to the first similarity and the second similarity, determining singing evaluation information of the to-be-evaluated singing.

Description

Technical Field

[0001] This application relates to the field of singing evaluation technology, and in particular to a method for acquiring singing evaluation information, a computer device, and a storage medium. Background Technology

[0002] With the development of computer technology, singing evaluation models can be used to perform multi-dimensional analysis of the singing of the evaluated object, and obtain singing evaluation results. One type of singing evaluation method is non-reference evaluation, which refers to using deep learning methods to allow the model to learn features for evaluating singing quality from a large amount of labeled training data.

[0003] In related technologies, when training a singing evaluation model for the absence of a reference evaluation, the singing quality of the singing samples is manually labeled. However, during labeling, staff can generally only classify the singing samples according to different quality levels. With a limited number of label levels, the singing evaluation model often struggles to obtain detailed and reliable singing evaluation information. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, computer device, and storage medium for acquiring singing evaluation information that can improve the accuracy of singing evaluation information, in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a method for obtaining singing evaluation information. The method includes:

[0006] Obtain the vocal characteristics corresponding to the singing voice to be evaluated;

[0007] The singing features are compared with the pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features; wherein, the singing quality of the positive sample singing is higher than that of the negative sample singing.

[0008] Based on the first similarity and the second similarity, the singing evaluation information of the singing voice to be evaluated is determined.

[0009] In one embodiment, the positive sample singing features and the negative sample singing features are obtained through the following steps:

[0010] Obtain a set of positive sample singing voices and a set of negative sample singing voices; the singing quality of each sample singing voice in the set of positive sample singing voices is higher than the singing quality of each sample singing voice in the set of negative sample singing voices;

[0011] Based on the singing characteristics of each sample singing voice in the positive sample singing voice set, the feature center of each sample singing voice in the positive sample singing voice set is obtained as the positive sample singing voice feature; and based on the singing characteristics of each sample singing voice in the negative sample singing voice set, the feature center of each sample singing voice in the negative sample singing voice set is obtained as the negative sample singing voice feature.

[0012] In one embodiment, determining the singing evaluation information of the singing voice to be evaluated based on the first similarity and the second similarity includes:

[0013] If the first similarity is greater than the second similarity, then an adjustment value for the evaluation score is determined based on the first similarity, and the adjustment value for the evaluation score is positively correlated with the first similarity.

[0014] If the first similarity is less than the second similarity, then an adjustment value for the evaluation score is determined based on the second similarity, wherein the adjustment value for the evaluation score is negatively correlated with the second similarity;

[0015] Based on the baseline evaluation score and the adjusted evaluation score, a singing evaluation score is obtained, and this singing evaluation score is used as singing evaluation information.

[0016] In one embodiment, the vocal features corresponding to the vocal performance to be evaluated include: vocal features corresponding to each of the multiple vocal segments of the vocal performance to be evaluated; each vocal segment corresponds to a baseline evaluation score and an adjustment value for the evaluation score;

[0017] The process of obtaining the singing evaluation score based on the baseline evaluation score and the evaluation score adjustment value includes:

[0018] The evaluation score for each singing segment is determined based on the baseline evaluation score and the evaluation score adjustment value for each singing segment.

[0019] The singing evaluation score is determined based on the individual singing segment evaluation scores of the multiple singing segments.

[0020] In one embodiment, the vocal features of the singing voice to be evaluated are obtained through a vocal feature extraction model; the vocal feature extraction model is trained through the following steps:

[0021] Obtain sample singing pairs; each sample singing pair includes two singing samples with different singing qualities;

[0022] Multiple vocal segments from each of the sample vocal pairs are input into the vocal feature extraction model to be trained to obtain vocal segment features corresponding to each vocal segment.

[0023] The vocal features of the sample singing are determined based on the features of the multiple vocal segments.

[0024] Based on the similarity between the features of each singing segment and the singing features of the corresponding sample singing, a first feature similarity is determined, and a second feature similarity is obtained between the singing features of each sample singing.

[0025] Based on the first feature similarity and the second feature similarity, the model parameters of the singing feature extraction model are adjusted to obtain a trained singing feature extraction model.

[0026] In one embodiment, determining the first feature similarity based on the similarity between the features of each vocal segment and the vocal features of the corresponding sample vocal segment includes:

[0027] Determine the similarity between the features of each singing segment and the singing features of the corresponding sample singing;

[0028] The first feature similarity is obtained by summing the similarities described above.

[0029] In one embodiment, determining the vocal features of the sample vocalization based on the features of the plurality of vocal segments includes:

[0030] For each sample singing voice, the mean value of the features of multiple singing voice segments corresponding to the sample singing voice is obtained;

[0031] The vocal characteristics of the sample singing can be obtained based on the mean value.

[0032] In one embodiment, adjusting the model parameters of the singing evaluation model based on the first feature similarity and the second feature similarity includes:

[0033] A first loss value is determined based on the first feature similarity; the first loss value is negatively correlated with the first feature similarity.

