Voiceprint recognition model generation, voiceprint recognition method and system

Abstract

This application proposes a method for generating a voiceprint recognition model, which includes: S1, performing initial pseudo-labeling on an unlabeled first audio data set U; S2, segmenting each audio data in U into multiple audio data segments of the same duration; inheriting the pseudo-labeling before segmentation to generate a second audio data set S3, which is fed into the voiceprint recognition model, and introducing a label correction strategy to train the voiceprint recognition model to obtain a third audio data set S4 with corrected labels, which is fed into the voiceprint recognition model for training; S5, extracting the representation vector of the voiceprint recognition model and feeding it into a clustering module to generate a second audio data set with updated labels. Repeat S3-S5 until the specified training rounds are completed. The above method uses a correction strategy to improve the accuracy of pseudo-labels and alleviate the impact of incorrect pseudo-labels on training; at the same time, a representation learning module is constructed based on the substructure (Sub-PTM) of the speech pre-training model to further improve the voiceprint recognition performance of the model.

Description

Technical Field

[0001] The present application belongs to the field of audio data processing technology, and specifically relates to a voiceprint recognition model generation, a voiceprint recognition method and a system. Background Art

[0002] Voiceprint recognition is a technology used to extract a speaker's voice features and automatically verify their identity. When a trained voiceprint recognition model is applied to a new domain, performance often degrades significantly. Because existing large-scale, precisely labeled datasets typically only cover mainstream languages, the aforementioned domain mismatch problem is even more common in voiceprint recognition for minority languages. Re-labeling a minority language dataset is costly, while unlabeled audio is readily available in large quantities for all languages. Therefore, leveraging large amounts of unlabeled speech data to build robust, unsupervised voiceprint recognition systems is a key development direction in this field.

[0003] The framework of an unsupervised voiceprint recognition system in the existing technology usually includes two stages: the initial speaker representation learning stage and the clustering-based iterative training stage. In the first stage, unlabeled audio is used to train an unsupervised model, which encodes speaker information into a representation vector. In the second stage, the speaker representation vector corresponding to each audio is annotated with a pseudo-label after passing through the clustering module, and the resulting pseudo-labeled dataset is used to train the voiceprint model. The trained voiceprint model learns a more speaker-discriminative representation vector, so it can replace the initial speaker representation vector for a new round of clustering. The above process is repeated many times until convergence. Among them, the final performance of the model depends largely on the quality of the pseudo-labels, and incorrect labels will bring negative effects.

[0004] To address the issue of pseudo-label correction, existing techniques have proposed methods for filtering pseudo-label data based on cluster confidence, thereby filtering out a certain proportion of unreliable pseudo-label data. However, if a class contains fewer samples, all samples clustered into that class will also be filtered out. Other techniques effectively mitigate the performance loss caused by incorrect labels by artificially setting a loss threshold in each round of training to screen out poor-quality pseudo-label data. Patent publication number CN114707668A introduces a Gaussian mixture model to model the distribution of loss values ​​for each round, setting the corresponding loss threshold accordingly.

[0005] However, the aforementioned solutions, as well as most existing work, mitigate the impact of erroneous pseudo-labels on voiceprint recognition model training by removing data. This requires optimal hyperparameter selection for each round, which is inflexible. Furthermore, simply removing data underutilizes some unlabeled audio training data, preventing the full value of pseudo-labels from being fully exploited.

[0006] In addition, supervised voiceprint recognition methods based on speech pre-training models have shown great potential, but in the absence of speaker labels, how to use speech pre-training models to improve supervised voiceprint recognition models remains to be studied. Summary of the Invention

[0007] In response to the above problems, the first aspect of this application proposes a method for generating a voiceprint recognition model, which includes:

[0008] S1, perform initial pseudo-labeling on the unlabeled first audio dataset U;

[0009] S2: Each audio data in the first audio data set U is segmented into multiple audio data segments of the same length; each audio data segment inherits the pseudo-label annotation before segmentation to generate a second audio data set

[0010] S3, the second audio dataset Feed it into the voiceprint recognition model and introduce the label correction strategy to train the voiceprint recognition model to obtain the third audio data set with corrected labels.

[0011] S4, the third audio dataset Send it to the voiceprint recognition model for training;

[0012] S5: Extract the speaker representation vector of the voiceprint recognition model and send it to the clustering module to generate the second audio dataset with updated labels. Repeat S3-S5 until the specified training round is completed.

