Audio data processing method, device and equipment and readable storage medium

By segmenting and re-labeling the voiceprint recognition dataset and using compactness variation values ​​to filter registered audio, the problem of poor voiceprint recognition performance under high noise conditions was solved, thus achieving both dataset integrity and improved recognition performance.

CN119028358APending Publication Date: 2024-11-26CHINA MOBILE COMM LTD RES INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310609617.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-26
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing voiceprint recognition technologies suffer from significant data loss under high noise conditions. Explicit processing methods result in severe data gaps, while implicit processing methods struggle to design suitable loss functions, thus affecting voiceprint recognition performance.

Method used

By segmenting and relabeling the voiceprint recognition dataset, the registered audio is filtered using the compactness variation value of the voiceprint feature vector, and label noise is filtered out to form an audio set without label noise, and the voiceprint recognition model is trained.

Benefits of technology

It effectively filters label noise, preserves the integrity of the dataset, and improves the accuracy and stability of voiceprint recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028358A_ABST
    Figure CN119028358A_ABST
Patent Text Reader

Abstract

The invention discloses an audio data processing method, device and equipment and a readable storage medium, relates to the technical field of communication, and aims to ensure a voiceprint recognition effect. The method comprises the steps that a voiceprint recognition data set with a label is acquired, the voiceprint recognition data set comprises a first audio set and a second audio set, the first audio set comprises the audio with the longest duration of each user, and the second audio set comprises the audios, except the first audio set, in the voiceprint recognition data set; segmenting audios in the first audio set to obtain a third audio set; based on the third audio set, obtaining a voiceprint feature vector set of the voiceprint recognition data set; determining a registered audio based on the voiceprint feature vector set and the first audio set; and based on the registered audio, the audio in the second audio set is marked again to obtain a fourth audio set, and the fourth audio set does not comprise label noise. According to the embodiment of the invention, the voiceprint recognition effect can be ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to an audio data processing method, device, equipment and readable storage medium. Background Art

[0002] Voiceprint recognition refers to extracting voiceprint features from the user's voice signal and verifying and identifying the user's identity. Currently, voiceprint recognition models based on deep learning have become mainstream. Developing voiceprint recognition models based on deep learning requires a large amount of labeled data. However, in reality, the problem of incorrect labeling of training data is inevitable.

[0003] Noise data is usually divided into two categories: feature noise and label noise. Feature noise refers to the deviation between the training sample instance feature itself and its true feature, such as artificially adding Gaussian noise to existing features. Label noise refers to the deviation between the target label used for training and the true label of the corresponding instance itself, such as incorrectly setting the category to which the label belongs when setting the label. Both types of noise will cause the performance of the system to degrade, but label noise is more harmful. For example, in voiceprint recognition, the audio of multiple different users is labeled with the same label (N:1) or the audio of the same user is labeled with different labels (1:N), which will cause the neural network to learn the wrong voiceprint features, affecting the final voiceprint recognition effect.

[0004] At present, there are two main methods for dealing with label noise: explicit processing and implicit processing. Explicit processing is to model the noise in some way and manipulate the input stream of the model. Among them, cleaning label noise and eliminating or correcting incorrectly labeled examples in training data is the most commonly used method. This step can be completed before training or simultaneously with the main model training. Implicit processing is not to explicitly model the noise, but to design a more general algorithm to eliminate the negative impact of noise. The use of inherent noise tolerance methods can achieve noise robustness.

[0005] However, under high noise conditions, explicit processing methods usually cause serious data loss, which ultimately leads to a decrease in the performance verification of the algorithm. The implicit processing method needs to design a suitable loss function according to different application environments. The main problem with this method is that it is difficult to design a suitable loss function under conditions of high noise levels.

[0006] Therefore, it is necessary to design an audio data processing solution that can ensure the voiceprint recognition effect. Summary of the invention

[0007] The embodiments of the present application provide an audio data processing method, apparatus, device and readable storage medium to ensure the voiceprint recognition effect.

[0008] In a first aspect, an embodiment of the present application provides an audio data processing method, comprising:

[0009] Acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set;

[0010] Segmenting the audio in the first audio set to obtain a third audio set;

[0011] Based on the third audio set, obtaining a voiceprint feature vector set of the voiceprint recognition data set;

[0012] Determine the registration audio based on the voiceprint feature vector set and the first audio set;

[0013] Based on the registered audio, the audio in the second audio set is re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

[0014] Optionally, dividing the audio in the first audio set to obtain a third audio set includes:

[0015] Each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set, wherein the predetermined time length is greater than or equal to 1.5 seconds and less than or equal to 2.5 seconds.

[0016] Optionally, obtaining a voiceprint feature vector set of the voiceprint recognition data set based on the third audio set includes:

[0017] Performing model training based on the third audio set to obtain a first voiceprint recognition model;

[0018] Extracting a voiceprint feature vector of the audio in the voiceprint recognition data set using the first voiceprint recognition model;

[0019] The voiceprint feature vector set is formed by using the voiceprint feature vector.