[0034] A second loss value is determined based on the second feature similarity; the second loss value is positively correlated with the second feature similarity.

[0035] The model loss value is determined based on the sum of the first loss value and the second loss value, and the model parameters of the singing feature extraction model are adjusted based on the model loss value.

[0036] Secondly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0037] Obtain the vocal characteristics corresponding to the singing voice to be evaluated;

[0038] The singing features are compared with the pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features; wherein, the singing quality of the positive sample singing is higher than that of the negative sample singing.

[0039] Based on the first similarity and the second similarity, the singing evaluation information of the singing voice to be evaluated is determined.

[0040] Thirdly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0041] Obtain the vocal characteristics corresponding to the singing voice to be evaluated;

[0042] The singing features are compared with the pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features; wherein, the singing quality of the positive sample singing is higher than that of the negative sample singing.

[0043] Based on the first similarity and the second similarity, the singing evaluation information of the singing voice to be evaluated is determined.

[0044] Fourthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0045] Obtain the vocal characteristics corresponding to the singing voice to be evaluated;

[0046] The singing features are compared with the pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features; wherein, the singing quality of the positive sample singing is higher than that of the negative sample singing.

[0047] Based on the first similarity and the second similarity, the singing evaluation information of the singing voice to be evaluated is determined.

[0048] The aforementioned method, computer equipment, and storage medium for acquiring singing evaluation information can obtain the singing features corresponding to the singing to be evaluated. These features are then compared with pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features. The singing quality of the positive sample singing is higher than that of the negative sample singing. Furthermore, the singing evaluation information of the singing to be evaluated can be determined based on the first and second similarities. In this embodiment, there is no need for strict hierarchical differentiation and scoring of the sample singing; it only requires simple differentiation of singing quality and acquisition of positive and negative sample singing features. The singing evaluation information can then be determined based on the first similarity between the singing features and the positive sample features, and the second similarity between the singing features and the negative sample features. By comparing the singing features with the positive and negative sample singing features used as reference benchmarks, subtle differences between different singing can be objectively reflected, effectively improving the accuracy of the singing evaluation information. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a method for obtaining singing evaluation information in one embodiment;

[0050] Figure 2 This is a flowchart illustrating the steps of training a singing voice feature extraction model in one embodiment;

[0051] Figure 3 This is a schematic diagram of the feature distribution of a singing segment in one embodiment;

[0052] Figure 4 This is a flowchart illustrating another step in training a singing voice feature extraction model in one embodiment;

[0053] Figure 5 This is an internal structural diagram of a computer device in one embodiment;

[0054] Figure 6 This is an internal structural diagram of another computer device in one embodiment. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] To facilitate understanding of the embodiments of this application, the relevant technologies for singing evaluation will be introduced below.

[0057] Vocal evaluation refers to the process of analyzing a subject's singing from multiple dimensions to obtain a comprehensive evaluation result. For example, a vocal evaluation model can be used to analyze the subject's vocal audio in dimensions such as timbre, pitch, rhythm, breath control, vibrato, glissando, and emotion to obtain a comprehensive score. Evaluating vocal performance through a vocal evaluation model not only provides objective results but also allows users to understand the strengths and weaknesses of their singing, while also increasing the enjoyment of the singing process.

[0058] In practical applications, singing evaluation models can be divided into referenced evaluation and non-referenced evaluation. Referenced evaluation involves processing and analyzing the audio signal based on the vocal characteristics of the subject being evaluated, comparing the processed signal with the vocal characteristics of a template song, and then determining the evaluation result. Non-referenced evaluation, on the other hand, uses deep learning methods to allow the model to learn features for evaluating singing quality from a large amount of labeled training data. For example, it learns the commonalities and differences between good and bad singing. Compared to referenced evaluation, non-referenced evaluation can provide a more comprehensive and multi-dimensional assessment of singing.

[0059] In related technologies, for songs without reference evaluation, the quality of the singing samples is manually labeled, and then a singing evaluation model is trained using these labeled singing samples. In some exemplary training methods, a regression model can be trained using mean squared error to obtain the singing evaluation model. That is, the signal features of the singing samples are input into the singing evaluation model, and the model outputs a probability value between 0 and 1 for each singing sample based on the signal features. The higher the probability value, the higher the score of the singing sample. For this, the labels of the singing samples need to be quantized. For example, if the labels of the singing samples include four types: "good," "not bad," "average," and "not good," then when quantizing the labels, these four types of labels correspond to the following values ​​respectively: 1, 0.75, 0.5, and 0.25. Of course, other methods can also be used to quantize these four types of labels to other values ​​in the 0-1 range. Then, during training, the model's output probability value and the preset labels of the singing samples are combined, and a loss function is used to determine the model's loss value for iterative optimization. After training is complete, the probability values ​​output by the singing evaluation model can be transformed according to a preset method (such as multiplying by a preset factor), and the transformation result can be used as the final singing evaluation result.

