Training method of audio feature extraction model and cover audio recognition method
By employing self-supervised training and data augmentation, and using unlabeled original song fragments for pre-training combined with labeled cover song fragments for fine-tuning, the performance bottleneck of the cover song recognition technology was solved, achieving efficient recognition of cover songs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TENCENT MUSIC ENTERTAINMENT TECH (SHENZHEN) CO LTD
- Filing Date
- 2023-10-30
- Publication Date
- 2026-06-05
AI Technical Summary
Existing cover song recognition technology suffers from performance bottlenecks due to the difficulty in labeling training data, making it difficult to accurately match the differences between the cover song and the original work.
By employing a self-supervised training method, we pre-trained the system using unlabeled original song clips, constructed song clips for data augmentation, and fine-tuned it by combining labeled cover song clips to build an audio feature extraction model.
The performance of the audio feature extraction model has been improved, enabling effective identification of cover songs with a small number of labeled samples, thus avoiding performance bottlenecks caused by insufficient label data.
Smart Images
Figure CN117636898B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of audio technology, and in particular to a training method for an audio feature extraction model, a method for recognizing cover songs, a computer device, and a computer-readable storage medium. Background Technology
[0002] Cover song recognition technology is an important supplement to song recognition technology. However, when it is necessary to identify some new cover songs or original works, it is difficult to make an accurate match because there are certain differences between the cover songs and the original works.
[0003] Currently, cover song recognition typically uses short-segment training to model the song. This requires a large number of labeled samples for model training, which makes it difficult to collect and label training data. The limited amount of labeled sample data leads to performance bottlenecks in the model. Summary of the Invention
[0004] Therefore, it is necessary to provide a training method for an audio feature extraction model, a cover song audio recognition method, a computer device, and a computer-readable storage medium that can improve the performance of the audio feature extraction model, in response to the above-mentioned technical problems.
[0005] Firstly, this application provides a method for training an audio feature extraction model, including:
[0006] The first audio feature extraction model is trained using a first original song fragment, a constructed song fragment, and first other song fragments to obtain the trained first audio feature extraction model; the constructed song fragment is obtained by data augmentation of the first original song fragment.
[0007] The trained first audio feature extraction model is used as an encoder to construct the second audio feature extraction model;
[0008] The second audio feature extraction model is trained using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments to obtain a trained second audio feature extraction model; the trained second audio feature extraction model is used to identify songs from input cover audio fragments.
[0009] The first original song fragment is a song fragment from an unlabeled original song; the second original song fragment is a song fragment from a labeled original song, and the second original song fragment and the corresponding cover song fragment have a pre-labeled association relationship.
[0010] In one embodiment, the step of training the first audio feature extraction model using a first original song fragment, a constructed song fragment, and a first set of other song fragments to obtain the trained first audio feature extraction model includes:
[0011] Based on the first original song without a tag, obtain a fragment of the first original song, and generate the constructed song fragment based on the fragment of the first original song;
[0012] The constructed song fragment is taken as the positive fragment corresponding to the first original song fragment, and the first other song fragment is taken as the negative fragment corresponding to the first original song fragment;
[0013] The first original song fragment, the positive fragment corresponding to the first original song fragment, and the negative fragment corresponding to the first original song fragment are input into the first audio feature extraction model to obtain the first audio feature extraction combination.
[0014] The first audio feature extraction combination is used for model training to obtain the trained first audio feature extraction model.
[0015] In one embodiment, generating the constructed song fragment based on the first original song fragment includes:
[0016] The target data augmentation method is randomly selected from the candidate data augmentation methods;
[0017] The first original song segment is augmented using the target data augmentation method to obtain the constructed song segment;
[0018] The candidate data augmentation methods include any one or more of the following:
[0019] Speed change, pitch shift, pitch offset, noise addition, random masking.
[0020] In one embodiment, the step of training the model using the first audio feature extraction combination to obtain the trained first audio feature extraction model includes:
[0021] Based on the first audio feature extraction combination, determine the target loss value;
[0022] The first audio feature extraction model is adjusted using the target loss value until the first training termination condition is met, thus obtaining the trained first audio feature extraction model.
[0023] In one embodiment, the step of using the trained first audio feature extraction model as an encoder to construct a second audio feature extraction model includes:
[0024] The trained first audio feature extraction model is used as an encoder, and the parameters of the encoder are fixed.
[0025] A preset convolutional network and a fully connected layer are concatenated at the output of the encoder to obtain the second audio feature extraction model.
[0026] In one embodiment, the step of training the second audio feature extraction model using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and a second other song fragment, to obtain the trained second audio feature extraction model, includes:
[0027] Based on the second original song and its cover song carrying tags, obtain the second original song fragment and the cover song fragment;
[0028] Based on the second original song fragment, the cover song fragment, and the second other song fragments, the parameters of the second audio feature extraction model are adjusted until the second training termination condition is met, thus obtaining the trained second audio feature extraction model.
[0029] Secondly, this application also provides a method for recognizing cover song audio, including:
[0030] Obtain the cover song audio segment to be identified, input the cover song audio segment into the trained second audio feature extraction model, and obtain the cover song audio features corresponding to the cover song audio segment;
[0031] Based on the degree of feature similarity between the cover song audio features and multiple song audio features, target song audio features that meet preset similarity conditions are determined; the song audio features are obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model;
[0032] The songs in the music library corresponding to the audio features of the target song are used as the song recognition results of the cover audio segment;
[0033] The trained second audio feature extraction model is obtained by training a second original song fragment, a corresponding cover song fragment, and two other song fragments. The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model is obtained by training a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment is obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment from an unlabeled original song. The second original song fragment is a song fragment from a labeled original song, and there is a pre-labeled association between the second original song fragment and the corresponding cover song fragment.