[0020] Optionally, determining the registration audio based on the voiceprint feature vector set and the first audio set includes:

[0021] S1. For each first audio in the first audio set, calculate a compactness change value of the voiceprint feature vector of the first audio;

[0022] S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio; if the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set;

[0023] S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

[0024] Optionally, for each first audio in the first audio set, calculating the compactness change value of the voiceprint feature vector of the first audio includes:

[0025] For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner:

[0026] DeltaC_vi=Ci_pre-Ci_post

[0027] Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

[0028] Optionally, the compactness corresponding to the vi is determined according to the following formula:

[0029]

[0030] Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0031] Optionally, the re-labeling the audio in the second audio set based on the registered audio to obtain a fourth audio set includes:

[0032] Performing user registration on the audio of each user in the finally obtained first audio set to obtain a first registration vector of each user;

[0033] The audio in the second audio set is re-labeled based on the first registration vector to obtain a fourth audio set.

[0034] Optionally, performing user registration on the audio of each user in the finally obtained first audio set to obtain a first registration vector of each user includes:

[0035] Obtaining a mean vector of all voiceprint feature vectors in the voiceprint feature vector set;

[0036] Obtaining a mean value of the voiceprint feature vector of the audio of each user in the third audio set;

[0037] The difference between the mean of the voiceprint feature vector of each user's audio and the mean vector is used as the first registration vector.

[0038] Optionally, the re-labeling the audio in the second audio set based on the first registration vector to obtain a fourth audio set includes:

[0039] For a second audio in the second audio set, performing voiceprint recognition on the second audio and a first registration vector of a user corresponding to the label of the second audio;

[0040] If the second audio matches the first registration vector of the user corresponding to the label of the second audio, it indicates that the labeling of the second audio is accurate, and the second audio is added to the first audio set and deleted from the second audio set;

[0041] If the second audio does not match the first registration vector of the user corresponding to the label of the second audio, matching the second audio with audios of other users to obtain audio of the target user matching the second audio;

[0042] Based on the second audio and the audio of the target user, obtaining a compactness change value corresponding to the second audio;

[0043] If the compactness change value corresponding to the second audio is greater than 0, the second audio is marked with the label of the target user, and the second audio is added to the first audio set and deleted from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, the second audio is retained in the second audio set.

[0044] Optionally, the method further includes:

[0045] Using the obtained first audio set to train a second voiceprint recognition model;

[0046] The method of claim 8 is performed using the second voiceprint recognition model until the number of audios in the first audio set remains unchanged.

[0047] In a second aspect, an embodiment of the present application provides an audio data processing device, including:

[0048] A first acquisition module is used to acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set;

[0049] A second acquisition module, configured to segment the audio in the first audio set to obtain a third audio set;

[0050] A third acquisition module, configured to obtain a voiceprint feature vector set of the voiceprint recognition data set based on the third audio set;

[0051] A first determining module, configured to determine the registration audio based on the voiceprint feature vector set and the first audio set;

[0052] The fourth acquisition module is used to re-label the audio in the second audio set based on the registered audio to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

[0053] Optionally, the second acquisition module is further used for:

[0054] Each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set, wherein the predetermined time length is greater than or equal to 1.5 seconds and less than or equal to 2.5 seconds.

[0055] Optionally, the third acquisition module includes:

[0056] A first training submodule, configured to perform model training based on the third audio set to obtain a first voiceprint recognition model;

[0057] A first extraction submodule, configured to extract a voiceprint feature vector of the audio in the voiceprint recognition data set by using the first voiceprint recognition model;

[0058] The first acquisition submodule is used to form the voiceprint feature vector set using the voiceprint feature vector.

[0059] Optionally, the fourth acquisition module is further used to:

[0060] S1. For each first audio in the first audio set, calculate a compactness change value of the voiceprint feature vector of the first audio;

[0061] S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio; if the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set;

[0062] S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

[0063] Optionally, the fourth acquisition module is further used to:

[0064] For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner:

[0065] DeltaC_vi=Ci_pre-Ci_post

[0066] Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

[0067] Optionally, the compactness corresponding to the vi is determined according to the following formula:

[0068]

[0069] Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0070] Optionally, the fourth acquisition module includes:

[0071] A first registration submodule, configured to register the audio of each user in the finally obtained first audio set to obtain a first registration vector for each user;

[0072] The first acquisition submodule is configured to re-label the audio in the second audio set based on the first registration vector to obtain a fourth audio set.

[0073] Optionally, the first registration submodule includes:

[0074] A first acquisition unit, configured to acquire a mean vector of all voiceprint feature vectors in the voiceprint feature vector set;

[0075] A second acquisition unit, configured to acquire a mean value of a voiceprint feature vector of the audio of each user in the third audio set;

[0076] The third acquisition unit is used to use the difference between the mean of the voiceprint feature vector of each user's audio and the mean vector as the first registration vector.