[0060] In other ways, the model can be trained according to the concept of classification. That is, the singing evaluation model can classify singing samples according to the input information. Taking the four label types mentioned above as an example, when training using the classification method, the singing evaluation model can be trained as a four-class classification model. The final singing evaluation result is transformed by the classification probability value. For example, the type with the highest probability value can be determined and the score corresponding to that type can be used as the singing evaluation result. Alternatively, the probability values ​​of each type and the score corresponding to each type can be weighted and summed to obtain the singing evaluation result based on the weighted summation result.

[0061] In the aforementioned no-reference evaluation methods, both regression and classification methods have relatively strict requirements for the labeling of singing samples. However, when labeling singing samples, singing evaluation and labeling involve a fusion of subjective and objective factors, making it difficult to obtain a large number of objectively labeled samples. Furthermore, staff can often only roughly classify and label singing quality at different levels, making it difficult to provide highly granular labels for singing samples. For example, manual labeling often cannot produce such a detailed score as 76.9. This is understandable, as the types of labels for different levels of singing samples are limited, and scores other than labeled scores will not appear in the training set. Singing evaluation information between levels must be obtained by the model through fitting, which obviously makes it difficult to guarantee the accuracy of scores between levels. Moreover, the further the obtained singing evaluation information is from the level provided during labeling, the more difficult it is to guarantee the reliability of the singing evaluation information. For example, for the four labels mentioned above, if the probability value output by the model is 0.625, it falls between the quantization values ​​of 0.5 and 0.75, and the result has a certain degree of ambiguity. Furthermore, even when evaluating singing through classification, subtle differences between samples at the same level are often overlooked, making it difficult to accurately distinguish the singing quality among multiple samples at the same level.

[0062] It is evident that singing evaluation models trained using relevant technologies often fail to yield detailed and reliable evaluation results. Therefore, this application provides a method for acquiring singing evaluation information, a computer device, and a storage medium to at least improve the reliability of singing evaluation information.

[0063] In one embodiment, such as Figure 1 As shown, a method for obtaining singing evaluation information is provided. This embodiment illustrates the application of this method to a server. It is understood that this method can also be applied to a terminal, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0064] S101, Obtain the vocal features corresponding to the vocal performance to be evaluated.

[0065] In practical applications, the singing voice to be evaluated can be obtained. In some embodiments, the singing voice to be evaluated can be audio recordings uploaded by users. For example, it can be audio recordings of a user singing online solo, or music works containing vocals published by the user. Alternatively, it can be audio recordings of a singing competition, for example, for a singing competition involving at least two users, the audio recordings of each user's singing voice during the competition can be used as the singing voice to be evaluated.

[0066] In other embodiments, the singing to be evaluated can also be audio recordings stored in a music library. For example, when recommending songs to a user, each candidate song can be used as a song to be evaluated. Then, after obtaining the singing evaluation information for each candidate song, the candidate songs and their evaluation information can be associated and sent to the user's terminal. The user can then select the song to play based on the singing evaluation information. Similarly, when recommending or filtering songs, different versions of the same song can be compared to prioritize recommending or filtering out versions with higher singing quality.

[0067] In this step, after acquiring the singing voice to be evaluated, feature extraction can be performed on the singing voice to obtain its vocal features. In one example, the vocal features can be represented in the form of embeddings. In some optional embodiments, the singing voice to be evaluated can be preprocessed. For example, after acquiring the audio containing the singing voice to be evaluated, the audio can be denoised and de-accompanied, removing blank segments from the audio and retaining only the vocal segments that include human voices. Then, feature extraction can be performed on the vocal segments to obtain the corresponding vocal features.

[0068] S102, compare the singing features with the pre-acquired positive sample singing features and negative sample singing features to obtain the first similarity between the singing features and the positive sample singing features, and the second similarity between the singing features and the negative sample singing features; wherein, the singing quality of the positive sample singing is higher than that of the negative sample singing.

[0069] In practical applications, positive and negative sample singing features can be pre-obtained. Positive sample singing features are obtained by extracting features from multiple positive samples, and negative sample singing features are obtained by extracting features from multiple negative samples. By extracting features from both positive and negative samples, we can obtain the common features of high-quality singing (positive sample singing features) and the common features of low-quality singing (negative sample singing features).

[0070] Furthermore, the singing features corresponding to the singing to be evaluated can be compared with the pre-acquired positive sample singing features and negative sample singing features respectively. The first similarity is obtained based on the difference or similarity between the singing features and the positive sample singing features, and the second similarity is obtained based on the difference or similarity between the singing features and the negative sample singing features.

[0071] S103, determine the singing evaluation information of the singing voice to be evaluated based on the first similarity and the second similarity.