[0034] In one embodiment, determining the target song audio features that meet preset similarity conditions based on the feature similarity between the cover song audio features and multiple song audio features includes:
[0035] The songs in each music library are segmented into audio segments according to a preset duration to obtain a set of audio segments for each of the aforementioned music libraries.
[0036] Each audio segment in the audio segment set is input into the trained second audio feature extraction model to obtain the multiple song audio features;
[0037] The cosine distance between the cover song audio features and any song audio features is used as the feature similarity.
[0038] The song audio feature that is closest to the cover song audio feature is taken as the target song audio feature.
[0039] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the training method for the audio feature extraction model as described in the first aspect, and / or the steps of the cover song audio recognition method as described in the second aspect.
[0040] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the training method for the audio feature extraction model as described in the first aspect, and / or the steps of the cover song audio recognition method as described in the second aspect.
[0041] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of a training method for an audio feature extraction model as described in the first aspect, and / or the steps of a cover song audio recognition method as described in the second aspect.
[0042] The aforementioned training method for an audio feature extraction model, a cover song audio recognition method, a computer device, and a computer-readable storage medium train a first audio feature extraction model using a first original song fragment, a constructed song fragment, and a first set of other song fragments. The constructed song fragment is obtained by data augmentation of the first original song fragment. Then, the trained first audio feature extraction model is used as an encoder to construct a second audio feature extraction model. This second model is trained using a second original song fragment, a corresponding cover song fragment, and a second set of other song fragments. The trained second audio feature extraction model is then used to recognize the input cover song fragment. This method optimizes the training of the audio feature extraction model. Data augmentation allows the first audio feature extraction model to adapt to information perturbations in the fragments for self-supervised training. Furthermore, labeled cover song fragment samples are used for fine-tuning, eliminating the need for a large number of labeled samples and avoiding performance bottlenecks caused by insufficient labeled data, thus effectively improving model performance. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0044] Figure 1 This is a flowchart illustrating the training method of an audio feature extraction model in one embodiment;
[0045] Figure 2a This is a schematic diagram of the pre-training stage processing flow in one embodiment;
[0046] Figure 2b This is a schematic diagram of the model fine-tuning stage processing flow in one embodiment;
[0047] Figure 3 This is a flowchart illustrating a cover song audio recognition method in one embodiment;
[0048] Figure 4 This is a flowchart illustrating the training method of the audio feature extraction model in another embodiment;
[0049] Figure 5 This is a structural block diagram of a training device for an audio feature extraction model in one embodiment;
[0050] Figure 6 This is a structural block diagram of a cover song audio recognition device in one embodiment;
[0051] Figure 7 This is an internal structural diagram of a computer device in one embodiment;
[0052] Figure 8 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation
[0053] 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.
[0054] Cover song recognition technology is an important supplement to song recognition technology. Song recognition technology achieves accurate matching through audio fingerprints, and can still accurately identify songs even with background noise when the recording sample is the original. However, when it is necessary to identify new cover songs, audio fingerprint matching is difficult to achieve ideal results due to the differences between the cover and the original.
[0055] Traditional cover song recognition technology is mainly applied to full-song covers, identifying covers from the song level. However, song recognition involves short segments, with user-recorded audio files typically under 15 seconds in length, making recognition more challenging. Traditional methods use short segments for training and modeling, which, due to the difficulty in labeling training data and the limited amount of supervised data, leads to performance bottlenecks. To address these issues, this application employs a pre-training method, using a large number of unlabeled samples for self-supervised training, followed by fine-tuning based on a limited labeled dataset. This allows for overcoming performance bottlenecks with a small number of labeled samples, further improving the performance of the audio feature extraction model.
[0056] In one embodiment, such as Figure 1 As shown, a training method for an audio feature extraction model is provided. This embodiment illustrates the method by applying it to a server. The method includes the following steps S101 to S103:
[0057] In step S101, the first audio feature extraction model is trained using the first original song fragment, the constructed song fragment, and the first other song fragments to obtain the trained first audio feature extraction model.
[0058] As an example, the first original song fragment can be a song fragment from an unlabeled original song, that is, the first original song fragment can be obtained based on the first original song without labels. For example, unlabeled song samples can be collected from the music library as the first original song, so as to train the first audio feature extraction model in a self-supervised manner based on the unlabeled samples; the first other song fragments can be audio fragments from other songs besides the first original song. For example, the first original song fragment is obtained from song A (original version), and the first other song fragments are obtained from song B. Song A and song B are two completely different songs.
[0059] The constructed song fragment can be obtained by data augmentation of the first original song fragment, such as by using data augmentation methods such as speed change, pitch shift, pitch shift, noise addition, and random masking.
[0060] In practical applications, by collecting unlabeled song samples from the music library, we can obtain the first original song fragment, the constructed song fragment, and the first other song fragments to construct a self-supervised pre-trained triplet, which can then be used to train the first audio feature extraction model to obtain the trained first audio feature extraction model.
[0061] Specifically, a large number of song samples can be collected from the music library as pre-training corpus. For example, each song can be sliced into segments of a preset duration (such as 3 seconds) to construct self-supervised pre-training triples for model pre-training.
