CN106157959B - voiceprint model updating method and system - Google Patents

Abstract

The invention discloses a method and system for updating a voiceprint model. The method includes: obtaining the current login time of a target speaker and the last update time of the target speaker's voiceprint model; and obtaining the last update time of the target speaker's voiceprint model The time period from the current login time of the target speaker is divided into multiple time clusters; the voice data of the target speaker’s voiceprint authentication in each time cluster is obtained; Select voice data in the data as the target speaker's voiceprint model update data; use the target speaker's voiceprint model update data and the original voiceprint model training data to perform voiceprint model training again, and obtain a new voiceprint model after training; The new voiceprint model is used to update the voiceprint model of the target speaker. The present invention can effectively prevent the impersonator speaker from continuously updating the voiceprint model of the target speaker in a short period of time, and ensure the correctness of updating the voiceprint model of the target speaker.

Description

Method and system for updating voiceprint model

technical field

The invention relates to the technical field of voiceprint recognition, in particular to a method and system for updating a voiceprint model.

Background technique

The voiceprint feature is one of the important biological characteristics of the human body. It is not only specific, but also relatively stable. It is often used in voiceprint recognition, voiceprint authentication and other fields.

The method of identity authentication through voiceprint features is usually to represent different individuals by establishing a voiceprint model, and then use the voiceprint model to identify different individuals. In general, when training the voiceprint model, very little training data is obtained, and the accuracy and adaptability of the trained voiceprint model are poor. Therefore, in order to increase the practicability of the voiceprint model, it is necessary to continuously update the voiceprint model in the later stage to increase its accuracy and adaptability. The existing voiceprint model update method is usually to directly add the newly recorded voice data of the successfully authenticated user to the training database of the original voiceprint model, and re-train the voiceprint model to complete the update of the voiceprint model.

However, when a false speaker with similar voiceprint characteristics to the target speaker continues to impersonate in a short period of time, a large amount of speech data will be continuously entered into the training database of the original voiceprint model, and the original voiceprint model will be continuously updated. As a result, the updated voiceprint model deviates from the actual voiceprint model of the target speaker, and ultimately the target speaker cannot be authenticated.

Contents of the invention

Embodiments of the present invention provide a method and system for updating a voiceprint model to solve the problem of the updated voiceprint caused by false speakers whose voiceprint features are similar to the target speaker in the prior art. The problem of the model deviating from the actual voiceprint model of the target speaker.

For this reason, the embodiment of the present invention provides following technical scheme:

A method for updating a voiceprint model, comprising:

Obtain the current login time of the target speaker and the last update time of the voiceprint model of the target speaker;

Divide the time period from the last update time of the voiceprint model of the target speaker to the current login time of the target speaker into multiple time clusters;

Obtain the voice data when the voiceprint authentication of the target speaker in each time cluster is successful;

Select voice data from the voice data successfully authenticated in each time cluster, and use the selected voice data as the voiceprint model update data of the target speaker;

Using the voiceprint model update data of the target speaker and the original voiceprint model training data to re-train the voiceprint model to obtain a new voiceprint model after training;

Utilizing the new voiceprint model to update the voiceprint model of the target speaker.

Preferably, the selecting voice data from voice data successfully authenticated in each time cluster includes:

For each time cluster, obtain the matching degree of the voice data of the successful authentication in the time cluster relative to the voiceprint model of the target speaker;

Filter out the voice data corresponding to the matching degree greater than the set first threshold; or

The pieces of voice data are sorted in descending order of matching degree, and the voice data with a set number of pieces are screened out.

Preferably, the selecting voice data from the voice data successfully authenticated in each time cluster further includes:

The filtered speech data within each temporal cluster is sampled.

Preferably, the method also includes:

When performing voiceprint authentication on a user, obtain the voice data of the current user;

Calculate the matching degree between the voiceprint feature of the voice data of the current user and the target speaker's voiceprint model, and use the matching degree as the matching degree to be regularized;

Carrying out zero regularization to the matching degree to be regularized to obtain the regularized matching degree;

If the regularized matching degree is greater than the set second threshold, it is determined that the current user is the target speaker.