[0072] After obtaining the first and second similarity scores, the first similarity score can be used to quantitatively and meticulously determine the degree of similarity between the singing to be evaluated and the singing of positive samples. In other words, it can be used to quantitatively determine whether the singing to be evaluated possesses the common characteristics of high-quality singing. Similarly, the second similarity score can be used to quantitatively and meticulously determine the degree of similarity between the singing to be evaluated and the singing of negative samples. This means it can be used to quantitatively determine whether the singing to be evaluated possesses the common characteristics of low-quality singing. It can be understood that the first and second similarity scores determined during the feature comparison process are continuous intervals, rather than a number of discrete specific values. Therefore, the first and second similarity scores can accurately and meticulously reflect the quality of the singing to be evaluated.

[0073] Based on this, the singing evaluation information of the singing to be evaluated can be determined according to the first similarity and the second similarity. Specifically, for example, a preset score mapping relationship can be obtained. This score mapping relationship can characterize the mapping relationship between the singing evaluation score and the first similarity and / or the second similarity. In this score mapping relationship, there can be a one-to-one corresponding singing evaluation score for each first similarity and each second similarity. Then, based on the currently determined first similarity, second similarity, and preset score mapping relationship, the singing evaluation score of the singing to be evaluated can be determined, and this score can be used as the singing evaluation information.

[0074] In the above-described method for obtaining singing evaluation information, the server can acquire the singing features corresponding to the singing to be evaluated, and compare these features with pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features. The singing quality of the positive sample singing is higher than that of the negative sample singing. Therefore, the singing evaluation information of the singing to be evaluated can be determined based on the first and second similarities. In this embodiment, there is no need for strict hierarchical differentiation and scoring of the sample singing; it is only necessary to simply differentiate between high and low singing quality and acquire positive and negative sample singing features. Then, the singing evaluation information can be determined based on the first similarity between the singing features and the positive sample singing features, and the second similarity between the singing features and the negative sample singing features. By comparing the singing features with the positive and negative sample singing features used as reference benchmarks, the subtle differences between different singing can be objectively reflected, effectively improving the accuracy of the singing evaluation information.

[0075] In one embodiment, the vocal features of the singing voice to be evaluated are obtained through a vocal feature extraction model, such as... Figure 2 As shown, this singing voice feature extraction model can be trained through the following steps:

[0076] S201, Obtain sample singing pairs; each sample singing pair includes two singing samples with different singing qualities.

[0077] In this step, multiple sample vocal pairs can be obtained. Each sample vocal pair includes two sample vocals with different vocal qualities. The two sample vocals in the same sample vocal pair can be vocals collected from singing the same song.

[0078] In some optional embodiments, the songs sung by users when participating in a singing competition can be used as sample vocal pairs. In the case of a singing competition that is a two-person competition, the songs sung by two users participating in the same competition can constitute a sample vocal pair. Alternatively, multiple song audios of different versions of the same song can be obtained from the music library. These multiple different versions can be sung by different users, or they can be sung by the same user but with different performance styles. Then, two song audios can be randomly selected from the multiple song audios as a sample vocal pair.

[0079] In some optional embodiments, for each sample singing pair, staff can listen to the two sample singing voices and compare their singing quality. For example, they can comprehensively evaluate and compare factors such as the singer's breath, timbre, and emotion, and then add comparison result labels. The comparison result labels can be used to distinguish the quality of the two sample singing voices in the same sample singing pair. For example, one sample singing voice can be marked as "winning" and the other sample singing voice as "losing".

[0080] S202, input multiple singing segments of each sample singing voice pair into the singing voice feature extraction model to be trained, and the singing voice feature extraction model obtains the singing segment features corresponding to each singing segment.

[0081] After obtaining sample vocal pairs, they can be input into the vocal feature extraction model to be trained, on a pair-by-pair basis, for feature extraction. During input, the two sample vocal pairs can be segmented into phrases to obtain multiple vocal segments. These multiple vocal segments can then be input into the vocal feature extraction model to be trained, whereby the model extracts the features corresponding to each vocal segment. In some optional embodiments, the spectral features of the vocal segments can be obtained. After inputting the spectral features into the vocal feature extraction model, the model determines the vocal features of the vocal segments based on the spectral features.

[0082] S203, determine the singing characteristics of the sample singing based on the characteristics of multiple singing segments.

[0083] Specifically, since the vocal segment features corresponding to each vocal segment in each sample vocal performance have been obtained, in this step, for each sample vocal performance, the vocal features of the sample vocal performance can be determined by combining the vocal segment features corresponding to each vocal segment in the sample vocal performance.

[0084] In one embodiment, step S203 may include the following steps: for each sample singing voice, obtain the mean value of multiple singing voice segment features corresponding to the sample singing voice; obtain the singing voice features of the sample singing voice based on the mean value.

[0085] In practical applications, for each sample singing voice, multiple singing voice segments corresponding to the sample singing voice can be determined, and the mean value of the singing voice segment features corresponding to each of the multiple singing voice segments can be obtained. This mean value can also be called the feature center of the sample singing voice. Then, the mean value of the features of multiple singing voice segments can be used as the singing voice features of the sample singing voice.