[0062] In step S102, the trained first audio feature extraction model is used as an encoder to construct the second audio feature extraction model;
[0063] In practice, after pre-training, the trained first audio feature extraction model can be used as an encoder to construct a second audio feature extraction model. For example, a small convolutional network can be spliced at the output of the pre-trained model (i.e., the trained first audio feature extraction model) to further adjust the model.
[0064] In step S103, the second audio feature extraction model is trained using the second original song fragment, the cover song fragment corresponding to the second original song fragment, and the second other song fragments, to obtain the trained second audio feature extraction model.
[0065] The trained second audio feature extraction model can be used to identify songs from input cover audio clips.
[0066] The second original song fragment can be a song fragment from an original song with annotations, and the second original song fragment has a pre-annotated association with the corresponding cover song fragment.
[0067] After obtaining the second audio feature extraction model, based on the original song and its cover versions with labels, fragments of the original song and cover versions can be obtained. This second audio feature extraction model can then be trained using other song fragments to obtain the trained model. By fine-tuning the model with labeled cover sample fragments, the model can learn the melodic connections between the original and cover segments, as well as the differences between different song fragments, based on a small number of labeled samples. After fine-tuning, the model (i.e., the trained second audio feature extraction model) can be used for further database construction and recognition.
[0068] Compared to traditional methods that require labeling a large number of samples, which is time-consuming and labor-intensive, the technical solution in this embodiment avoids the performance bottleneck caused by insufficient labeled data for cover songs by using a pre-training method. Furthermore, the self-supervised training method used in the pre-training stage can include all songs in the music library into the training samples, which can avoid the problem of the model not recognizing songs it has never seen before, and further improve the model's performance.
[0069] In the training method of the above-mentioned audio feature extraction model, the first audio feature extraction model is trained by using a first original song fragment, a constructed song fragment, and a first other song fragment, resulting in a trained first audio feature extraction model. Then, the trained first audio feature extraction model is used as an encoder to construct a second audio feature extraction model. The second audio feature extraction model is trained by using a second original song fragment, a corresponding cover song fragment, and a second other song fragment, resulting in a trained second audio feature extraction model. This achieves training optimization of the audio feature extraction model. Data augmentation allows the first audio feature extraction model to adapt to the information perturbation of the fragments for self-supervised training. Furthermore, labeled cover song fragment samples are used for fine-tuning, eliminating the need for a large number of labeled samples and avoiding the model performance bottleneck caused by insufficient labeled data, thus effectively improving model performance.
[0070] In one embodiment, step S101, using a first original song fragment, a constructed song fragment, and a first other song fragment, to train the first audio feature extraction model to obtain the trained first audio feature extraction model, may include the following steps:
[0071] Based on the first original song without tags, obtain a fragment of the first original song, and generate a constructed song fragment based on the first original song fragment; use the constructed song fragment as the positive fragment corresponding to the first original song fragment, and use the first other song fragments as the negative fragments corresponding to the first original song fragment; input the first original song fragment, the positive fragment corresponding to the first original song fragment, and the negative fragment corresponding to the first original song fragment into the first audio feature extraction model to obtain the first audio feature extraction combination; use the first audio feature extraction combination to train the model to obtain the trained first audio feature extraction model.
[0072] In practical applications, such as Figure 2a As shown, a large number of untagged songs (i.e., the first original songs without tags) can be filtered from the music library by sorting based on play count, such as... Figure 2a Song 1, Song 2, ..., Song n are used as a pre-training dataset. Then, samples can be randomly drawn from the pre-training dataset according to the size of each training batch. For example, when the batch size is 32, 32 songs can be drawn. For each song, it can be divided into multiple 3-second segments, which are the first original song segments.
[0073] In one alternative embodiment, song fragments can be obtained by extracting raw audio or by extracting spectral features to obtain data for the input model. No specific limitations are imposed in this embodiment.
[0074] For example, by collecting a large number of song samples from a music library as pre-training corpus, segments of the original song (i.e., the first original song segment) can be used as anchors, and new segments (i.e., constructed song segments) can be constructed using data augmentation as positive segments, while other song segments (i.e., the first other song segments) can be used as negative segments to construct self-supervised pre-training triples (such as...). Figure 2a The model is pre-trained using the original segment x, the enhanced segment x, and other segments y. This not only allows the model to adapt to information perturbations in the segments through data augmentation, but also enables the model to acquire a large number of audio samples. Since a large number of samples have appeared during training, they are easier to recognize when applied.
[0075] In one example, such as Figure 2a As shown, after obtaining the enhanced segment x, any segment from other songs can be randomly selected as other segments y to construct triples. Then, a batch of triples can be input into SampleCNN (convolutional neural networks) for feature extraction, thus obtaining the first audio feature extraction combination.
[0076] In yet another example, such as Figure 2a As shown, the first audio feature extraction model can be the neural network SampleCNN, or other convolutional neural networks such as ResNet can be used for processing. No specific restrictions are imposed in this embodiment.
[0077] In this embodiment, a first original song fragment is obtained based on a first original song without a label, and a constructed song fragment is generated based on the first original song fragment. Then, the constructed song fragment is used as the positive fragment corresponding to the first original song fragment, and other song fragments are used as the negative fragments corresponding to the first original song fragment. The first original song fragment, the positive fragment corresponding to the first original song fragment, and the negative fragment corresponding to the first original song fragment are input into a first audio feature extraction model to obtain a first audio feature extraction combination. The first audio feature extraction combination is then used for model training to obtain a trained first audio feature extraction model. This model can use a self-supervised training method in the pre-training stage, avoiding the problem that the model is not familiar with the song and therefore cannot easily identify it, which helps to further improve the model performance.