[0077] Optionally, the first acquisition submodule includes:

[0078] A first processing unit is configured to perform voiceprint recognition on a second audio in the second audio set by using the second audio and a first registration vector of a user corresponding to a label of the second audio;

[0079] a second processing unit, configured to, if the second audio matches the first registration vector of the user corresponding to the label of the second audio, indicate that the labeling of the second audio is accurate, and add the second audio to the first audio set and delete it from the second audio set;

[0080] a third processing unit, configured to, if the second audio does not match the first registration vector of the user corresponding to the label of the second audio, match the second audio with audios of other users to obtain audio of a target user matching the second audio;

[0081] A first acquisition unit, configured to acquire a compactness change value corresponding to the second audio based on the second audio and the audio of the target user;

[0082] The fourth processing unit, if the compactness change value corresponding to the second audio is greater than 0, marks the second audio with the label of the target user, adds the second audio to the first audio set, and deletes it from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, retains the second audio in the second audio set.

[0083] Optionally, the device further comprises:

[0084] A training module, used for training a second voiceprint recognition model using the obtained first audio set;

[0085] The aforementioned method is performed using the second voiceprint recognition model until the number of audios in the first audio set remains unchanged.

[0086] In a third aspect, an embodiment of the present application provides an audio data processing device, including: a processor and a transceiver;

[0087] Wherein, the processor is used for:

[0088] Acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set;

[0089] Segmenting the audio in the first audio set to obtain a third audio set;

[0090] Based on the third audio set, obtaining a voiceprint feature vector set of the voiceprint recognition data set;

[0091] Determine the registration audio based on the voiceprint feature vector set and the first audio set;

[0092] Based on the registered audio, the audio in the second audio set is re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

[0093] Optionally, the processor is further configured to:

[0094] Each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set, wherein the predetermined time length is greater than or equal to 1.5 seconds and less than or equal to 2.5 seconds.

[0095] Optionally, the processor is further configured to:

[0096] Performing model training based on the third audio set to obtain a first voiceprint recognition model;

[0097] Extracting a voiceprint feature vector of the audio in the voiceprint recognition data set using the first voiceprint recognition model;

[0098] The voiceprint feature vector set is formed by using the voiceprint feature vector.

[0099] Optionally, the processor is further configured to:

[0100] S1. For each first audio in the first audio set, calculate a compactness change value of the voiceprint feature vector of the first audio;

[0101] S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio; if the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set;

[0102] S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

[0103] Optionally, the processor is further configured to:

[0104] For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner:

[0105] DeltaC_vi=Ci_pre-Ci_post

[0106] Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

[0107] Optionally, the compactness corresponding to the vi is determined according to the following formula:

[0108]

[0109] Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0110] Optionally, the processor is further configured to:

[0111] Performing user registration on the audio of each user in the finally obtained first audio set to obtain a first registration vector of each user;

[0112] The audio in the second audio set is re-labeled based on the first registration vector to obtain a fourth audio set.

[0113] Optionally, the processor is further configured to:

[0114] Obtaining a mean vector of all voiceprint feature vectors in the voiceprint feature vector set;

[0115] Obtaining a mean value of the voiceprint feature vector of the audio of each user in the third audio set;

[0116] The difference between the mean of the voiceprint feature vector of each user's audio and the mean vector is used as the first registration vector.

[0117] Optionally, the processor is further configured to:

[0118] For a second audio in the second audio set, performing voiceprint recognition on the second audio and a first registration vector of a user corresponding to the label of the second audio;

[0119] If the second audio matches the first registration vector of the user corresponding to the label of the second audio, it indicates that the labeling of the second audio is accurate, and the second audio is added to the first audio set and deleted from the second audio set;

[0120] If the second audio does not match the first registration vector of the user corresponding to the label of the second audio, matching the second audio with audios of other users to obtain audio of the target user matching the second audio;

[0121] Based on the second audio and the audio of the target user, obtaining a compactness change value corresponding to the second audio;

[0122] If the compactness change value corresponding to the second audio is greater than 0, the second audio is marked with the label of the target user, and the second audio is added to the first audio set and deleted from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, the second audio is retained in the second audio set.

[0123] Optionally, the processor is further configured to:

[0124] Using the obtained first audio set to train a second voiceprint recognition model;

[0125] The aforementioned method is performed using the second voiceprint recognition model until the number of audios in the first audio set remains unchanged.

[0126] In a fourth aspect, an embodiment of the present application further provides a communication device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor implements the steps in the audio data processing method as described above when executing the program.

[0127] In a fifth aspect, an embodiment of the present application further provides a readable storage medium, on which a program is stored, and when the program is executed by a processor, the steps in the audio data processing method as described above are implemented.