[0086] For example, Figure 3 The diagram illustrates the characteristics of a vocal segment. For the first and second sample vocal samples, such as... Figure 3 As shown, center A is the feature center of the first sample singing voice, and center B is the feature center of the second sample singing voice. After obtaining the features of each singing voice segment, the mean of the features of multiple singing voice segments of the first sample singing voice can be determined as the feature center of the first sample singing voice; for the second sample singing voice, the mean of the features of multiple singing voice segments of the second sample singing voice can also be determined as the feature center of the second sample singing voice.

[0087] Since the sample singing voice is composed of multiple singing voice segments, in this embodiment, the singing voice features of the sample singing voice can be quickly obtained by obtaining the average value of the features of multiple singing voice segments corresponding to the sample singing voice.

[0088] S204, determine the first feature similarity based on the similarity between the features of each singing segment and the singing features of the corresponding sample singing, and obtain the second feature similarity between the singing features of each sample singing.

[0089] Specifically, for each vocal segment of a sample song, since it originates from that sample song (or the sample song contains that vocal segment), the overall sample song and its segments will share a similar or identical performance style. Consequently, the features of the vocal segments will be quite similar to the vocal features of the sample song. However, for different sample songs, since they are sung by different users, there will be differences in timbre, breath control, emotion, and one or more performance techniques. Therefore, the vocal features acquired by the model for different sample songs will vary significantly.

[0090] To address this, on one hand, after obtaining the vocal features corresponding to the sample singing voices and the vocal segment features corresponding to each vocal segment, the vocal segment features of each vocal segment can be compared with the vocal features of the corresponding sample singing voice to obtain the similarity between the vocal segment features and the vocal features of the sample singing voice. Here, the corresponding sample singing voice refers to the sample singing voice containing the given vocal segment. Therefore, the first feature similarity can be determined based on the similarity between the vocal segment features and the vocal features of the corresponding sample singing voice. On the other hand, the second feature similarity can be obtained by acquiring the similarity between the vocal features of two sample singing voices.

[0091] In some optional embodiments, when obtaining the similarity between singing segment features and singing features, or when obtaining the similarity between two singing features, the similarity between features can be obtained based on the cosine distance or Euclidean distance between the features. Figure 3 For example, the second feature similarity can be obtained based on the cosine distance or Euclidean distance between feature center A and feature center B, where the second segment feature similarity is negatively correlated with this distance.

[0092] S205. Based on the first feature similarity and the second feature similarity, adjust the model parameters of the singing feature extraction model to obtain the trained singing feature extraction model.

[0093] After obtaining the first feature similarity and the second feature similarity, the current model training effect can be determined by the first feature similarity and the second feature similarity. Based on the first feature similarity and the second feature similarity, the model loss value is determined. The model parameters are adjusted based on the model loss value. Then, the process returns to step S201 and repeats the above process until the training termination condition is met (such as the loss function converges), and the trained singing feature extraction model is obtained.

[0094] In this embodiment, there is no need to strictly distinguish and score the sample singing voices. It is only necessary to provide sample singing voice pairs containing sample singing voices of different qualities. By evaluating the first feature similarity between the singing voice segment features and the corresponding singing voice features, as well as the second feature between the singing voice features of different sample singing voices, it is possible to determine whether the feature extraction method of the singing voice feature extraction model is appropriate, whether the singing voice features extracted by it can distinguish sample singing voices of different qualities, and adjust the model parameters accordingly so that the model can learn the differences between sample singing voice pairs, thereby extracting singing voice features for distinguishing singing voice quality more accurately.

[0095] In one embodiment, step S204, determining the first feature similarity based on the similarity between the features of each singing segment and the singing features of the corresponding sample singing, may include the following steps: determining the similarity between the features of each singing segment and the singing features of the corresponding sample singing; obtaining the first feature similarity based on the sum of the similarities.

[0096] After obtaining the vocal features of each sample song, the similarity between each vocal segment feature and the corresponding sample song feature is determined. Then, the first segment feature similarity can be obtained by summing the similarities of each segment feature. Specifically, for example, after obtaining the similarity between each vocal segment feature of the first sample song and the vocal features of the first sample song, and the similarity between each vocal segment feature of the second sample song and the vocal features of the second sample song, the various similarities can be summed to obtain the first feature similarity.

[0097] In this embodiment, by obtaining the similarity between the features of each singing segment and the corresponding singing features and summing them, the similarity between the singing segment features and the singing features can be determined quickly and accurately, so that the singing segment features extracted by the model are as close as possible to the feature center of the singing, and the feature differences between different singing voices are increased.

[0098] In one embodiment, step S204, adjusting the model parameters of the singing feature extraction model based on the feature similarity of the first segment and the feature similarity of the second segment, may include the following steps:

[0099] A first loss value is determined based on the feature similarity of the first segment; a second loss value is determined based on the feature similarity of the second segment; a model loss value is determined based on the sum of the first loss value and the second loss value; and the model parameters of the singing feature extraction model are adjusted based on the model loss value.

