CN110379433A - Method, apparatus, computer equipment and the storage medium of authentication - Google Patents

Abstract

The application relates to an identity verification method, device, computer equipment and storage medium. The method includes: acquiring the voice data input by the target user according to the target dynamic verification code; dividing the voice data into at least one voice frame according to a preset segmentation algorithm; and extracting each voice frame according to a preset acoustic feature extraction algorithm. The acoustic feature vector corresponding to the speech frame; the acoustic feature vector corresponding to the speech frame is input to the pre-trained identity verification multi-task model, and the intermediate user feature vector and the first posterior probability set corresponding to the speech frame are output; according to each speech The intermediate user feature vector corresponding to the frame and the preset pooling algorithm determine the first user feature vector corresponding to the target user; according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame, the The target user is authenticated. By adopting the application, the computing complexity of the server can be reduced, and the processing efficiency of the server can be improved.

Description

Identity verification method, device, computer equipment and storage medium

technical field

The present application relates to the field of security technology, in particular to an identity verification method, device, computer equipment and storage medium.

Background technique

At present, identity verification technologies based on biometrics (such as fingerprints, faces, voices, etc.) and identity verification technologies based on dynamic verification codes are two commonly used identity verification technologies. In order to further improve the security of identity verification, a combination of voice-based identity verification technology and dynamic verification code can be used to verify user identity.

In the traditional combination method, the user terminal collects the voice data input by the user according to the dynamic verification code through the voice collection device, and sends the voice data to the server deployed with the voiceprint recognition model and the voice recognition model. After receiving the voice data, the server may input the voice data into the voiceprint recognition model, and output the target user feature vector corresponding to the user. At the same time, the server can also input the speech data into the speech recognition model, and output the target text corresponding to the speech data. If the target user feature vector is similar to the pre-stored user feature vector of the user, and the target text is the same as the dynamic verification code, the server can determine that the user is a legitimate user. Otherwise, the server may determine that the user is an illegal user.

Based on the traditional combination method, the server needs two sets of voiceprint recognition models and speech recognition models with different deployment structures and parameters. The voiceprint recognition model and the speech recognition model need to independently process the user's voice data to obtain the user's feature vector and target text, resulting in high computational complexity of the server and low processing efficiency of the server.

Contents of the invention

Based on this, it is necessary to provide an identity verification method, device, computer equipment and storage medium for the above technical problems.

In a first aspect, a method for identity verification is provided, the method comprising:

Obtain the voice data input by the target user according to the target dynamic verification code;

dividing the speech data into at least one speech frame according to a preset segmentation algorithm;

For each speech frame, extract an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm;

The acoustic feature vector corresponding to the voice frame is input to the pre-trained identity verification multi-task model, and the intermediate user feature vector and the first posterior probability set corresponding to the voice frame are output, and the first posterior probability set includes each preset The posterior probability corresponding to the pronunciation unit;

According to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm, determine the first user feature vector corresponding to the target user;

Perform identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame.

As an optional implementation, the identity verification multi-task model includes a multi-task shared hidden layer, a voiceprint recognition network and a speech recognition network;

The acoustic feature vector corresponding to the voice frame is input to the pre-trained identity verification multi-task model, and the intermediate user feature vector and the first posterior probability set corresponding to the voice frame are output, including:

The acoustic feature vector corresponding to the speech frame is input to the multi-task sharing hidden layer, and the intermediate feature vector corresponding to the speech frame is output;

The intermediate feature vector corresponding to the speech frame is input to the speech recognition network, and the pronunciation feature vector and the first posterior probability set corresponding to the speech frame are output;

The intermediate feature vector and pronunciation feature vector corresponding to the speech frame are input to the voiceprint recognition network, and the intermediate user feature vector corresponding to the speech frame is output.