[0128] In the embodiment of the present application, the data in the first audio set in the voiceprint recognition data set is segmented to obtain the third audio data, and based on the third audio set, the voiceprint feature vector set and the registration audio of the voiceprint recognition data set are obtained. Therefore, based on the registration audio, the audio in the second audio set can be re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise. Therefore, in the embodiment of the present application, through a series of processes such as segmentation, recognition, and re-labeling, the label noise of the voiceprint recognition data set can be filtered, the integrity of the data set is retained, and the voiceprint recognition effect is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0129] Figure 1 It is one of the flowcharts of the audio data processing method provided in the embodiment of the present application;

[0130] Figure 2 is a schematic diagram of the audio data processing process provided by an embodiment of the present application;

[0131] Figure 3 is one of the structural diagrams of the audio data processing device provided in the embodiment of the present application;

[0132] Figure 4 This is the second structural diagram of the audio data processing device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0133] In the embodiments of the present application, the term "and / or" describes the association relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0134] In the embodiments of the present application, the term "plurality" refers to two or more than two, and other quantifiers are similar.

[0135] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0136] See also Figure 1 , Figure 1It is a flowchart of the audio data processing method provided by an embodiment of the present application. As Figure 1 shown, it includes the following steps:

[0137] Step 101: Obtain a voiceprint recognition dataset with annotations. Among them, the voiceprint recognition dataset includes a first audio set and a second audio set. The first audio set includes the audio with the longest duration for each user, and the second audio set includes the audio in the voiceprint recognition dataset except the first audio set.

[0138] Among them, the voiceprint recognition dataset D may include the audio of M users (labels are s1, s2......sM), the number of audio for each user is (n1, n2......nM) respectively, and the total number of audio is N. For a certain user mi, it contains ni audio (u1, u2......uni), and there may be j (0≤j<ni) audio that do not belong to user mi, that is, for user mi, among the ni audio that belong to him, there are j audio whose users are not mi.

[0139] In the embodiment of the present application, the following operations are performed on M users in sequence: For the i-th (1≤i≤M) user si, select the audio u_si_max with the longest duration from the ni audio belonging to si. Combine the u_si_max of each user among the M users to obtain an audio set D_max of size M (u_s1_max, u_s2_max......u_sM_max) (that is, the first audio set). After removing D_max from the voiceprint recognition dataset D, the remaining audio forms a set D_max- (that is, the second audio set).

[0140] Step 102: Segment the audio in the first audio set to obtain a third audio set.

[0141] In the embodiment of the present application, each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set D_max_seg, where the value of the predetermined time length is greater than or equal to 1.5s and less than or equal to 2.5s.

[0142] In this way, due to the segmentation, therefore, the features of the user can be obtained by using the audio with a shorter time after segmentation. At the same time, according to the characteristics of the audio, there will only be one user in the vast majority of audio with a duration of 1.5s - 2.5s. Therefore, through this way of segmentation, it can be basically ensured that D_max_seg does not contain label noise. In addition, an extended dataset can be obtained through segmentation, so as to be used as data for training the basic model.

[0143] Step 103: Based on the third audio set, obtain a voiceprint feature vector set of the voiceprint recognition data set.

[0144] In this step, model training is performed based on the third audio set to obtain a first voiceprint recognition model, and the first voiceprint recognition model is used to extract voiceprint feature vectors of the audio in the voiceprint recognition data set, and the voiceprint feature vectors are used to form the voiceprint feature vector set.

[0145] Specifically, a basic voiceprint recognition model M_max_seg can be trained through the third audio set D_max_seg. Since there is no influence of label noise, the performance of M_max_seg will not be affected. Specifically, during the training process, for the audio in D_max_seg, the frequency domain fbank features are extracted. Then the extracted fbank features are sent to the feature extraction network based on TDNN (Time Delay Neural Network) to output the voiceprint feature vector. Then, the voiceprint feature vector is sent to the fully connected classification network for classification.

[0146] After the model M_max_seg is trained, voiceprint feature vectors vi are extracted from all the audio in the voiceprint recognition dataset D, and each vi is used to form a voiceprint feature vector set Vi.

[0147] Step 104: Determine registration audio based on the voiceprint feature vector set and the first audio set.

[0148] In the embodiment of the present application, the concept of compactness of a voiceprint vector cluster (voiceprint feature vector set) is introduced to evaluate the similarity within the entire voiceprint vector cluster (voiceprint feature vector set). The meaning of compactness refers to the average similarity of all vectors in the voiceprint vector cluster of users with the same label, where the similarity calculation formula uses cosine distance for calculation.

[0149] For example, in a certain voiceprint feature vector set, the compactness of a certain voiceprint feature vector can be expressed as:

[0150]

[0151] Among them, C i represents the compactness of the voiceprint feature vector, M is the number of voiceprint feature vectors in the voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0152] Among them, for a voiceprint vector vi and a voiceprint vector cluster Fi, its compactness is Ci, and the compactness deltaC (compactness change value) of the vector vi is defined as the change value of Ci before and after vi is added to Fi, that is, DeltaC_vi = Ci_pre-Ci_post, Ci_pre is the compactness corresponding to vi before adding the vi to Fi; Ci_post is the compactness corresponding to vi after adding the vi to Fi.