[0100] Among them, the first loss value is negatively correlated with the feature similarity of the first segment; the second loss value is positively correlated with the feature similarity of the second segment.

[0101] In practice, the appropriateness of the extracted vocal features depends on the following two aspects: First, the vocal features of the same song should be close to the feature center, which helps to effectively partition the characteristics of different vocals; Second, the feature centers of songs with different vocal qualities (such as the feature centers of the winning song and the losing song) should be as far apart as possible, so that the winning and losing songs can be clearly distinguished.

[0102] To address this, a first loss value can be determined based on the first feature similarity. By making the first loss value negatively correlated with the first feature similarity, the singing feature extraction model can make the features of singing segments belonging to the same sample singing more closely related when extracting singing features. Furthermore, a second loss value can be determined based on the second feature similarity. By making the second loss value positively correlated with the second feature similarity, the feature centers of different samples singing can be further separated, increasing the distinguishability between singing voices.

[0103] Then, the model loss value can be determined based on the sum of the first and second loss values. In one example, the model loss value can be determined according to the following mapping relationship (also known as the clause center loss function). :

[0104]

[0105] in, The sum of the distances from all the vocal fragment features (such as embedding) of the first sample vocal voice to the feature center A is given. The first sample vocal voice has a total of N vocal fragment features. d represents the sum of distances from all vocal segment features of the second sample singing to the feature center B. The second sample singing has a total of M vocal segment features. ab This represents the distance between feature center A and feature center B; the coefficient k is a preset value used to control the weight of the first and second loss values. Furthermore, to ensure the loss function is greater than 0, optimization can be performed only on songs with a difference greater than 0 between the two terms.

[0106] In this embodiment, by determining the model loss value based on the sum of the first loss value and the second loss value, the singing features extracted by the singing feature extraction model for the same song can be made closer to each other, and the singing features of different songs can be made more consistent, thereby increasing the distinguishability between singing voices of different quality and helping to obtain more accurate singing features.

[0107] To enable those skilled in the art to better understand the above steps, the following example illustrates the embodiments of this application, but it should be understood that the embodiments of this application are not limited thereto.

[0108] like Figure 4 As shown, multiple sample singing voices (i.e., multiple training samples) can be obtained, and each sample singing voice can be an audio recording of the same song sung by different users. For multiple sample singing voices, two sample singing voices can be randomly selected, and the user can manually identify which sample singing voice sounds better and label it to form a sample singing voice pair.

[0109] Then, the melodic features of each sample song in the sample song pair can be extracted by segmenting the song into phrases. The melodic features of each phrase of each sample song can then be input into the song feature extraction model to be trained. This model can be a CRNN network (Convolutional Recurrent Neural Network) to obtain the song segment features of each sample song. Based on the song segment features of each sample song, the loss function of the phrase center is substituted to calculate the model loss value.

[0110] In one embodiment, positive sample singing features and negative sample singing features can be obtained through the following steps:

[0111] S1, obtain the positive sample singing voice set and the negative sample singing voice set; the singing quality of each sample singing voice in the positive sample singing voice set is higher than the singing quality of each sample singing voice in the negative sample singing voice set.

[0112] In practice, two sets of songs that have been manually labeled can be obtained: a set of positive sample songs and a set of negative sample songs. The quality of all songs in the positive sample song set is higher than that of any song in the negative sample song set.

[0113] S2, based on the singing features of each sample singing in the positive sample singing set, obtain the feature center of each sample singing in the positive sample singing set as the positive sample singing feature; and based on the singing features of each sample singing in the negative sample singing set, obtain the feature center of each sample singing in the negative sample singing set as the negative sample singing feature.

[0114] In some embodiments, after obtaining the set of positive sample singing voices, each sample singing voice in the set can be segmented and its spectral features extracted. The spectral features of each segment are input into a trained singing voice feature extraction model. The singing voice feature extraction model determines the singing features corresponding to each segment based on the spectral features of that segment. Then, a feature center can be determined based on the singing voice features of each sample singing voice segment. For example, the mean of the singing voice features of each segment can be used as the feature center. Furthermore, this feature center can be determined as the positive sample singing voice feature.

[0115] Similarly, after obtaining the set of negative sample singing voices, each sample singing voice in the set can be segmented and its spectral features extracted. The spectral features of each segment are input into a trained singing voice feature extraction model. The model determines the singing voice features corresponding to each segment based on its spectral features. Then, a feature center can be determined based on the singing voice features of each sample singing voice segment. For example, the mean of the singing voice features of each segment can be used as the feature center. Furthermore, this feature center can be determined as the negative sample singing voice feature.

[0116] In this embodiment, by obtaining a set of positive sample singing voices and a set of negative sample singing voices, and ensuring that the singing quality of each sample singing voice in the set of positive sample singing voices is higher than that of each sample singing voice in the set of negative sample singing voices, the positive sample singing voice features and negative sample singing voice features are determined based on the singing voice features of the sample singing voices in each set. This approach can comprehensively determine the positive sample singing voice features and negative sample singing voice features by integrating multiple samples, thereby improving the accuracy and comprehensiveness of the positive sample singing voice features and negative sample singing voice features used as subsequent reference benchmarks.