[0013] Preferably, the initial pseudo-label annotation specifically includes: using the substructure of the speech pre-training model to extract the representation sequence of each audio in the audio data set U; performing a statistical pooling operation on the representation sequence to obtain a representation vector for each audio, which is regarded as the initial speaker representation vector; and sending it to the clustering module to obtain the initial pseudo-label.

[0014] Preferably, the voiceprint recognition model includes a representation learning module and a classification neural network module. The representation learning module is built based on the substructure of the speech pre-training model and is used to convert audio data into a multi-frame representation sequence; the classification neural network module is used to aggregate the multi-frame representation sequence into a representation vector to further realize speaker classification.

[0015] Preferably, the label correction strategy includes introducing the predicted labels of the voiceprint recognition model into the loss function for online label correction; and adopting a dynamic confidence strategy to gradually increase the confidence of the voiceprint recognition model in the predicted labels in the loss function during training. This method utilizes the model's inherent ability to predict incorrectly labeled samples as correctly classified for label correction.

[0016] Preferably, the label correction strategy includes performing online label correction by rescaling the posterior probability of reliable labels, wherein the reliable labels are pseudo labels that are consistent with the predicted labels of the voiceprint recognition model.

[0017] Preferably, the label correction strategy includes offline label correction, specifically including: counting the frequency of occurrence of the predicted label of the corresponding voiceprint recognition model in the L segments of audio data cut from a piece of audio data x annotated with a pseudo-label, where H is the number of audio data segments corresponding to the predicted label with the highest frequency of occurrence; if H / L>0.5, the pseudo-label of the L segments of audio data is updated to the predicted label with the highest frequency of occurrence; otherwise, the L segments of audio data are discarded.

[0018] Preferably, the clustering module is constructed using an algorithm based on graph clustering.

[0019] A second aspect of the present application provides a voiceprint recognition method, which performs voiceprint recognition using a voiceprint recognition model generated by any of the methods in the first aspect.

[0020] The third aspect of the present application provides a voiceprint recognition model generation system, comprising:

[0021] a pseudo-label generation module configured to perform initial pseudo-labeling on the unlabeled first audio data set U;

[0022] The audio segmentation module is configured to segment each audio data in the first audio data set U into multiple audio data segments of the same length; each audio data segment inherits the pseudo-label annotation before segmentation to generate a second audio data set

[0023] The first model training module is configured to train the second audio data set Feed it into the voiceprint recognition model and introduce the label correction strategy to train the voiceprint recognition model to obtain the third audio data set with corrected labels.

[0024] The second model training module is configured to train a third audio data set Send it to the voiceprint recognition model for training;

[0025] The loop training module is configured to extract the representation vector of the voiceprint recognition model and feed it into the clustering module to generate a second audio data set with updated labels. Loop until the specified number of training rounds is completed.

[0026] The fourth aspect of the present application proposes a voiceprint recognition system, which performs voiceprint recognition using the voiceprint recognition model generated by the system described in the third aspect.

[0027] The technical effects of the generation process of the voiceprint recognition model proposed in this application include: first, using online and offline label correction strategies to improve the accuracy of pseudo-labels and alleviate the impact of incorrect pseudo-labels on training. Compared with the existing methods such as setting thresholds, this label correction strategy is more flexible and can make greater use of all unlabeled audio for model training; second, using the substructure of the speech pre-training model (Sub-PTM) to extract the initial speaker representation vector of the unlabeled audio dataset; third, constructing a voiceprint recognition model based on Sub-PTM. The representation vector output by Sub-PTM contains richer phonetic and semantic information and can significantly improve the generalization of the trained voiceprint recognition model, thereby further improving the model's voiceprint recognition performance. In tasks such as voiceprint recognition and voiceprint verification, the voiceprint recognition model generated by the above method can be used to extract speaker representation vectors, and combined with algorithms such as cosine similarity, it can be used to determine whether two sentences come from the same speaker. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings help further understand the present application. The elements of the drawings are not necessarily in scale with each other. For ease of description, only the parts relevant to the invention are shown in the drawings.