[0153] Based on this, in the embodiment of the present application, the registration audio is determined as follows:

[0154] S1. For each first audio in the first audio set, calculate a compactness change value of a voiceprint feature vector of the first audio. The first audio is any audio in the first audio set.

[0155] For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner:

[0156] DeltaC_vi=Ci_pre-Ci_post

[0157] Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

[0158] The compactness corresponding to the vi is determined according to the following formula:

[0159]

[0160] Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0161] S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio, that is, the replacement audio can be used as the first audio to perform the process of S1-S2. If the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set. The replacement audio can be any audio of the audio corresponding to the user of the first audio.

[0162] S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

[0163] Step 105: re-label the audio in the second audio set based on the registered audio to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

[0164] The audio of each user si in the finally obtained first audio set is registered to obtain a first registration vector si_enroll of each user si. Then, the audio in the second audio set is re-labeled based on the first registration vector to obtain a fourth audio set. Here, the method of registering the audio of each user si is not limited.

[0165] In order to solve the interference of channel information on the registration vector, in an embodiment of the present application, the mean vector Vi_mean of all voiceprint feature vectors in the voiceprint feature vector set can be obtained, and the mean vi_d_max_mean of the voiceprint feature vector vi_d_max of each user's audio in the third audio set D_max_seg can be obtained. Afterwards, the difference between the mean of the voiceprint feature vector of each user's audio and the mean vector is used as the first registration vector, expressed as: si_enroll = vi_d_max_mean-Vi_mean. The voiceprint feature vector vi_d_max of each user's audio in the third audio set D_max_seg can also be extracted using the first voiceprint recognition model obtained by the aforementioned training.

[0166] In the process of re-labeling the audio in the second audio set based on the first registration vector to obtain the fourth audio set, for the second audio in the second audio set, the voiceprint recognition can be performed on the first registration vector of the user corresponding to the second audio and the label of the second audio. The second audio is any audio in the second audio set.

[0167] If the second audio matches the first registration vector of the user corresponding to the label of the second audio, it means that the label of the second audio is accurate, and the second audio is added to the first audio set and deleted from the second audio set. The meaning of matching can be understood as being the same or most similar.

[0168] If the second audio does not match the first registration vector of the user corresponding to the tag of the second audio, the second audio is matched with the audio of other users to obtain the audio of the target user matching the second audio. The other users may be one or more. Here, the audio of the target user matching the second audio refers to the audio of the user who is most similar to the user of the second audio.

[0169] After determining the audio of the target user, the compactness change value corresponding to the second audio is obtained based on the second audio and the audio of the target user. Specifically, the voiceprint feature vector of each audio in the audio set formed by the second audio and the audio of the target user matching the second audio can be calculated in the above manner to form a target voiceprint feature vector set. Then, based on the voiceprint feature vector set and the voiceprint feature vector of the second audio, the compactness change value corresponding to the second audio is obtained in the above method of calculating the compactness change value.

[0170] If the compactness change value corresponding to the second audio is greater than 0, the second audio is marked with the label of the target user, and the second audio is added to the first audio set and deleted from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, the second audio is retained in the second audio set.

[0171] In the embodiment of the present application, on the basis of the above, the obtained first audio set can also be used to train the second voiceprint recognition model M_d, and the second voiceprint recognition model can be used to perform the process shown in the aforementioned step 105 until the number of audios in the first audio set remains unchanged. The first audio set finally obtained is used as the fourth audio set.

[0172] In the embodiment of the present application, the data in the first audio set in the voiceprint recognition data set is segmented to obtain the third audio data, and based on the third audio set, the voiceprint feature vector set and the registration audio of the voiceprint recognition data set are obtained. Therefore, based on the registration audio, the audio in the second audio set can be re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise. Therefore, in the embodiment of the present application, through a series of processes such as segmentation, recognition, and re-labeling, the label noise of the voiceprint recognition data set can be filtered, the integrity of the data set is retained, and the voiceprint recognition effect is guaranteed.

[0173] See also Figure 2 , Figure 2 is a schematic diagram of the audio data processing process provided by the embodiment of the present application, such as Figure 2 As shown, the process may include:

[0174] The longest audio of each user is selected from the voiceprint recognition data set D to obtain a first audio set speech and a second audio set D_max-. The concepts of the first audio set and the second audio set can refer to the description of the above embodiments.

[0175] Each audio in the first audio set is segmented to obtain a set consisting of multiple segmented audios, namely D_max_seg, and D_max_seg is used for model training to obtain a voiceprint model M. Afterwards, user registration is performed on D_max_seg, the registered audio is determined, and the registration vector of each user tag is calculated.