[0117] In one embodiment, step S103, determining the singing evaluation information of the singing voice to be evaluated based on the first similarity and the second similarity, may include:

[0118] If the first similarity is greater than the second similarity, the evaluation score adjustment value is determined based on the first similarity; if the first similarity is less than the second similarity, the evaluation score adjustment value is determined based on the second similarity; based on the benchmark evaluation score and the evaluation score adjustment value, the singing evaluation score is obtained, and the singing evaluation score is used as the singing evaluation information.

[0119] Among them, the adjusted evaluation score is positively correlated with the first similarity and negatively correlated with the second similarity.

[0120] In practical applications, if the first similarity score is greater than the second similarity score, it indicates that the singing voice to be evaluated is more similar to the positive sample singing voice. This can be interpreted as the singing quality of the singing voice to be evaluated being above average. Therefore, an adjustment value for the evaluation score can be determined based on the first similarity score. This adjustment value is used to adjust the baseline evaluation score. The adjustment value for the evaluation score is positively correlated with the first similarity score, meaning that the closer the singing characteristics of the singing voice to the positive sample singing voice, the higher the singing quality of the singing voice to be evaluated, and the larger the value of the adjustment value for the evaluation score, the greater the increase in the evaluation score.

[0121] If the first similarity score is less than the second similarity score, then the singing voice to be evaluated is more similar to the negative sample singing voice. This can be interpreted as the singing quality of the singing voice to be evaluated being below average. Therefore, the evaluation score adjustment value can be determined based on the second similarity score. The evaluation score adjustment value is negatively correlated with the second similarity score, meaning that the closer the singing characteristics of the singing voice to be evaluated are to the characteristics of the negative sample singing voice, the lower the singing quality of the singing voice to be evaluated, the smaller the value of the evaluation score adjustment value, and the greater the reduction in the evaluation score.

[0122] In practical applications, the evaluation score adjustment value can correspond one-to-one with the first similarity or the second similarity. That is, under different first similarity or second similarity, there can be a unique corresponding evaluation score adjustment value. Thus, subtle differences between singing voices can be reflected in the evaluation score adjustment value through the first similarity or the second similarity.

[0123] Then, the singing evaluation score can be determined based on the baseline evaluation score and the evaluation score adjustment value. Specifically, the baseline evaluation score can be adjusted using the evaluation score adjustment value; for example, the baseline evaluation score and the evaluation score adjustment value can be summed, and the sum can be used as the singing evaluation score. For example, if the first similarity is greater than the second similarity, the singing evaluation score can be determined as follows: :

[0124]

[0125] in, The distance between the singing features and the singing features of the positive samples is negatively correlated with the first similarity.

[0126] If the first similarity score is less than the second similarity score, the singing evaluation score can be determined as follows. :

[0127]

[0128] in, The distance between the singing features and the singing features of the negative samples is negatively correlated with the second similarity.

[0129] In this embodiment, by comparing the first similarity and the second similarity, an evaluation score adjustment value that is positively correlated with the first similarity or negatively correlated with the second similarity can be determined. This allows subtle differences between singing voices to be reflected in the evaluation score adjustment value through the first or second similarity, resulting in an accurate singing evaluation score.

[0130] In one embodiment, the vocal features corresponding to the vocal performance to be evaluated may include vocal features corresponding to each of multiple vocal segments of the vocal performance to be evaluated.

[0131] In this embodiment, vocal features can be extracted from the singing to be evaluated on a segment-by-segment basis. In some embodiments, for each vocal audio file, the audio can be segmented into sentences, resulting in multiple vocal segments. These segments can then be input into a trained vocal feature extraction model, which extracts the vocal features corresponding to each segment. Segmenting into sentences aligns with the user's habit of singing in segments, ensuring the integrity of information within each segment and enabling analysis. In one example, the spectral characteristics of the vocal segments can be input into the vocal feature extraction model, which then determines the vocal features of the segments based on the input spectral characteristics.

[0132] Accordingly, the singing evaluation score is obtained based on the baseline evaluation score and the evaluation score adjustment value, which may include the following steps:

[0133] The evaluation score for each singing segment is determined based on the baseline evaluation score and the evaluation score adjustment value for each singing segment; the overall singing evaluation score is determined based on the individual singing segment evaluation scores of the multiple singing segments.

[0134] Specifically, for each vocal segment, an adjustment value for the evaluation score of that vocal segment can be determined according to the aforementioned embodiments. This adjustment value is then used to adjust the baseline evaluation score, resulting in the vocal segment evaluation score for that vocal segment. Furthermore, the overall vocal evaluation score can be determined by combining the individual vocal segment evaluation scores of multiple vocal segments to be evaluated. For example, the average of the vocal segment evaluation scores of multiple vocal segments can be taken as the overall vocal evaluation score.