[0029] Figure 1 This is a schematic diagram of the steps of a method for generating a voiceprint recognition model in one embodiment of the present application;

[0030] Figure 2 This is a schematic diagram of the algorithm flow of the voiceprint recognition model in another embodiment of the present application;

[0031] Figure 3 This is a structural diagram of a voiceprint recognition model generation system in another embodiment of the present application. DETAILED DESCRIPTION

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention.

[0033] Figure 1 The figure is a schematic diagram of the steps of a method for generating a voiceprint recognition model in one embodiment of the present invention. The method specifically includes:

[0034] S1: Perform initial pseudo-labeling on an unlabeled first audio dataset U.

[0035] The initial audio data set U is unlabeled. This step is used to obtain the initial pseudo-labels of the unlabeled audio data set U, which can be implemented using various existing unsupervised voiceprint recognition models. In a preferred embodiment, a method based on a substructure of a speech pre-training model (Sub-PTM) is adopted. The speech pre-training model (Pre-trained Models, PTM) can learn high-level speech representations from a large amount of unlabeled data, and the middle layer of its model contains speaker-discriminative information. PTM is an open source and available generalization model. This embodiment uses it to further construct Sub-PTM, which consists of a convolutional network module and several Transformer modules. The original audio is converted into a multi-frame representation sequence after Sub-PTM. Subsequently, the representation sequence is statistically pooled to obtain the initial speaker representation vector for each audio; it is sent to the clustering module to obtain the initial pseudo-label.

[0036] In a further preferred embodiment, the clustering module employs the Infomap algorithm based on graph clustering. Clustering methods commonly used in clustering modules in existing unsupervised voiceprint recognition models include K-Means and AHC. K-Means optimizes to minimize the distance between samples and class centers. However, clustering performance deteriorates when the dataset is complex, such as when the classes are imbalanced or the dataset has a non-convex distribution. Furthermore, K-Means requires specifying the number of classes, K, and different K values ​​significantly impact the algorithm's results. AHC combines the two most similar samples or classes in the dataset, repeatedly iterating this process to create a cluster tree from the bottom up. However, AHC also requires a predefined number of classes or a threshold to stop merging. Each iteration requires calculating pairwise distances and can only merge two subclasses, resulting in high computational complexity and slow clustering when applied to large datasets. The clustering module in this embodiment employs the Infomap algorithm based on graph clustering. This algorithm combines class and information encoding through a two-layer encoding scheme. Its graph-based construction enables it to handle complex relationships and is suitable for clustering large datasets. In addition, this clustering algorithm does not require setting hyperparameters, making it more reasonable. Specifically, the speaker category name encoding and the intra-category speaker encoding are used as a double-layer encoding, the speaker representation vectors are used as nodes, and the cosine similarity between vectors is used as the edge weight value. The average encoding length Len is used to achieve a better clustering result:

[0037]

[0038] in, Indicates the proportion of category name codes in all codes, represents the proportion of the code of the node belonging to category i in all codes, E represents the average code length required to encode the category name, and E i=(\frac{\text{Len}}} represents the average encoding length required to encode all nodes of category i. Calculating the average encoding length is essentially calculating entropy. The clustering steps include: First, initialize each node as an independent class; second, perform a random walk on the constructed speaker representation graph and generate a sequence, attempting to assign each node to the class of the adjacent node that causes the largest decrease in Len; repeat the second step until Len no longer decreases.

[0039] S2: Each audio data in the first audio data set U is segmented into multiple audio data segments of the same length; each audio data segment inherits the pseudo-label annotation before segmentation to generate a second audio data set with labels

[0040] S3, the second audio dataset Feed it into the voiceprint recognition model and introduce the label correction strategy to train the voiceprint recognition model to obtain the third audio data set with corrected labels. This audio dataset also contains labels.

[0041] In a preferred embodiment, the voiceprint recognition model includes a representation learning module and a classification neural network module. The representation learning module can also be built based on the substructure (Sub-PTM) of the speech pre-training model to convert audio data into a multi-frame representation sequence; the classification neural network module is used to aggregate the multi-frame representation sequence into a representation vector. This part can be constructed using some mainstream voiceprint models, such as ECAPA-TDNN.

[0042] The initial pseudo-labels obtained in S1 inevitably contain erroneous pseudo-labels, i.e., noisy labels, which will limit the performance of the model. Therefore, the present invention designs multiple label correction strategies to perform label correction.