[0176] Re-label the audio in D_max- through voiceprint model M: (1) For audio U-max-_i in D_max-, extract the voiceprint feature vector through voiceprint model M, and perform one-to-one (1vs1) voiceprint recognition on it and the user si_enroll corresponding to its label. If a match is made, it means that the original label of audio U-max-_i is accurate. (2) If the voiceprint recognition fails to pass 1vs1, the voiceprint feature vector of audio U-max-_i is matched one by one with other users (1vs(N-1)) to obtain the most similar user S-max_N-1. Then, calculate the compactness DeltaC of U-max-_i. If DeltaC is greater than 0, U-max-_i is placed in the user cluster corresponding to S-max_N-1 (that is, the label is set to the label corresponding to S-max_N-1), that is, it is added to the first audio set; if it is less than 0, no processing is done and it continues to stay in D_max-. Retrain the voiceprint recognition model M_d with the new D_max, and then repeat the above process until the audio in D_max no longer increases, discard D_max-, and obtain the data set D_max with label noise removed.

[0177] From the above description, it can be seen that by utilizing the solution of the embodiments of the present application, the label noise problem of voiceprint recognition training data can be filtered through the segmentation-recognition-relabeling method, which retains the integrity of the data set, is simple to implement, has strong applicability, performs well under high levels of label noise, and can be used as a general label noise processing method in the field of voiceprint recognition.

[0178] See also Figure 3 , Figure 3 is a structural diagram of an audio data processing device provided in an embodiment of the present application, which is applied to a terminal. Figure 3 As shown, the audio data processing device includes:

[0179] A first acquisition module 301 is used to acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set;

[0180] A second acquisition module 302, configured to segment the audio in the first audio set to obtain a third audio set;

[0181] A third acquisition module 303, configured to obtain a voiceprint feature vector set of the voiceprint recognition data set based on the third audio set;

[0182] A first determination module 304, configured to determine the registration audio based on the voiceprint feature vector set and the first audio set;

[0183] The fourth acquisition module 305 is configured to re-label the audio in the second audio set based on the registered audio to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

[0184] Optionally, the second acquisition module is further used for:

[0185] Each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set, wherein the predetermined time length is greater than or equal to 1.5 seconds and less than or equal to 2.5 seconds.

[0186] Optionally, the third acquisition module includes:

[0187] A first training submodule, configured to perform model training based on the third audio set to obtain a first voiceprint recognition model;

[0188] A first extraction submodule, configured to extract a voiceprint feature vector of the audio in the voiceprint recognition data set by using the first voiceprint recognition model;

[0189] The first acquisition submodule is used to form the voiceprint feature vector set using the voiceprint feature vector.

[0190] Optionally, the fourth acquisition module is further used to:

[0191] S1. For each first audio in the first audio set, calculate a compactness change value of the voiceprint feature vector of the first audio;

[0192] S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio; if the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set;

[0193] S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

[0194] Optionally, the fourth acquisition module is further used to:

[0195] For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner:

[0196] DeltaC_vi=Ci_pre-Ci_post

[0197] Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

[0198] Optionally, the compactness corresponding to the vi is determined according to the following formula:

[0199]

[0200] Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0201] Optionally, the fourth acquisition module includes:

[0202] A first registration submodule, configured to register the audio of each user in the finally obtained first audio set to obtain a first registration vector for each user;

[0203] The first acquisition submodule is configured to re-label the audio in the second audio set based on the first registration vector to obtain a fourth audio set.

[0204] Optionally, the first registration submodule includes:

[0205] A first acquisition unit, configured to acquire a mean vector of all voiceprint feature vectors in the voiceprint feature vector set;

[0206] A second acquisition unit, configured to acquire a mean value of a voiceprint feature vector of the audio of each user in the third audio set;

[0207] The third acquisition unit is used to use the difference between the mean of the voiceprint feature vector of each user's audio and the mean vector as the first registration vector.

[0208] Optionally, the first acquisition submodule includes:

[0209] A first processing unit is configured to perform voiceprint recognition on a second audio in the second audio set by using the second audio and a first registration vector of a user corresponding to a label of the second audio;

[0210] a second processing unit, configured to, if the second audio matches the first registration vector of the user corresponding to the label of the second audio, indicate that the labeling of the second audio is accurate, and add the second audio to the first audio set and delete it from the second audio set;

[0211] a third processing unit, configured to, if the second audio does not match the first registration vector of the user corresponding to the label of the second audio, match the second audio with audios of other users to obtain audio of a target user matching the second audio;

[0212] A first acquisition unit, configured to acquire a compactness change value corresponding to the second audio based on the second audio and the audio of the target user;

[0213] The fourth processing unit, if the compactness change value corresponding to the second audio is greater than 0, marks the second audio with the label of the target user, adds the second audio to the first audio set, and deletes it from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, retains the second audio in the second audio set.

[0214] Optionally, the device further comprises:

[0215] A training module, used for training a second voiceprint recognition model using the obtained first audio set;

[0216] The aforementioned method is performed using the second voiceprint recognition model until the number of audios in the first audio set remains unchanged.

[0217] The device provided in the embodiment of the present application can execute the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0218] See also Figure 4 , Figure 4 1 is a structural diagram of an audio data processing device provided in an embodiment of the present application, which is applied to a network device. Figure 4 As shown, the audio data processing device includes: a processor 401 and a transceiver 402 .