[0078] In one embodiment, generating a constructed song fragment based on a first original song fragment may include the following steps:
[0079] From the candidate data augmentation methods, a target data augmentation method is randomly selected; the target data augmentation method is used to augment the data of the first original song segment to obtain the constructed song segment.
[0080] Candidate data augmentation methods may include any one or more of the following:
[0081] Speed shifting, pitch shifting, pitch offsetting, noise addition, random masking; other data enhancement methods may also be included, but are not specifically limited in this embodiment.
[0082] In specific implementations, such as Figure 2a As shown, after obtaining the original fragment x, for any original fragment x, an augmentation method can be randomly selected from speed variation, pitch variation, pitch shift, noise addition, and random mask (i.e., candidate data augmentation methods) for data augmentation. The purpose of randomly selecting the augmentation method is to ensure that various perturbations can fully cover various samples, thereby ensuring the diversity of training data and allowing the model to learn the invariance between samples after various perturbations and the original samples.
[0083] For example, due to the introduction of various complex noises in real recording scenarios, the adapted version may also have changes in speed and pitch, and there may be interruptions in the recording. Corresponding to the interruption and the random mask, the random mask can randomly set the audio sampling points to zero.
[0084] In this embodiment, by randomly selecting a target data augmentation method from the candidate data augmentation methods, and then using the target data augmentation method to augment the data of the first original song segment, a constructed song segment is obtained. Data augmentation can enable the model to adapt to the information perturbation of the segment, which helps to improve the accuracy of model recognition.
[0085] In one embodiment, training a model using a first audio feature extraction combination to obtain a trained first audio feature extraction model may include the following steps:
[0086] Based on the first audio feature extraction combination, determine the target loss value; adjust the first audio feature extraction model using the target loss value until the first training termination condition is met, and obtain the trained first audio feature extraction model.
[0087] In one example, the following loss function can be used for calculation:
[0088] L=max(d(a,p)-d(a,n)+margin,0)
[0089] Where d(a, p) is the cosine distance between the original anchor (i.e., the first original song segment) and the enhanced positive (i.e., the constructed song segment), d(a, n) is the cosine distance between the original anchor and other segments negative (i.e., the first other song segments), and margin is an adjustable degree coefficient. For the first audio feature extraction model, the training goal is to make the original segment and the enhanced segment as close as possible in the embedding space, and as far away as possible from other segments.
[0090] In another example, the network can be saved after training until the loss function value (i.e., the target loss value) is below a certain threshold, that is, until the first training termination condition is met, to obtain the first audio feature extraction model after training.
[0091] In this embodiment, a target loss value is determined based on the first audio feature extraction combination, and then the first audio feature extraction model is adjusted using the target loss value until the first training termination condition is met, thus obtaining the trained first audio feature extraction model. This allows for a self-supervised training method during the pre-training stage, further improving the model's performance.
[0092] In one embodiment, step S102, using the trained first audio feature extraction model as an encoder to construct the second audio feature extraction model, may include the following steps:
[0093] The trained first audio feature extraction model is used as the encoder, and the encoder parameters are fixed. A pre-defined convolutional network and a fully connected layer are spliced at the output of the encoder to obtain the second audio feature extraction model.
[0094] In practical applications, such as Figure 2b As shown, the audio feature extraction model saved in the pre-training stage (i.e., the first audio feature extraction model after training) can be used as the encoder in the model fine-tuning stage. In the fine-tuning stage, the parameters of the encoder can be frozen, followed by three layers of convolutional neural network and one fully connected layer to obtain the second audio feature extraction model.
[0095] In this embodiment, the trained first audio feature extraction model is used as the encoder. The parameters of the encoder are fixed, and then a preset convolutional network and a fully connected layer are spliced at the output of the encoder to obtain the second audio feature extraction model, which provides support for further model fine-tuning.
[0096] In one embodiment, step S103, using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments, to train the second audio feature extraction model to obtain the trained second audio feature extraction model, may include the following steps:
[0097] Based on the original song and its cover song with tags, obtain the original song fragment and the cover song fragment; based on the original song fragment, the cover song fragment, and other song fragments, adjust the parameters of the second audio feature extraction model until the second training termination condition is met, and obtain the trained second audio feature extraction model.
[0098] In practical implementation, labeled cover song samples can be used to train the model, for example, such as Figure 2b As shown, song fragment A can be the first line of the original version of song A, and the cover version of song fragment A can be the first line of the adapted version of song A sung by the cover singer. The lyrics of the two are the same. Other fragments B can be any line of other songs B (not song A), that is, the second original song fragment, the cover song fragment, and the second other song fragment.
[0099] Optionally, the loss function can be adjusted using triplet data to fine-tune the model, enabling the pre-trained model to quickly learn the melodic characteristics of cover songs. The model can be saved after the loss function stabilizes below a certain threshold, i.e., the parameters of the second audio feature extraction model are adjusted until the second training termination condition is met, resulting in the trained second audio feature extraction model.
[0100] In this embodiment, by obtaining a second original song fragment and a cover song fragment based on the second original song and its cover song with tags, the parameters of the second audio feature extraction model are adjusted based on the second original song fragment, the cover song fragment, and other second song fragments until the second training termination condition is met, resulting in the trained second audio feature extraction model. This model can be fine-tuned based on the cover song fragment samples with tags, without the need for a large number of labeled samples. This avoids the model performance bottleneck caused by insufficient labeled data and effectively improves the model performance.