[0043] In a preferred embodiment, the label correction strategy includes designing an LNC loss function (Label NoiseCorrection Loss) and introducing the labels predicted by the voiceprint recognition model into the loss function. During the training process of the neural network, the model will learn the intrinsic feature distribution of the data set and then be forced to memorize the noise label samples. Before being seriously affected by the noise labels, the model can learn in the right direction, that is, the model has the ability to re-predict the wrong labels of the samples as correct labels. Based on this feature, this embodiment introduces the pseudo labels predicted by the voiceprint recognition model into the loss function for training. The loss function L LNC Use the following expression:

[0044]

[0045] in, B is the size of the mini-batch in model training, T is the number of training steps in each round, represents the pseudo label of the i-th sample, Indicates that the i-th sample is recognized by the model as The posterior probability of is the predicted label of the model corresponding to the input i-th sample, α t and β t Represents the confidence of the model for the pseudo label and its own predicted label, satisfying α t +β t =1. α t and β t can be simply set to a fixed value, but if β t If it is set too large, it means that the model will over-trust its own prediction ability, which is unreasonable in the early stages of training. t If it is set too small, the training process will still be dominated by pseudo labels, resulting in limited label noise suppression capabilities of the model. t The value of should be able to reasonably balance the influence of pseudo labels and predicted labels. In this embodiment, it is preferred to adopt a dynamic confidence strategy, that is, β t It will gradually increase as the training progresses, and finally reach β after T iterations T , which means that the model is increasingly confident in its own predictive ability. From another perspective, this strategy is equivalent to updating soft false labels in real time, because the model has two labels to fit. As the model is trained, the possibility that the predicted label is correct becomes greater and greater. The increase in its weight value achieves label correction to a certain extent.

[0046] In another preferred embodiment, the label correction strategy includes rescaling the posterior probability of reliable labels. When the pseudo label of a sample is consistent with the model predicted label, it will be considered a reliable label. For pseudo-label datasets, these reliable samples are particularly important. They are the basis for the model to optimize in the right direction. In order to further enable the model to fully utilize the reliable labels, during the training process, in order to increase the contribution of the reliable labels in the reverse gradient propagation, the posterior probability of the reliable labels is rescaled according to:

[0047]

[0048] The above formula is a variant of the posterior probability of AM-Softmax, where s and m are the parameters of AM-Softmax, representing the angle penalty factor and scale factor respectively. The speaker representation vector and W of the i-th sample output by the representative model j The angle, W j represents the j-th center vector of the fully connected layer matrix W, M represents the total number of categories covered by the pseudo-label, and d represents the output dimension of the fully connected layer. Let r ≥ 0, then therefore This means that reliable samples contribute more to the update of model parameters.

[0049] In another preferred embodiment, the label correction strategy includes offline label correction. The voiceprint recognition model trained using the LNC loss function has a certain robustness to label noise and can be used for offline label correction. The offline label correction strategy is proposed based on the label consistency before and after segmentation. Specifically, after all the segmented segment-level audios are marked with predicted speaker labels after passing through the voiceprint recognition model, it is assumed that there are L segment-level audios corresponding to a certain complete audio, and there are H predicted labels with the highest frequency in these L audios. If H / L>0.5, the predicted label will be regarded as the dominant label of this complete audio. Then the pseudo-labels of the segment-level audios corresponding to this complete audio will be updated to the dominant labels. If a complete audio does not have a dominant label, all the corresponding segment-level audios will be discarded.

[0050] S4, audio dataset The pseudo-labeled dataset after correction and filtering will be used to retrain the voiceprint recognition model. Since the accuracy of the pseudo-labels is greatly improved at this time, the retraining only uses the traditional loss function.

[0051] S5, extract the speaker representation vector of the voiceprint recognition model and send it to the clustering module to generate an audio dataset with updated labels Repeat S3-S5 until the specified training round is completed.

[0052] The output of the last fully connected layer of the voiceprint recognition model is used as the speaker representation vector and clustered. The clustering module here can also be constructed based on the Infomap algorithm of graph clustering in the aforementioned preferred embodiment.