[0219] The processor 401 is configured to:

[0220] Acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set;

[0221] Splitting the audio in the first audio set to obtain a third audio set;

[0222] Based on the third audio set, obtaining a voiceprint feature vector set of the voiceprint recognition data set;

[0223] Determine the registration audio based on the voiceprint feature vector set and the first audio set;

[0224] Based on the registered audio, the audio in the second audio set is re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

[0225] Optionally, the processor 401 is further configured to:

[0226] Each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set, wherein the predetermined time length is greater than or equal to 1.5 seconds and less than or equal to 2.5 seconds.

[0227] Optionally, the processor 401 is further configured to:

[0228] Performing model training based on the third audio set to obtain a first voiceprint recognition model;

[0229] Extracting a voiceprint feature vector of the audio in the voiceprint recognition data set using the first voiceprint recognition model;

[0230] The voiceprint feature vector set is formed by using the voiceprint feature vector.

[0231] Optionally, the processor 401 is further configured to:

[0232] S1. For each first audio in the first audio set, calculate a compactness change value of the voiceprint feature vector of the first audio;

[0233] S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio; if the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set;

[0234] S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

[0235] Optionally, the processor 401 is further configured to:

[0236] For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner:

[0237] DeltaC_vi=Ci_pre-Ci_post

[0238] Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

[0239] Optionally, the compactness corresponding to the vi is determined according to the following formula:

[0240]

[0241] Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

[0242] Optionally, the processor 401 is further configured to:

[0243] Performing user registration on the audio of each user in the finally obtained first audio set to obtain a first registration vector of each user;

[0244] The audio in the second audio set is re-labeled based on the first registration vector to obtain a fourth audio set.

[0245] Optionally, the processor 401 is further configured to:

[0246] Obtaining a mean vector of all voiceprint feature vectors in the voiceprint feature vector set;

[0247] Obtaining a mean value of the voiceprint feature vector of the audio of each user in the third audio set;

[0248] The difference between the mean of the voiceprint feature vector of each user's audio and the mean vector is used as the first registration vector.

[0249] Optionally, the processor 401 is further configured to:

[0250] For a second audio in the second audio set, performing voiceprint recognition on the second audio and a first registration vector of a user corresponding to the label of the second audio;

[0251] If the second audio matches the first registration vector of the user corresponding to the label of the second audio, it indicates that the labeling of the second audio is accurate, and the second audio is added to the first audio set and deleted from the second audio set;

[0252] If the second audio does not match the first registration vector of the user corresponding to the label of the second audio, matching the second audio with audios of other users to obtain audio of the target user matching the second audio;

[0253] Based on the second audio and the audio of the target user, obtaining a compactness change value corresponding to the second audio;

[0254] If the compactness change value corresponding to the second audio is greater than 0, the second audio is marked with the label of the target user, and the second audio is added to the first audio set and deleted from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, the second audio is retained in the second audio set.

[0255] Optionally, the processor 401 is further configured to:

[0256] Using the obtained first audio set to train a second voiceprint recognition model;

[0257] The aforementioned method is performed using the second voiceprint recognition model until the number of audios in the first audio set remains unchanged.

[0258] The device provided in the embodiment of the present application can execute the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0259] It should be noted that the division of units in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0260] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0261] An embodiment of the present application provides a communication device, including: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is used to read the program in the memory to implement the steps in the audio data processing method as described above.

[0262] The embodiment of the present application also provides a readable storage medium, on which a program is stored. When the program is executed by the processor, each process of the above-mentioned audio data processing method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. Among them, the readable storage medium can be any available medium or data storage device that can be accessed by the processor, including but not limited to magnetic storage (such as floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical storage (such as CD, DVD, BD, HVD, etc.), and semiconductor storage (such as ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)), etc.

[0263] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0264] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0265] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. An audio data processing method, characterized in that: include: Acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audios of different users, and the second audio set includes the audios in the voiceprint recognition data set except the first audio set; Segmenting the audio in the first audio set to obtain a third audio set; Based on the third audio set, obtaining a voiceprint feature vector set of the voiceprint recognition data set; Determine the registration audio based on the voiceprint feature vector set and the first audio set; Based on the registered audio, the audio in the second audio set is re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

2. The method according to claim 1, characterized in that The step of dividing the audio in the first audio set to obtain a third audio set includes: Each audio in the first audio set is segmented according to a predetermined time length to obtain the third audio set, wherein the predetermined time length is greater than or equal to 1.5 seconds and less than or equal to 2.5 seconds.

3. The method according to claim 1, characterized in that The step of obtaining a voiceprint feature vector set of the voiceprint recognition data set based on the third audio set includes: Performing model training based on the third audio set to obtain a first voiceprint recognition model; Extracting a voiceprint feature vector of the audio in the voiceprint recognition data set using the first voiceprint recognition model; The voiceprint feature vector set is formed by using the voiceprint feature vector.