Preferably, the method also includes:

Pre-collect a large amount of speech data from different speakers as seed data and put them into the seed collection;

Calculate the matching degree of the voiceprint feature of each piece of voice data in the seed set and the target speaker's voiceprint model to obtain a matching degree set;

Calculate the mean value and standard deviation of all matching degrees in the matching degree set, and use the calculated mean value and standard deviation as the distribution of the matching degree distribution between the voiceprint feature of the voice data of the impostor speaker and the voiceprint model of the target speaker mean and standard deviation;

The said matching degree to be adjusted is zero-adjusted, and the matching degree obtained after adjustment includes:

Using the mean value and standard deviation to zero-adjust the matching degree to be adjusted to obtain the adjusted matching degree.

A voiceprint model updating system, comprising:

The update period acquisition module is used to obtain the current login time of the target speaker and the last update time of the target speaker's voiceprint model;

The division module is used to divide the time period from the last update time of the target speaker's voiceprint model to the current login time of the target speaker into multiple time clusters;

Voice data acquisition module, used to acquire the voice data when the target speaker's voiceprint authentication is successful in each time cluster;

The model update data acquisition module is used to select voice data from the voice data successfully authenticated in each time cluster, and use the selected voice data as the voiceprint model update data of the target speaker;

The training module is used to re-train the voiceprint model by using the update data of the voiceprint model of the target speaker and the training data of the original voiceprint model to obtain a new voiceprint model after training;

A model updating module, configured to update the voiceprint model of the target speaker by using the new voiceprint model.

Preferably, the model update data acquisition module includes:

The obtaining unit is used to obtain, for each time cluster, the matching degree of the voice data successfully authenticated in the time cluster relative to the voiceprint model of the target speaker;

The filtering unit is used to filter out voice data corresponding to a matching degree greater than a set first threshold; or sort each piece of voice data in descending order of matching degree, and filter out a set number of voice data.

Preferably, the model update data acquisition module also includes:

A sampling unit, configured to sample the speech data screened out by the screening unit in each time cluster.

Preferably, the system also includes:

The receiving module is used to obtain the voice data of the current user when performing voiceprint authentication on the user;

A matching degree calculation module, configured to calculate the matching degree between the voiceprint feature of the voice data of the current user and the target speaker's voiceprint model, and use the matching degree as the matching degree to be regularized;

A regularizing module, configured to perform zero regularization on the matching degree to be regularized to obtain the regularized matching degree;

A judging module, configured to determine whether the current user is the target speaker according to the adjusted matching degree, and determine that the current user is the target speaker if the adjusted matching degree is greater than a set second threshold.

Preferably, the system also includes:

The seed data acquisition module is used to pre-collect a large amount of speech data from different speakers as seed data and put them into the seed collection;

The first calculation module is used to calculate the matching degree between the voiceprint feature of each piece of speech data in the seed set and the voiceprint model of the target speaker, to obtain a matching degree set;

The second calculation module is used to calculate the mean value and standard deviation of all matching degrees in the matching degree set, and use the calculated mean value and standard deviation as the voiceprint feature of the voice data of the impostor speaker and the target speaker The mean and standard deviation of the matching degree distribution of the voiceprint model;

The normalizing module is specifically configured to use the mean value and standard deviation to perform zero normalization on the matching degree to be normalized to obtain the normalized matching degree.

The method and system for updating the voiceprint model provided by the embodiments of the present invention divide the current login time of the target speaker and the last update time of the voiceprint model of the target speaker into multiple time clusters, and obtain the information in each time cluster. The voice data of the target speaker's voiceprint authentication is successful, and then select the voice data from the voice data of the successful authentication in the cluster at different times as the update data of the target speaker's voiceprint model, and then use these target speaker's voiceprint model to update the data and The original voiceprint model training data is re-trained for the voiceprint model, and the new voiceprint model is used to update the target speaker's voiceprint model, which can effectively prevent the impostor speaker from continuously updating the target speaker's voiceprint model in a short period of time. Ensure the correctness of the target speaker's voiceprint model update.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings that are required in the embodiments. Obviously, the accompanying drawings in the following description are only described in the present invention For some embodiments of the present invention, those skilled in the art can also obtain other drawings according to these drawings.