[0053] Figure 2 This is a schematic diagram of the algorithm flow of the voiceprint recognition model in another embodiment of the present application. Specifically, the voiceprint recognition model It includes a representation learning module and a classification neural network module. The representation learning module contains a substructure of the speech pre-training model (Sub-PTM) M θ , the top layer of Sub-PTM is connected with a classification neural network M ζ ,refer to Figure 2 , the algorithm flow of this model specifically includes:

[0054] S201, the unlabeled audio dataset is denoted as There are N audios in total; the substructure (Sub-PTM) of the speech pre-training model is used θ Extract the representation sequence of each audio in U, and obtain the representation vector e of each audio by statistical pooling these representation sequences i , which is used as the initial speaker representation vector;

[0055] S202, in Perform clustering operations on the dataset to obtain pseudo label sets

[0056] S203, cut U into clip-level audio sets There are K fragment-level audios in total, according to U * Annotate and obtain a pseudo-label set at the segment level And pseudo-labeled audio datasets The audio duration of each segment is the same, for example, 2s or 5s. If the last segment does not meet the duration requirement, zero padding or repeating the operation can be performed as appropriate.

[0057] S204, use M θ Parameters to initialize the voiceprint recognition model The representation learning module, and the classification neural network module M ζ The network parameters are randomly initialized and then annotated with pseudo-labeled audio datasets Come to training An online label correction strategy is introduced to improve the training effect by improving the loss function and reducing the negative impact of label noise on training;

[0058] S205, after training Used to predict U * Speaker labels, the predicted label set is Introducing offline label correction strategies to filter and correct Obtaining a label-corrected audio dataset

[0059] S206: Reinitialize the voiceprint recognition model according to the method of S204. Then use the audio dataset To retrain At this time, the quality of the pseudo-label is high, so the traditional AM-Softmax is used for model training;

[0060] S207, extraction The speaker representation vector E is generated from the output of the last fully connected layer, and S202-S207 are repeated until the set number of training rounds is reached.

[0061] In voiceprint recognition, voiceprint verification and other tasks, the trained voiceprint recognition model It can be used to extract the final speaker representation vector and, combined with algorithms such as cosine similarity, determine whether two sentences come from the same speaker.

[0062] In a specific embodiment, a two-stage fine-tuning method is used to train the voiceprint recognition model. The total number of cycles in each round of training is N epoch , it means that the model will have seen N completely during the entire training process epoch Secondary dataset, in the first 2 / 3·N epoch In the training, only the parameters of the classification neural network will be updated by back propagation. The learning rate parameter is lr. epoch During the training, all parameters of Sub-PTM and classification neural network will be optimized and updated with a learning rate of 1 / 2·lr.

[0063] In another preferred embodiment, data augmentation is combined with training to improve model generalization. This augmentation can be performed offline by pre-caching a multiple-fold dataset, or by performing online data augmentation during training. Online data augmentation offers advantages in increasing data diversity. Common methods include noise addition, reverberation, variable-speed perturbation, and spectral enhancement. When applying online data augmentation, each augmentation method can be assigned a certain probability, while also ensuring a certain probability of retaining the original data.

[0064] The voiceprint recognition models generated in the above-mentioned multiple embodiments use an online label correction strategy to adjust the confidence of pseudo-labels and model-predicted labels in real time during the training process, greatly reducing the impact of label noise on the model. An offline label correction strategy is used to perform label correction based on the consistency of sentence-level and segment-level labels, further improving the quality of pseudo-labels. The model framework that organically combines the two can effectively complete unsupervised voiceprint recognition tasks, improve the label noise problem existing in pseudo-label datasets in existing technologies, and effectively improve the quality of pseudo-labels and model performance. At the same time, the speech pre-training model is used for unsupervised speaker recognition. The speech pre-training model does not rely on the speaker negative example hypothesis and can be directly used to extract the initial speaker representation. Furthermore, the voiceprint recognition system built based on the speech pre-training model has better generalization ability and can further improve voiceprint recognition performance.

[0065] Figure 3 FIG. 3 is a schematic diagram of the structure of a voiceprint recognition model generation system 300 in another embodiment of the present application, which specifically includes:

[0066] A pseudo-label generation module 301 is configured to perform initial pseudo-labeling on an unlabeled first audio data set U;

[0067] The audio segmentation module 302 is configured to segment each audio data in the first audio data set U into multiple audio data segments of the same duration; each audio data segment inherits the pseudo-label annotation before segmentation to generate a second audio data set

[0068] The first model training module 303 is configured to train the second audio data set Feed it into the voiceprint recognition model and introduce the label correction strategy to train the voiceprint recognition model to obtain the third audio data set with corrected labels.