4. The method according to claim 1, characterized in that: The determining the registration audio based on the voiceprint feature vector set and the first audio set includes: S1. For each first audio in the first audio set, calculate a compactness change value of the voiceprint feature vector of the first audio; S2. If the compactness change value of the first audio is greater than 0, select a replacement audio from the audio corresponding to the user of the first audio in the voiceprint recognition data set, and use the replacement audio to replace the first audio; if the compactness change value of the first audio is less than or equal to 0, retain the first audio in the first audio set; S1-S2 are repeatedly performed until the compactness change value of each audio in the first audio set is less than or equal to 0, and the audio in the first audio set finally obtained is used as the registered audio.

5. The method according to claim 3, characterized in that: The step of calculating, for each first audio in the first audio set, a compactness change value of the voiceprint feature vector of the first audio includes: For the first audio, the compactness change value of the voiceprint feature vector of the first audio is determined in the following manner: DeltaC_vi=Ci_pre-Ci_post Among them, DeltaC_vi represents the compactness change value of the voiceprint feature vector vi of the first audio; Ci_pre is the compactness corresponding to the vi before adding the vi to the voiceprint feature vector set; Ci_post is the compactness corresponding to the vi after adding the vi to the voiceprint feature vector set.

6. The method according to claim 5, characterized in that The compactness corresponding to the vi is determined according to the following formula: Among them, C i represents the compactness corresponding to the said vi, N is the number of voiceprint feature vectors in the said voiceprint feature vector set, s(x i,j ,x i,k ) represents the voiceprint feature vector x in the voiceprint feature vector set i,j With x i,k The similarity between .

7. The method according to claim 4, characterized in that The step of re-labeling the audio in the second audio set based on the registered audio to obtain a fourth audio set includes: Performing user registration on the audio of each user in the finally obtained first audio set to obtain a first registration vector of each user; The audio in the second audio set is re-labeled based on the first registration vector to obtain a fourth audio set.

8. The method according to claim 7, characterized in that The step of registering the audio of each user in the finally obtained first audio set to obtain a first registration vector of each user includes: Obtaining a mean vector of all voiceprint feature vectors in the voiceprint feature vector set; Obtaining a mean value of the voiceprint feature vector of the audio of each user in the third audio set; The difference between the mean of the voiceprint feature vector of each user's audio and the mean vector is used as the first registration vector.

9. The method according to claim 8, characterized in that The step of re-labeling the audio in the second audio set based on the first registration vector to obtain a fourth audio set includes: For a second audio in the second audio set, performing voiceprint recognition on the second audio and a first registration vector of a user corresponding to the label of the second audio; If the second audio matches the first registration vector of the user corresponding to the label of the second audio, it indicates that the labeling of the second audio is accurate, and the second audio is added to the first audio set and deleted from the second audio set; If the second audio does not match the first registration vector of the user corresponding to the label of the second audio, matching the second audio with audios of other users to obtain audio of the target user matching the second audio; Based on the second audio and the audio of the target user, obtaining a compactness change value corresponding to the second audio; If the compactness change value corresponding to the second audio is greater than 0, the second audio is marked with the label of the target user, and the second audio is added to the first audio set and deleted from the second audio set; if the compactness change value corresponding to the second audio is less than or equal to 0, the second audio is retained in the second audio set.

10. The method according to claim 9, characterized in that The method further comprises: Using the obtained first audio set to train a second voiceprint recognition model; The method of claim 9 is performed using the second voiceprint recognition model until the number of audios in the first audio set remains unchanged.

11. An audio data processing device, characterized in that: include: A first acquisition module is used to acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set; A second acquisition module, configured to segment the audio in the first audio set to obtain a third audio set; A third acquisition module, configured to obtain a voiceprint feature vector set of the voiceprint recognition data set based on the third audio set; A first determining module, configured to determine the registration audio based on the voiceprint feature vector set and the first audio set; The fourth acquisition module is used to re-label the audio in the second audio set based on the registered audio to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

12. An audio data processing device, characterized in that: include: processors and transceivers; Wherein, the processor is used for: Acquire a voiceprint recognition data set with annotations, wherein the voiceprint recognition data set includes a first audio set and a second audio set, the first audio set includes the longest audio of each user, and the second audio set includes the audio in the voiceprint recognition data set except the first audio set; Segmenting the audio in the first audio set to obtain a third audio set; Based on the third audio set, obtaining a voiceprint feature vector set of the voiceprint recognition data set; Determine the registration audio based on the voiceprint feature vector set and the first audio set; Based on the registered audio, the audio in the second audio set is re-labeled to obtain a fourth audio set, wherein the fourth audio set does not include label noise.

13. A communication device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; wherein the processor is used to read the program in the memory to implement the steps in the method for processing signed noise data as described in any one of claims 1 to 9.

14. A readable storage medium for storing a program, characterized in that: When the program is executed by a processor, the steps of the method for processing signed noise data according to any one of claims 1 to 9 are implemented.