As an optional implementation manner, the authentication of the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame includes:

According to the first posterior probability set corresponding to each speech frame, determine the target dynamic verification code score corresponding to the target user;

If the similarity between the first user feature vector and the pre-stored second user feature vector corresponding to the target user is greater than or equal to a preset similarity threshold, and the target dynamic verification code score is greater than or equal to a preset dynamic verification code score threshold, then determine that the target user is a legitimate user;

If the similarity between the first user feature vector and the second user feature vector is smaller than the preset similarity threshold, or the target dynamic verification code score is smaller than the preset dynamic verification code score threshold, then It is determined that the target user is an illegal user.

As an optional implementation manner, the determining the target dynamic verification code score corresponding to the target user according to the first posterior probability set corresponding to each speech frame includes:

Acquiring the pronunciation unit sequence corresponding to the target dynamic verification code;

Determine the target pronunciation unit corresponding to each voice frame according to the first posterior probability set corresponding to each voice frame, the pronunciation unit sequence and the preset forced alignment algorithm;

For each speech frame, in the first posterior probability set corresponding to the speech frame, determine the posterior probability of the target pronunciation unit corresponding to the speech frame, and determine the posterior probability of the target pronunciation unit and the pre-stored target The product of the prior probability of the pronunciation unit is used as the likelihood value of the target pronunciation unit;

The target dynamic verification code score corresponding to the target user is determined according to the likelihood value of the target pronunciation unit corresponding to each speech frame.

As an optional implementation manner, the acquiring the pronunciation unit sequence corresponding to the target dynamic verification code includes:

Determine the word set corresponding to the target dynamic verification code according to the target dynamic verification code and a preset word segmentation algorithm;

For each word in the word set, determine the pronunciation unit sequence corresponding to the word according to the correspondence between the pre-stored word and the pronunciation unit sequence;

The pronunciation unit sequence corresponding to each word is sorted according to the order of each word in the target dynamic verification code to obtain the pronunciation unit sequence corresponding to the target dynamic verification code.

As an optional implementation manner, the determining the target dynamic verification code score corresponding to the target user according to the likelihood value of the target pronunciation unit corresponding to each speech frame includes:

For each speech frame, determine the difference between the likelihood value of the target pronunciation unit corresponding to the speech frame and the likelihood value of each preset pronunciation unit corresponding to the speech frame, as the speech frame The corresponding dynamic verification code score;

Determine the average value of the dynamic verification code scores corresponding to the speech frames as the target dynamic verification code score corresponding to the target user.

As an optional implementation, the method also includes:

Obtaining a pre-stored first training sample set, the first training sample set including a plurality of sample user IDs and first sample voice data corresponding to each sample user ID;

For each first sample speech data in the first training sample set, divide the first sample speech data into at least one first sample speech frame according to a preset segmentation algorithm;

For each first sample speech frame corresponding to the first sample speech data, extract an acoustic feature vector corresponding to the first sample speech frame according to a preset acoustic feature extraction algorithm;

The acoustic feature vectors of each first sample speech frame corresponding to the first sample speech data are input to the identity verification multi-task model to be trained, and the second posterior probability set corresponding to the first sample speech data is output, so The second posterior probability set includes the posterior probability corresponding to each sample user identifier;

Determining a first cost function corresponding to the first training sample set according to the posterior probability of the sample user identification corresponding to each first sample voice data;

According to the first cost function and the preset first parameter update algorithm, update the parameters corresponding to the multi-task shared hidden layer and the parameters corresponding to the voiceprint recognition network in the identity verification multi-task model to be trained Parameters corresponding to the speech recognition network.

As an optional implementation, the method also includes:

Obtaining a pre-stored second training sample set, the second training sample set including a plurality of second sample speech frames and sample pronunciation units corresponding to each second sample speech frame;

For each second sample speech frame in the second training sample set, extract an acoustic feature vector corresponding to the second sample speech frame according to the preset acoustic feature extraction algorithm;

The acoustic feature vector corresponding to the second sample speech frame is input to the identity verification multi-task model to be trained, and the third posterior probability set corresponding to the second sample speech frame is output, and the third posterior probability set includes each sample The posterior probability corresponding to the pronunciation unit;

Determining a second cost function corresponding to the second training sample set according to the posterior probability of the sample pronunciation unit corresponding to each second sample speech frame;

According to the second cost function and the preset second parameter update algorithm, update the parameters corresponding to the multi-task shared hidden layer and the parameters corresponding to the speech recognition network in the identity verification multi-task model to be trained.