[0101] In one embodiment, such as Figure 3 As shown, a method for recognizing cover song audio is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S301 to S303. Wherein:
[0102] In step S301, the cover audio segment to be identified is obtained, and the cover audio segment is input into the trained second audio feature extraction model to obtain the cover audio features corresponding to the cover audio segment.
[0103] In step S302, the target song audio features that meet the preset similarity conditions are determined based on the feature similarity between the cover song audio features and the audio features of multiple songs.
[0104] Among them, the song audio features can be obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model.
[0105] In step S303, the songs in the music library corresponding to the audio features of the target song are used as the song recognition results of the cover audio segment.
[0106] The trained second audio feature extraction model can be obtained by training the second original song segment, the cover song segment corresponding to the second original song segment, and other song segments.
[0107] The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model can be trained using a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment can be obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment in an unlabeled original song. The second original song fragment is a song fragment in a labeled original song, and the second original song fragment has a pre-labeled association with the corresponding cover song fragment.
[0108] In practical applications, to address the problem of limited labeled training samples, a self-supervised training method can be used to add songs from the music library, regardless of whether they have song grouping information or timestamp information, to the model for training. The resulting pre-trained model (i.e., the first audio feature extraction model after training) can be equivalent to a dictionary of the music library, which records a large amount of song information. Then, it can be fine-tuned using a small number of labeled cover song fragments, so that the model (i.e., the second audio feature extraction model after training) can adapt to the cover song recognition task, focusing on extracting the melody information of the audio fragments, and further improving the cover song audio recognition effect.
[0109] In the above-mentioned cover song audio recognition method, the cover song audio segment to be recognized is obtained, and the cover song audio segment is input into the trained second audio feature extraction model to obtain the cover song audio features corresponding to the cover song audio segment. Then, based on the feature similarity between the cover song audio features and the audio features of multiple songs, the target song audio features that meet the preset similarity conditions are determined. The song audio features are obtained by inputting the audio segments of songs in the music library into the trained second audio feature extraction model. Then, the music library songs corresponding to the target song audio features are used as the song recognition results of the cover song audio segment. This realizes the optimization of the audio feature extraction model based on training, which can avoid the model performance bottleneck problem caused by insufficient label data, effectively improve the model performance, and improve the cover song audio recognition effect.
[0110] In one embodiment, step S302, determining the target song audio features that meet preset similarity conditions based on the feature similarity between the cover song audio features and multiple song audio features, may include the following steps:
[0111] The songs in each music library are segmented into audio segments according to a preset duration to obtain a set of audio segments for each music library. Each audio segment in the set of audio segments is input into the trained second audio feature extraction model to obtain multiple song audio features. The cosine distance between the cover song audio feature and any song audio feature is used as the feature similarity. The song audio feature with the smallest distance from the cover song audio feature is used as the target song audio feature.
[0112] In one example, such as Figure 2bAs shown, song segments from the music library can be input into the fine-tuned model (i.e., the trained second audio feature extraction model). The library is built using the embedding features (i.e., multiple song audio features) output from the fully connected layer. Then, after obtaining the cover song audio to be predicted, the same cover song audio can be input into the fine-tuned model to extract the embedding features (i.e., cover song audio features). The cosine distance between the embedding feature of the sample to be predicted and the embeddings of other segments in the search library can be calculated. Then, the song corresponding to the closest segment can be selected as the cover song recognition result, that is, the song in the music library corresponding to the target song audio features is used as the song recognition result of the cover song audio segment.
[0113] In this embodiment, the songs in each music library are segmented into audio segments according to a preset duration to obtain a set of audio segments for each music library. Then, each audio segment in the audio segment set is input into the trained second audio feature extraction model to obtain multiple song audio features. The cosine distance between the cover song audio feature and any song audio feature is used as the feature similarity. The song audio feature with the smallest distance from the cover song audio feature is then used as the target song audio feature, which can improve the cover song audio recognition effect.
[0114] In one embodiment, such as Figure 4 The diagram illustrates a training method for another audio feature extraction model. In this embodiment, the method includes the following steps:
[0115] In step 401, a first original song fragment is obtained based on the first original song without tags. A target data augmentation method is randomly selected from candidate data augmentation methods, and the first original song fragment is augmented using the target data augmentation method to obtain a constructed song fragment. In step 402, the constructed song fragment is used as the positive fragment corresponding to the first original song fragment, and other song fragments are used as the negative fragments corresponding to the first original song fragment. In step 403, the first original song fragment, the corresponding positive fragment, and the corresponding negative fragment are input into the first audio feature extraction model to obtain a first audio feature extraction combination. In step 404, the first audio feature extraction combination is used for model training to obtain the trained first audio feature extraction model. In step 405, the trained first audio feature extraction model is used as an encoder. The encoder parameters are fixed, and a preset convolutional network and a fully connected layer are concatenated at the encoder output to obtain a second audio feature extraction model. In step 406, a second original song fragment and a cover song fragment are obtained based on the second original song with tags and its cover song. In step 407, the parameters of the second audio feature extraction model are adjusted based on the second original song fragment, the cover song fragment, and the second other song fragments until the second training termination condition is met, resulting in the trained second audio feature extraction model. It should be noted that the specific limitations of the above steps can be found in the specific limitations of a cover song audio recognition method described above, and will not be repeated here.