[0069] The second model training module 304 is configured to train the third audio data set Send it to the voiceprint recognition model for training;

[0070] The loop training module 305 is configured to extract the representation vector of the voiceprint recognition model and send it to the clustering module to generate a second audio data set with updated labels. Loop until the specified number of training rounds is completed.

[0071] Although the contents of the present application are specifically shown and described in conjunction with the preferred embodiments, those skilled in the art should understand that various changes made to the present application in form and details without departing from the spirit and scope of the present application as defined by the appended claims and without making any creative work are within the scope of protection of the present application.

Claims

1. A method for generating a voiceprint recognition model, characterized in that: include: S1, perform initial pseudo-labeling on the unlabeled first audio dataset U; S2: Segment each audio data in the first audio data set U into multiple audio data segments of the same duration; each audio data segment inherits the pseudo-label annotation before segmentation to generate a second audio data set ; S3, the second audio data set The voiceprint recognition model is fed into the dataset and a label correction strategy is introduced to train the voiceprint recognition model to obtain a third audio dataset with corrected labels. ; S4, the third audio data set Send it to the voiceprint recognition model for training; S5: Extract the speaker representation vector of the voiceprint recognition model and send it to the clustering module to generate the second audio dataset with updated labels. Repeat S3-S5 until the specified training round is completed.

2. The method for generating a voiceprint recognition model according to claim 1, wherein: The initial pseudo-label annotation specifically includes: The substructure of the speech pre-training model is used to extract the representation sequence of each audio in the audio dataset U; Performing a statistical pooling operation on the representation sequence to obtain a representation vector for each audio; The representation vector is sent to the clustering module to obtain the initial pseudo label.

3. The method for generating a voiceprint recognition model according to claim 1, wherein: The voiceprint recognition model includes a representation learning module and a classification neural network module. The representation learning module is built based on the substructure of the speech pre-training model and is used to convert audio data into a multi-frame representation sequence; the classification neural network module is used to aggregate the multi-frame representation sequence into a representation vector.

4. The method for generating a voiceprint recognition model according to claim 1, wherein: The label correction strategy includes introducing the predicted label of the voiceprint recognition model into the loss function to perform online label correction; and adopting a dynamic confidence strategy to gradually increase the confidence of the voiceprint recognition model in the predicted label in the loss function during the training process.

5. The method for generating a voiceprint recognition model according to claim 1, wherein: The label correction strategy includes performing online label correction by rescaling the posterior probability of a reliable label, wherein the reliable label is a pseudo label consistent with the predicted label of the voiceprint recognition model.

6. The method for generating a voiceprint recognition model according to claim 1, wherein: The label correction strategy includes offline label correction, specifically including: counting the frequency of occurrence of the corresponding predicted label of the voiceprint recognition model in the L segments of audio data cut from a piece of audio data x annotated with a pseudo-label, where H is the number of audio data segments corresponding to the predicted label with the highest frequency of occurrence; if H / L>0.5, then updating the pseudo-label of the L segments of audio data to the predicted label with the highest frequency of occurrence; otherwise, discarding the L segments of audio data.

7. The method for generating a voiceprint recognition model according to claim 1 or 2, wherein: The clustering module is constructed using an algorithm based on graph clustering.

8. A voiceprint recognition method, characterized in that: Voiceprint recognition is performed using a voiceprint recognition model generated by the method according to any one of claims 1 to 7.

9. A voiceprint recognition model generation system, characterized in that: include: A pseudo-label generation module is configured to perform initial pseudo-labeling on the unlabeled audio dataset U; The audio segmentation module is configured to segment each audio data in the audio data set U into multiple audio data segments of the same length; each audio data segment inherits the pseudo label annotation before segmentation to generate an audio data set ; The first model training module is configured to train the audio data set Input the voiceprint recognition model and introduce the label correction strategy to train the voiceprint recognition model to obtain the audio data set with label correction ; The second model training module is configured to train the audio data set Send it to the voiceprint recognition model for training; The loop training module is configured to extract the representation vector of the voiceprint recognition model and feed it into the clustering module to generate an audio dataset with updated labels. , loop until the specified training rounds are completed.

10. A voiceprint recognition system, characterized in that: Voiceprint recognition is performed using the voiceprint recognition model generated by the system described in claim 9.

Images