FIG. 1 is a flowchart of a method for updating a voiceprint model according to an embodiment of the present invention;

Fig. 2 is a kind of flowchart of user authentication in the embodiment of the present invention;

3 is a schematic structural diagram of a voiceprint model updating system according to an embodiment of the present invention;

Fig. 4 is a partial structural diagram of authenticating a user in the system for updating a voiceprint model according to an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention will be further described in detail below in conjunction with the drawings and implementations.

The present invention is different from the prior art: the voice data of all users who have successfully authenticated are added to the training data of the original voiceprint model as model update data, resulting in the voice data of the impostor being included in the update data, and the voiceprint model of the target speaker is updated. , making it deviate from the target speaker's voiceprint characteristics. The present invention divides the current login time of the target speaker and the last update time of the voiceprint model of the target speaker into multiple time clusters, and selects the successfully authenticated voice data in each time cluster as the update of the voiceprint model of the target speaker data, and then use these data and the original voiceprint model training data to update the voiceprint model of the target speaker, which can effectively solve the problem of continuous impersonation of speakers who are similar to the voiceprint of the target speaker, resulting in continuous updating of the target speaker. The voiceprint model makes it deviate from the target speaker's voiceprint characteristics, and ensures the correctness of the update of the target speaker's voiceprint model, thereby ensuring the correctness of subsequent user authentication.

As shown in Figure 1, it is a flow chart of the method for updating the voiceprint model in the embodiment of the present invention, including the following steps:

Step 101, obtain the current login time of the target speaker and the last update time of the voiceprint model of the target speaker.

In this embodiment, by acquiring the current login time of the target speaker and the last update time of the target speaker's voiceprint model, the time range is used as the time range for collecting voiceprint model update data. The target speaker’s voiceprint model update cycle can be set according to the user’s habits. For example, if the user logs in frequently in a certain period of time, but usually logs in less often, the model update time can be set longer, such as updating once a month; If the user has been logging in frequently, the model update time can be set to be shorter, such as once a week. For another example, it can be set that the target speaker's voiceprint model will be updated after the user logs in for a certain number of times.

Step 102: Divide the time period from the last update time of the voiceprint model of the target speaker to the current login time of the target speaker into multiple time clusters.

The size of agglomerates at each time can be calculated according to the predetermined number N of agglomerates.

Assume that the current login time of the target speaker is T ₁ , the last update time of the voiceprint model is T ₂ , and the cluster size at the i-th time is t _i , as shown in formula (1):

Among them, λ _i is the impact factor, The impact factor depends on the actual application environment, and the specific value can be set manually or obtained through training with a large amount of data. If the actual application environment has poor network security and many hackers in a certain period of time, it is necessary to increase the influence factor λ _i accordingly to expand the time cluster size t _i to prevent the impostor speaker from continuously entering in a short period of time A large amount of speech data continuously updates the target speaker model. when , it is the average division of each time cluster size.

Step 103, acquiring voice data of the target speaker in each time cluster when the voiceprint authentication is successful.

In practical applications, when voiceprint authentication is performed on a user, after each successful authentication, the voice data of the target speaker whose authentication is successful this time can be recorded, so that when the voiceprint model is updated in the future, these Get the corresponding voice data in the record.

Step 104, select speech data from the speech data successfully authenticated in each time cluster as the update data of the voiceprint model of the target speaker.

Specifically, the authenticated voice data in each time cluster can be randomly selected, and the number of pieces of voice data selected in each time cluster can be the same or different, which is not limited in this embodiment of the present invention.

In addition, in order to further improve the effectiveness of the selected update data, the voice data in each time cluster can also be screened according to the matching degree, specifically, the relative target of each voice data successfully authenticated in the time cluster can be obtained The matching degree of the speaker's voiceprint model is then screened by selecting voice data according to the matching degree score, for example:

If the matching degree is greater than a set first threshold, then filter out the voice data corresponding to the matching degree; or

The pieces of voice data are sorted in descending order of matching degree, and the voice data with a set number of pieces are screened out.

The first threshold may be an experimental result or an empirical value.