[0116] 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.
[0117] Based on the same inventive concept, this application also provides a training apparatus for an audio feature extraction model to implement the training method for the audio feature extraction model described above. The solution provided by this apparatus is similar to the implementation described in the above method. Therefore, the specific limitations of the one or more audio feature extraction model training apparatus embodiments provided below can be found in the limitations of the audio feature extraction model training method described above, and will not be repeated here.
[0118] In one exemplary embodiment, such as Figure 5 As shown, a training device for an audio feature extraction model is provided, comprising: a first audio feature extraction model training module, a second audio feature extraction model construction module, and a second audio feature extraction model training module, wherein:
[0119] The first audio feature extraction model training module 501 is used to train the first audio feature extraction model using a first original song fragment, a constructed song fragment, and a first other song fragment, to obtain the trained first audio feature extraction model; the constructed song fragment is obtained by data augmentation of the first original song fragment.
[0120] The second audio feature extraction model construction module 502 is used to construct the second audio feature extraction model by using the trained first audio feature extraction model as an encoder.
[0121] The second audio feature extraction model training module 503 is used to train the second audio feature extraction model using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments, to obtain the trained second audio feature extraction model; the trained second audio feature extraction model is used to identify songs from input cover audio fragments.
[0122] The first original song fragment is a song fragment from an unlabeled original song; the second original song fragment is a song fragment from a labeled original song, and the second original song fragment and the corresponding cover song fragment have a pre-labeled association relationship.
[0123] In one embodiment, the first audio feature extraction model training module 501 includes:
[0124] The first song fragment acquisition submodule is used to obtain the first original song fragment based on the first original song without a tag, and to generate the constructed song fragment based on the first original song fragment.
[0125] The positive and negative segment acquisition submodule is used to treat the constructed song segment as the positive segment corresponding to the first original song segment, and to treat the first other song segments as the negative segments corresponding to the first original song segment.
[0126] The feature extraction submodule is used to input the first original song fragment, the positive fragment corresponding to the first original song fragment, and the negative fragment corresponding to the first original song fragment into the first audio feature extraction model to obtain the first audio feature extraction combination.
[0127] The first trained model obtains a sub-module, which is used to train the model using the first audio feature extraction combination to obtain the trained first audio feature extraction model.
[0128] In one embodiment, the song fragment acquisition submodule includes:
[0129] The target data augmentation method determination unit is used to randomly select the target data augmentation method from the candidate data augmentation methods;
[0130] A song fragment construction unit is used to perform data augmentation on the first original song fragment using the target data augmentation method to obtain the constructed song fragment.
[0131] The candidate data augmentation methods include any one or more of the following:
[0132] Speed change, pitch shift, pitch offset, noise addition, random masking.
[0133] In one embodiment, the submodule obtained by the first trained model includes:
[0134] The target loss value determination unit is used to determine the target loss value based on the first audio feature extraction combination;
[0135] The model adjustment unit is used to adjust the first audio feature extraction model using the target loss value until the first training termination condition is met, thereby obtaining the trained first audio feature extraction model.
[0136] In one embodiment, the second audio feature extraction model construction module 502 includes:
[0137] The parameter fixing submodule is used to fix the parameters of the encoder by using the trained first audio feature extraction model as an encoder.
[0138] The splicing submodule is used to splice a preset convolutional network and a fully connected layer at the output of the encoder to obtain the second audio feature extraction model.
[0139] In one embodiment, the second audio feature extraction model training module 503 includes:
[0140] The second song fragment acquisition submodule is used to obtain the second original song fragment and the cover song fragment based on the second original song with tags and its cover song.
[0141] The model parameter adjustment submodule is used to adjust the parameters of the second audio feature extraction model based on the second original song fragment, the cover song fragment, and the second other song fragments, until the second training termination condition is met, thereby obtaining the trained second audio feature extraction model.
[0142] Each module in the training device of the aforementioned audio feature extraction model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.
[0143] Based on the same inventive concept, this application also provides a cover song audio recognition device for implementing the cover song audio recognition method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more cover song audio recognition device embodiments provided below can be found in the limitations of the cover song audio recognition method described above, and will not be repeated here.
[0144] In one exemplary embodiment, such as Figure 6 As shown, a cover song audio recognition device is provided, comprising: a cover song audio feature acquisition module, a target song audio feature determination module, and a song recognition result acquisition module, wherein:
[0145] The cover song audio feature acquisition module 601 is used to acquire the cover song audio segment to be identified, input the cover song audio segment into the trained second audio feature extraction model, and obtain the cover song audio features corresponding to the cover song audio segment.
[0146] The target song audio feature determination module 602 is used to determine the target song audio features that meet preset similarity conditions based on the feature similarity between the cover song audio features and multiple song audio features; the song audio features are obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model;
[0147] The song recognition result acquisition module 603 is used to take the songs in the music library corresponding to the audio features of the target song as the song recognition result of the cover audio segment;
[0148] The trained second audio feature extraction model is obtained by training a second original song fragment, a corresponding cover song fragment, and two other song fragments. The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model is obtained by training a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment is obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment from an unlabeled original song. The second original song fragment is a song fragment from a labeled original song, and there is a pre-labeled association between the second original song fragment and the corresponding cover song fragment.
[0149] In one embodiment, the target song audio feature determination module 602 includes:
[0150] The audio segment segmentation module is used to segment the songs in each music library into audio segments according to a preset duration, so as to obtain a set of audio segments for each song in the music library.