[0135] In this embodiment, the quality of the singing is evaluated according to the singing segments, and a singing segment evaluation score is obtained for each singing segment. The singing evaluation result is determined by combining the evaluation scores of multiple singing segments. This allows for fine-grained analysis of the singing audio and improves the accuracy of the singing evaluation score.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores vocal data. The I / O interfaces allow the processor to exchange information with external devices. The communication interface allows communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for acquiring vocal evaluation information.

[0138] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for acquiring singing evaluation information. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0139] Those skilled in the art will understand that Figure 5 and Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0140] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0141] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0142] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0143] 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, data stored, data displayed, 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 related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0144] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media 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), magnetic 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. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0145] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0146] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for obtaining singing evaluation information, characterized in that, The method includes: Obtain the vocal characteristics corresponding to the singing voice to be evaluated; The singing features are compared with the pre-acquired positive sample singing features and negative sample singing features to obtain a first similarity between the singing features and the positive sample singing features, and a second similarity between the singing features and the negative sample singing features; wherein, the singing quality of the positive sample singing is higher than that of the negative sample singing. Based on the first similarity and the second similarity, the singing evaluation information of the singing voice to be evaluated is determined.

2. The method according to claim 1, characterized in that, The positive sample singing features and the negative sample singing features are obtained through the following steps: Obtain a set of positive sample singing voices and a set of negative sample singing voices; the singing quality of each sample singing voice in the set of positive sample singing voices is higher than the singing quality of each sample singing voice in the set of negative sample singing voices; Based on the singing characteristics of each sample singing voice in the positive sample singing voice set, the feature center of each sample singing voice in the positive sample singing voice set is obtained as the positive sample singing voice feature; and based on the singing characteristics of each sample singing voice in the negative sample singing voice set, the feature center of each sample singing voice in the negative sample singing voice set is obtained as the negative sample singing voice feature.

3. The method according to claim 1, characterized in that, The step of determining the singing evaluation information of the singing voice to be evaluated based on the first similarity and the second similarity includes: If the first similarity is greater than the second similarity, then an adjustment value for the evaluation score is determined based on the first similarity, and the adjustment value for the evaluation score is positively correlated with the first similarity. If the first similarity is less than the second similarity, then an adjustment value for the evaluation score is determined based on the second similarity, wherein the adjustment value for the evaluation score is negatively correlated with the second similarity; Based on the baseline evaluation score and the adjusted evaluation score, a singing evaluation score is obtained, and this singing evaluation score is used as singing evaluation information.

4. The method according to claim 3, characterized in that, The vocal features corresponding to the singing voice to be evaluated include: vocal features corresponding to each of the multiple vocal segments of the singing voice to be evaluated; each vocal segment corresponds to a baseline evaluation score and an adjustment value for the evaluation score; The process of obtaining the singing evaluation score based on the baseline evaluation score and the evaluation score adjustment value includes: The evaluation score for each singing segment is determined based on the baseline evaluation score and the evaluation score adjustment value for each singing segment. The singing evaluation score is determined based on the individual singing segment evaluation scores of the multiple singing segments.

5. The method according to any one of claims 1 to 4, characterized in that, The vocal features to be evaluated are obtained through a vocal feature extraction model; the vocal feature extraction model is trained through the following steps: Obtain sample singing pairs; each sample singing pair includes two singing samples with different singing qualities; Multiple vocal segments from each of the sample vocal pairs are input into the vocal feature extraction model to be trained to obtain vocal segment features corresponding to each vocal segment. The vocal features of the sample singing are determined based on the features of the multiple vocal segments. Based on the similarity between the features of each singing segment and the singing features of the corresponding sample singing, a first feature similarity is determined, and a second feature similarity is obtained between the singing features of each sample singing. Based on the first feature similarity and the second feature similarity, the model parameters of the singing feature extraction model are adjusted to obtain a trained singing feature extraction model.

6. The method according to claim 5, characterized in that, The step of determining the first feature similarity based on the similarity between the features of each vocal segment and the vocal features of the corresponding sample vocal segment includes: Determine the similarity between the features of each singing segment and the singing features of the corresponding sample singing; The first feature similarity is obtained by summing the similarities described above.

7. The method according to claim 5, characterized in that, The step of determining the vocal features of the sample vocalization based on the features of the multiple vocal segments includes: For each sample singing voice, the mean value of the features of multiple singing voice segments corresponding to the sample singing voice is obtained; The vocal characteristics of the sample singing can be obtained based on the mean value.

8. The method according to claim 5, characterized in that, The step of adjusting the model parameters of the singing feature extraction model based on the first feature similarity and the second feature similarity includes: A first loss value is determined based on the first feature similarity; the first loss value is negatively correlated with the first feature similarity. A second loss value is determined based on the second feature similarity; the second loss value is positively correlated with the second feature similarity. The model loss value is determined based on the sum of the first loss value and the second loss value, and the model parameters of the singing feature extraction model are adjusted based on the model loss value.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images