Similarly, when performing voiceprint authentication on a user, after each successful authentication, it is possible to record the matching degree of the voice data of this successful authentication relative to the voiceprint model of the target speaker, so that when the voiceprint model is updated later, that is Corresponding matching degrees can be obtained from these records.

Because the less data selected in each time period, the less likely it is for an impostor to impersonate. Therefore, in order to prevent impersonators from intensively impersonating in a short period of time, the speech data screened out in each time cluster can also be sampled to obtain the update data of the voiceprint model of the target speaker. If the speech data screened out in a time cluster is 0, no sampling will be performed, if it is 1, it will be used directly, if it is more than 1, one or more speech data can be selected from it, the specific sampling method can be For random sampling or other sampling methods.

Step 105, using the voiceprint model update data of the target speaker and the original voiceprint model training data to perform voiceprint model training again to obtain a new trained voiceprint model.

In this embodiment, the voiceprint model is retrained according to the obtained target speaker's voiceprint model update data and original voiceprint model training data to obtain a new voiceprint model.

Step 106, using the new voiceprint model to update the voiceprint model of the target speaker.

That is, the target speaker's voiceprint model is updated by using the new voiceprint model obtained in step 105 .

Since the method of the present invention divides the update period into multiple time clusters, and then selects the update data of the voiceprint model of the target speaker from the voice data successfully authenticated in each time cluster, avoiding the impersonator who is similar to the target speaker When the speaker continues to impersonate in a short period of time, the target speaker cannot be authenticated, which ensures the correctness of the update of the target speaker's voiceprint model, thereby ensuring the correctness of subsequent user authentication.

It should be noted that the method for updating the voiceprint model in the embodiment of the present invention can be applied to various voiceprint authentication schemes, which is not limited in the present invention.

Due to the different frequency of use by different users, the update degree of the voiceprint model is different. For users with high frequency of use, the model matching degree is often high, while for users with low frequency of use, the model matching degree is often low, that is, the model matching degree of different users The model fit distribution is inconsistent.

Considering this situation, in another embodiment of the method of the present invention, the identity authentication of the user can also be performed in a zero-regular manner, so as to solve the above-mentioned problem of inconsistency in the distribution of the model matching degree caused by the different updating degrees of the voiceprint models of different users . In this embodiment, the flow process of user authentication is as shown in Figure 2, including the following steps:

Step 201, when voiceprint authentication is performed on a user, voice data of the current user is acquired.

Step 202, calculating the matching degree between the voiceprint feature of the current user's voice data and the target speaker's voiceprint model, and using the matching degree as the matching degree to be regularized.

The calculation of the matching degree is regularized as shown in formula (2):

Among them, P(X|SM) is the likelihood of the voiceprint feature X relative to the speaker's voiceprint model SM (Speaker Model, SM), P(X|UBM) is the voiceprint feature X relative to the universal background model UBM (Universal Background Model, UBM) likelihood. The voiceprint model SM and the universal background model UBM can be constructed in advance by collecting data.

Step 203, performing zero rounding on the matching degree to be regularized to obtain the matching degree after regularization.

In the embodiment of the present invention, the above-mentioned matching degree can be regularized by using the mean μ and standard deviation σ of the matching degree distribution of the impostor speaker on the target speaker model to obtain the regularized matching degree S′.

The regularized matching degree distribution can roughly be regularized to a standard normal distribution with a mean of 0 and a variance of 1, as shown in formula (3):

Among them, S is the matching degree to be regularized.

The above mean value μ and standard deviation σ can be calculated as follows:

1. Pre-collect a large amount of speech data from different speakers as seed data and put them into the seed collection;

2. Calculate the matching degree QS of the voiceprint feature of each piece of voice data in the seed collection and the target speaker's voiceprint model, obtain the matching degree set, as shown in the following formula (4):

Among them, QS _i represents the matching degree of the voiceprint feature of the i-th voice data with respect to the target speaker model, and A _i represents the voiceprint feature of the i-th voice data.

3. Calculate the mean and standard deviation of all matching degrees in the matching degree set, and use the calculated mean and standard deviation as the matching degree between the voiceprint feature of the voice data of the impostor speaker and the voiceprint model of the target speaker The mean and standard deviation of the distribution.