[0151] A sub-module for obtaining multiple song audio features is used to input each audio segment in the audio segment set into the trained second audio feature extraction model to obtain the multiple song audio features.
[0152] The feature similarity determination submodule is used to take the cosine distance between the cover song audio features and any song audio features as the feature similarity.
[0153] The target song audio feature acquisition submodule is used to select the song audio feature with the smallest distance from the cover song audio feature as the target song audio feature.
[0154] Each module in the aforementioned cover song audio recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0155] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7As 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 operating system and computer programs stored in the non-volatile storage media. The database stores training data for an audio feature extraction model. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a training method for an audio feature extraction model.
[0156] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces 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 interfaces. 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 interfaces are 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 cover song audio recognition method.
[0157] Those skilled in the art will understand that Figure 7 and Figure 8 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.
[0158] In one exemplary 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 perform the following steps:
[0159] The first audio feature extraction model is trained using a first original song fragment, a constructed song fragment, and first other song fragments to obtain the trained first audio feature extraction model; the constructed song fragment is obtained by data augmentation of the first original song fragment.
[0160] The trained first audio feature extraction model is used as an encoder to construct the second audio feature extraction model;
[0161] The second audio feature extraction model is trained using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments to obtain a trained second audio feature extraction model; the trained second audio feature extraction model is used to identify songs from input cover audio fragments.
[0162] The first original song fragment is a song fragment from an unlabeled original song; the second original song fragment is a song fragment from a labeled original song, and the second original song fragment and the corresponding cover song fragment have a pre-labeled association relationship.
[0163] In one embodiment, the processor, when executing the computer program, also implements the steps of the training method for the audio feature extraction model in the other embodiments described above.
[0164] In one exemplary 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 perform the following steps:
[0165] Obtain the cover song audio segment to be identified, input the cover song audio segment into the trained second audio feature extraction model, and obtain the cover song audio features corresponding to the cover song audio segment;
[0166] Based on the degree of feature similarity between the cover song audio features and multiple song audio features, target song audio features that meet preset similarity conditions are determined; the song audio features are obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model;
[0167] The songs in the music library corresponding to the audio features of the target song are used as the song recognition results of the cover audio segment;
[0168] The trained second audio feature extraction model is obtained by training a second original song fragment, a corresponding cover song fragment, and two other song fragments. The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model is obtained by training a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment is obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment from an unlabeled original song. The second original song fragment is a song fragment from a labeled original song, and there is a pre-labeled association between the second original song fragment and the corresponding cover song fragment.
[0169] In one embodiment, the processor, when executing a computer program, also implements the steps of the cover song audio recognition method in the other embodiments described above.
[0170] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0171] The first audio feature extraction model is trained using a first original song fragment, a constructed song fragment, and first other song fragments to obtain the trained first audio feature extraction model; the constructed song fragment is obtained by data augmentation of the first original song fragment.
[0172] The trained first audio feature extraction model is used as an encoder to construct the second audio feature extraction model;
[0173] The second audio feature extraction model is trained using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments to obtain a trained second audio feature extraction model; the trained second audio feature extraction model is used to identify songs from input cover audio fragments.
[0174] The first original song fragment is a song fragment from an unlabeled original song; the second original song fragment is a song fragment from a labeled original song, and the second original song fragment and the corresponding cover song fragment have a pre-labeled association relationship.
[0175] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the training method for the audio feature extraction model in the other embodiments described above.
[0176] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0177] Obtain the cover song audio segment to be identified, input the cover song audio segment into the trained second audio feature extraction model, and obtain the cover song audio features corresponding to the cover song audio segment;
[0178] Based on the degree of feature similarity between the cover song audio features and multiple song audio features, target song audio features that meet preset similarity conditions are determined; the song audio features are obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model;
[0179] The songs in the music library corresponding to the audio features of the target song are used as the song recognition results of the cover audio segment;
[0180] The trained second audio feature extraction model is obtained by training a second original song fragment, a corresponding cover song fragment, and two other song fragments. The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model is obtained by training a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment is obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment from an unlabeled original song. The second original song fragment is a song fragment from a labeled original song, and there is a pre-labeled association between the second original song fragment and the corresponding cover song fragment.
[0181] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the cover song audio recognition method in the other embodiments described above.
[0182] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:
[0183] The first audio feature extraction model is trained using a first original song fragment, a constructed song fragment, and first other song fragments to obtain the trained first audio feature extraction model; the constructed song fragment is obtained by data augmentation of the first original song fragment.
[0184] The trained first audio feature extraction model is used as an encoder to construct the second audio feature extraction model;
[0185] The second audio feature extraction model is trained using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments to obtain a trained second audio feature extraction model; the trained second audio feature extraction model is used to identify songs from input cover audio fragments.
[0186] The first original song fragment is a song fragment from an unlabeled original song; the second original song fragment is a song fragment from a labeled original song, and the second original song fragment and the corresponding cover song fragment have a pre-labeled association relationship.
[0187] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the training method for the audio feature extraction model in the other embodiments described above.