Specifically, it is necessary to calculate the mean and standard deviation of the matching degree of each speech feature in the seed set relative to the target speaker model, and the specific calculation methods are shown in formula (5) and formula (6):

Among them, N represents the number of voice entries in the seed set.

Step 204, judging whether the matching degree after regularization is greater than the set second threshold; if yes, go to step 205; otherwise, go to step 206.

Step 205, the authentication is successful, and the current user is determined as the target speaker.

Step 206, the authentication fails.

Generally, the second threshold can be set according to experimental results or experience.

The voiceprint authentication method provided by the embodiment of the present invention calculates the matching degree between the voiceprint feature of the current user's voice data and the voiceprint model of the target speaker, and zero-adjusts the matching degree to obtain the adjusted matching degree. It is compared with the set second threshold to determine whether the current user is the target speaker. Due to the use of a large number of non-target speaker voice data and the target speaker voiceprint model matching average and standard deviation, the current user voice data The matching degree is adjusted to zero, which further avoids the problem of inconsistent model matching degree distribution caused by different update degrees of voiceprint models of different users.

Correspondingly, an embodiment of the present invention also provides a system for updating a voiceprint model, as shown in FIG. 3 , which is a schematic structural diagram of the system.

In this embodiment, the system includes:

The update period acquisition module 300 is used to acquire the current login time of the target speaker and the last update time of the voiceprint model of the target speaker;

The division module 301 is used to divide the time period from the last update time of the target speaker's voiceprint model to the current login time of the target speaker into a plurality of time clusters;

Voice data acquisition module 302, for acquiring the voice data when the voiceprint authentication of the target speaker in each time cluster is successful;

Model update data acquisition module 303, used to select voice data from the voice data of successful authentication in each time group, as the voiceprint model update data of the target speaker;

The training module 304 is used to re-train the voiceprint model by using the voiceprint model update data of the target speaker and the original voiceprint model training data to obtain a new voiceprint model after training;

A model updating module 305, configured to use the new voiceprint model to update the voiceprint model of the target speaker.

The above-mentioned model update data acquisition module 303 can specifically determine the number of pieces of voice data selected in the time cluster according to the number of pieces of voice data in the time cluster and randomly select the voiceprint authentication of the target speaker in the time cluster. voice data, the number of pieces of voice data selected in clusters at different times may be the same or different, which is not limited in this embodiment of the present invention.

In addition, in order to further make the selected speech data more effective, the model update data acquisition module 303 can also filter the speech data in each time cluster according to the matching degree. Correspondingly, a specific structure of the model update data acquisition module 303 includes:

The acquisition unit is configured to, for each temporal cluster, acquire the matching degree of each voice data in the temporal cluster with respect to the target speaker's voiceprint model.

The filtering unit is used to filter out voice data corresponding to a matching degree greater than a set first threshold; or sort each piece of voice data in descending order of matching degree, and filter out a set number of voice data.

In order to further prevent impersonators from intensive impersonation in a short period of time, in another embodiment of the model update data acquisition module 303, it may also include:

The sampling unit is configured to sample the speech data screened out in each time cluster to obtain update data of the target speaker's voiceprint model.

The voiceprint model update system in the embodiment of the present invention divides the voiceprint update period into multiple time clusters, and then selects the voiceprint model update data of the target speaker from the successfully authenticated voice data in each time cluster, and then uses It retrains a new voiceprint model with the training data of the original voiceprint model, and updates the original voiceprint model, which prevents the fake speaker who is similar to the voiceprint feature of the target speaker from continuously updating the voiceprint model of the target speaker in a short period of time. Making it deviate from the voiceprint characteristics of the target speaker leads to the problem that the target speaker cannot be successfully authenticated in the subsequent voiceprint authentication, which ensures the correctness of the update of the target speaker's voiceprint model, thereby ensuring the correctness of subsequent user authentication.

In addition, in practical applications, different users use voiceprint authentication at different frequencies. Generally, users with higher frequency of use have a higher degree of model matching, while users with lower frequency of use often have a lower degree of model matching. That is, different users The model matching degree distribution of is not consistent, therefore, in order to solve this problem, as shown in Figure 4, the system can further include:

The receiving module 400 is used to obtain the voice data of the current user when performing voiceprint authentication on the user;

The matching degree calculation module 401 is used to calculate the matching degree between the voiceprint feature of the voice data of the current user and the target speaker's voiceprint model, and use the matching degree as the matching degree to be regularized;

A regularizing module 402, configured to perform zero regularization on the matching degree to be regularized to obtain a regularized matching degree;

A judging module 403, configured to determine whether the current user is the target speaker according to the adjusted matching degree, and determine the current user as the target speaker if the adjusted matching degree is greater than a set second threshold.

Of course, in practical applications, the system may further include: a storage module (not shown), which is used to save the information of the successfully authenticated user (ie, the target speaker), such as: authentication time, voice data of successful authentication, successful authentication The matching degree of the voice data of the target speaker with the voiceprint model of the target speaker, etc. In this way, the above-mentioned voice data acquisition module 302 can directly acquire the voice data when the voiceprint authentication of the target speaker in each time cluster is successful from the storage module.

Wherein, the regularization module 402 can use the mean value and standard deviation of the matching degree of a large number of different speaker voice data and the voiceprint model of the target speaker to carry out zero regularization on the matching degree of the current user to obtain the regularized matching degree. Perform voiceprint authentication based on the regularized matching degree.

The mean value and standard deviation of the matching degree between the voice data of a large number of different speakers and the voiceprint model of the target speaker can be obtained through the following modules:

The seed data acquisition module is used to pre-collect a large amount of speech data from different speakers as seed data and put them into the seed collection;

The first calculation module is used to calculate the matching degree between the voiceprint feature of each piece of speech data in the seed set and the voiceprint model of the target speaker, to obtain a matching degree set;

The second calculation module is used to calculate the mean value and standard deviation of all matching degrees in the matching degree set, and use the calculated mean value and standard deviation as the voiceprint feature of the voice data of the impostor speaker and the target speaker The mean and standard deviation of the matching degree distribution of the voiceprint model.

The voiceprint authentication system provided by the embodiment of the present invention calculates the matching degree between the voiceprint feature of the current user's voice data and the target speaker's voiceprint model, and zero-adjusts the matching degree to obtain the adjusted matching degree. It is compared with the set second threshold to determine whether the current user is the target speaker. Since a large number of non-target speaker voice data and the target speaker's voiceprint model matching mean and standard deviation are used, the current user voice data The matching degree is adjusted to zero, which further avoids the problem of inconsistent model matching degree distribution caused by different update degrees of voiceprint models of different users.

Each embodiment in this specification is described in a progressive manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, as for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for related parts, please refer to part of the description of the method embodiment. The system embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without creative effort.

The embodiments of the present invention have been introduced in detail above, and the present invention has been described by using specific implementation methods in this paper. The description of the above embodiments is only used to help understand the method and system of the present invention; at the same time, for those of ordinary skill in the art, According to the idea of the present invention, there will be changes in the specific implementation and scope of application. To sum up, the contents of this specification should not be construed as limiting the present invention.

Claims (10)