[0188] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:
[0189] Obtain the cover song audio segment to be identified, input the cover song audio segment into the trained second audio feature extraction model, and obtain the cover song audio features corresponding to the cover song audio segment;
[0190] Based on the degree of feature similarity between the cover song audio features and multiple song audio features, target song audio features that meet preset similarity conditions are determined; the song audio features are obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model;
[0191] The songs in the music library corresponding to the audio features of the target song are used as the song recognition results of the cover audio segment;
[0192] The trained second audio feature extraction model is obtained by training a second original song fragment, a corresponding cover song fragment, and two other song fragments. The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model is obtained by training a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment is obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment from an unlabeled original song. The second original song fragment is a song fragment from a labeled original song, and there is a pre-labeled association between the second original song fragment and the corresponding cover song fragment.
[0193] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the cover song audio recognition method in the other embodiments described above.
[0194] 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 the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.
[0195] Those skilled in the art will understand that all or part of the processes in the methods of 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, and when executed, it can include the processes of the embodiments of the above methods. 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.
[0196] 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.
[0197] 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 training method for an audio feature extraction model, characterized in that, The method includes: The first audio feature extraction model is trained using a first original song fragment, a constructed song fragment, and first other song fragments to obtain the trained first audio feature extraction model; the constructed song fragment is obtained by data augmentation of the first original song fragment. The trained first audio feature extraction model is used as an encoder to construct the second audio feature extraction model; The second audio feature extraction model is trained using a second original song fragment, a cover song fragment corresponding to the second original song fragment, and other song fragments to obtain a trained second audio feature extraction model; the trained second audio feature extraction model is used to identify songs from input cover audio fragments. The first original song fragment is a song fragment from an unlabeled original song; the second original song fragment is a song fragment from a labeled original song, and the second original song fragment and the corresponding cover song fragment have a pre-labeled association relationship.
2. The method according to claim 1, characterized in that, The process of training the first audio feature extraction model using a first original song fragment, a constructed song fragment, and first other song fragments to obtain the trained first audio feature extraction model includes: Based on the first original song without a tag, obtain a fragment of the first original song, and generate the constructed song fragment based on the fragment of the first original song; The constructed song fragment is taken as the positive fragment corresponding to the first original song fragment, and the first other song fragment is taken as the negative fragment corresponding to the first original song fragment; The first original song fragment, the positive fragment corresponding to the first original song fragment, and the negative fragment corresponding to the first original song fragment are input into the first audio feature extraction model to obtain the first audio feature extraction combination. The first audio feature extraction combination is used for model training to obtain the trained first audio feature extraction model.
3. The method according to claim 2, characterized in that, The step of generating the constructed song fragment based on the first original song fragment includes: The target data augmentation method is randomly selected from the candidate data augmentation methods; The first original song segment is augmented using the target data augmentation method to obtain the constructed song segment; The candidate data augmentation methods include any one or more of the following: Speed change, pitch shift, pitch offset, noise addition, random masking.
4. The method according to claim 2, characterized in that, The step of training the model using the first audio feature extraction combination to obtain the trained first audio feature extraction model includes: Based on the first audio feature extraction combination, determine the target loss value; The first audio feature extraction model is adjusted using the target loss value until the first training termination condition is met, thus obtaining the trained first audio feature extraction model.
5. The method according to claim 1, characterized in that, The step of using the trained first audio feature extraction model as an encoder to construct a second audio feature extraction model includes: The trained first audio feature extraction model is used as an encoder, and the parameters of the encoder are fixed. A preset convolutional network and a fully connected layer are concatenated at the output of the encoder to obtain the second audio feature extraction model.
6. The method according to claim 1, characterized in that, The second audio feature extraction model is trained using a second original song fragment, a corresponding cover song fragment, and other song fragments to obtain the trained second audio feature extraction model, including: Based on the second original song and its cover song carrying tags, obtain the second original song fragment and the cover song fragment; Based on the second original song fragment, the cover song fragment, and the second other song fragments, the parameters of the second audio feature extraction model are adjusted until the second training termination condition is met, thus obtaining the trained second audio feature extraction model.
7. A method for recognizing cover song audio, characterized in that, The method includes: Obtain the cover song audio segment to be identified, input the cover song audio segment into the trained second audio feature extraction model, and obtain the cover song audio features corresponding to the cover song audio segment; Based on the degree of feature similarity between the cover song audio features and multiple song audio features, target song audio features that meet preset similarity conditions are determined; the song audio features are obtained by inputting audio segments of songs in the music library into the trained second audio feature extraction model; The songs in the music library corresponding to the audio features of the target song are used as the song recognition results of the cover audio segment; The trained second audio feature extraction model is obtained by training a second original song fragment, a corresponding cover song fragment, and two other song fragments. The second audio feature extraction model is constructed from the trained first audio feature extraction model. The trained first audio feature extraction model is obtained by training a first original song fragment, a constructed song fragment, and a first other song fragment. The constructed song fragment is obtained by data augmentation of the first original song fragment. The first original song fragment is a song fragment from an unlabeled original song. The second original song fragment is a song fragment from a labeled original song, and there is a pre-labeled association between the second original song fragment and the corresponding cover song fragment.
8. The method according to claim 7, characterized in that, The step of determining the target song audio features that meet preset similarity conditions based on the feature similarity between the cover song audio features and multiple song audio features includes: The songs in each music library are segmented into audio segments according to a preset duration to obtain a set of audio segments for each of the aforementioned music libraries. Each audio segment in the audio segment set is input into the trained second audio feature extraction model to obtain the multiple song audio features; The cosine distance between the cover song audio features and any song audio features is used as the feature similarity. The song audio feature that is closest to the cover song audio feature is taken as the target song audio feature.
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 6, and / or the steps of the method according to any one of claims 7 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 6, and / or the steps of the method according to any one of claims 7 to 8.