1. A method for updating a voiceprint model, comprising: Obtain the current login time of the target speaker and the last update time of the voiceprint model of the target speaker; Divide the time period from the last update time of the voiceprint model of the target speaker to the current login time of the target speaker into multiple time clusters; Obtain the voice data when the voiceprint authentication of the target speaker in each time cluster is successful; Select voice data from the voice data successfully authenticated in each time cluster, and use the selected voice data as the voiceprint model update data of the target speaker; Using the voiceprint model update data of the target speaker and the original voiceprint model training data to re-train the voiceprint model to obtain a new voiceprint model after training; Utilizing the new voiceprint model to update the voiceprint model of the target speaker. 2. The method according to claim 1, wherein the selecting voice data from the voice data successfully authenticated in each time cluster comprises: For each time cluster, obtain the matching degree of the voice data of the successful authentication in the time cluster relative to the voiceprint model of the target speaker; Filter out the voice data corresponding to the matching degree greater than the set first threshold; or The pieces of voice data are sorted in descending order of matching degree, and the voice data with a set number of pieces are screened out. 3. The method according to claim 2, characterized in that, selecting voice data from the voice data of successful authentication in each time group also includes: The filtered speech data within each temporal cluster is sampled. 4. The method according to any one of claims 1 to 3, characterized in that the method further comprises: When performing voiceprint authentication on a user, obtain the voice data of the current user; Calculate the matching degree between the voiceprint feature of the voice data of the current user and the target speaker's voiceprint model, and use the calculated matching degree as the matching degree to be regularized; Carrying out zero regularization to the matching degree to be regularized to obtain the regularized matching degree; If the regularized matching degree is greater than the set second threshold, it is determined that the current user is the target speaker. 5. method according to claim 4, is characterized in that, described method also comprises: Pre-collect a large amount of speech data from different speakers as seed data and put them into the seed collection; Calculate the matching degree of the voiceprint feature of each piece of voice data in the seed set and the target speaker's voiceprint model to obtain a matching degree set; Calculate the mean value and standard deviation of all matching degrees in the matching degree set, and use the calculated mean value and standard deviation as the distribution of the matching degree distribution between the voiceprint feature of the voice data of the impostor speaker and the voiceprint model of the target speaker mean and standard deviation; The said matching degree to be adjusted is zero-adjusted, and the matching degree obtained after adjustment includes: Using the mean value and standard deviation to zero-adjust the matching degree to be adjusted to obtain the adjusted matching degree. 6. A system for updating a voiceprint model, comprising: The update period acquisition module is used to obtain the current login time of the target speaker and the last update time of the target speaker's voiceprint model; The division module is used to divide the time period from the last update time of the target speaker's voiceprint model to the current login time of the target speaker into multiple time clusters; Voice data acquisition module, used to acquire the voice data when the target speaker's voiceprint authentication is successful in each time cluster; The model update data acquisition module is used to select voice data from the voice data successfully authenticated in each time cluster, and use the selected voice data as the voiceprint model update data of the target speaker; The training module is used to re-train the voiceprint model by using the update data of the voiceprint model of the target speaker and the training data of the original voiceprint model to obtain a new voiceprint model after training; A model updating module, configured to update the voiceprint model of the target speaker by using the new voiceprint model. 7. The system according to claim 6, wherein the model update data acquisition module comprises: The obtaining unit is used to obtain, for each time cluster, the matching degree of the voice data successfully authenticated in the time cluster relative to the voiceprint model of the target speaker; The filtering unit is used to filter out voice data corresponding to a matching degree greater than a set first threshold; or sort each piece of voice data in descending order of matching degree, and filter out a set number of voice data. 8. The system according to claim 7, wherein the model update data acquisition module further comprises: A sampling unit, configured to sample the speech data screened out by the screening unit in each time cluster. 9. The system according to any one of claims 6 to 8, wherein the system further comprises: The receiving module is used to obtain the voice data of the current user when performing voiceprint authentication on the user; The matching degree calculation module is used to calculate the matching degree between the voiceprint feature of the voice data of the current user and the target speaker's voiceprint model, and use the calculated matching degree as the matching degree to be regularized; A regularizing module, configured to perform zero regularization on the matching degree to be regularized to obtain the regularized matching degree; A judging module, configured to determine whether the current user is the target speaker according to the adjusted matching degree, and determine that the current user is the target speaker if the adjusted matching degree is greater than a set second threshold. 10. The system according to claim 9, further comprising: The seed data acquisition module is used to pre-collect a large amount of speech data from different speakers as seed data and put them into the seed collection; The first calculation module is used to calculate the matching degree between the voiceprint feature of each piece of speech data in the seed set and the voiceprint model of the target speaker, to obtain a matching degree set; The second calculation module is used to calculate the mean value and standard deviation of all matching degrees in the matching degree set, and use the calculated mean value and standard deviation as the voiceprint feature of the voice data of the impostor speaker and the target speaker The mean and standard deviation of the matching degree distribution of the voiceprint model; The normalizing module is specifically configured to use the mean value and standard deviation to perform zero normalization on the matching degree to be normalized to obtain the normalized matching degree.