As an optional implementation, the method also includes:

Obtaining a plurality of pre-stored verification sample sets, each verification sample set including a plurality of sample user IDs and second sample voice data corresponding to each sample user ID;

For each second sample speech data in each verification sample set, divide the second sample speech data into at least one third sample speech frame according to a preset segmentation algorithm;

For each third sample speech frame corresponding to the second sample speech data, extract an acoustic feature vector corresponding to the third sample speech frame according to a preset acoustic feature extraction algorithm;

Input the acoustic feature vectors of the third sample speech frames corresponding to the second sample speech data to the identity verification multi-task model to be verified, and output the fourth posterior probability set corresponding to the second sample speech data, the fourth The posterior probability set includes the posterior probability corresponding to each sample user identifier;

If the posterior probability of the sample user identification corresponding to the second sample voice data is the maximum value in the fourth posterior probability set corresponding to the second sample voice data, then determine that the second sample voice data is the target sample voice data;

Determine the ratio of the number of target sample speech data in the verification sample set to the total number of the second sample speech data in the verification sample set, as the accuracy rate of the verification sample set;

According to the accuracy rate of each verification sample set, determine the change rate of the accuracy rate corresponding to each verification sample set, if there is a continuous preset number of verification sample sets The change rate corresponding to the accuracy rate is less than or equal to the preset change rate threshold, then It is determined that the identity verification multi-task model to be verified has been trained.

In a second aspect, an identity verification device is provided, and the device includes:

The first obtaining module is used to obtain the voice data input by the target user according to the target dynamic verification code;

A first division module, configured to divide the voice data into at least one voice frame according to a preset segmentation algorithm;

The first extraction module is used for extracting the acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm for each speech frame;

The first output module is used to input the acoustic feature vector corresponding to the voice frame to the pre-trained identity verification multi-task model, and output the intermediate user feature vector and the first posterior probability set corresponding to the voice frame, the first posterior probability set The posterior probability set includes the posterior probability corresponding to each preset pronunciation unit;

The first determination module is used to determine the first user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and a preset pooling algorithm;

A verification module, configured to perform identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame.

In a third aspect, a computer device is provided, including a memory and a processor, the memory stores a computer program that can run on the processor, and the processor implements the following steps when executing the computer program:

Obtain the voice data input by the target user according to the target dynamic verification code;

dividing the speech data into at least one speech frame according to a preset segmentation algorithm;

For each speech frame, extract an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm;

The acoustic feature vector corresponding to the voice frame is input to the pre-trained identity verification multi-task model, and the intermediate user feature vector and the first posterior probability set corresponding to the voice frame are output, and the first posterior probability set includes each preset The posterior probability corresponding to the pronunciation unit;

According to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm, determine the first user feature vector corresponding to the target user;

Perform identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame.

In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

Obtain the voice data input by the target user according to the target dynamic verification code;

dividing the speech data into at least one speech frame according to a preset segmentation algorithm;

For each speech frame, extract an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm;

The acoustic feature vector corresponding to the voice frame is input to the pre-trained identity verification multi-task model, and the intermediate user feature vector and the first posterior probability set corresponding to the voice frame are output, and the first posterior probability set includes each preset The posterior probability corresponding to the pronunciation unit;

According to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm, determine the first user feature vector corresponding to the target user;

Perform identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame.

Embodiments of the present application provide an identity verification method, device, computer equipment, and storage medium. The server acquires the voice data input by the target user according to the target dynamic verification code, and divides the voice data into at least one voice frame according to a preset segmentation algorithm. Then, for each speech frame, the server determines an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm. Afterwards, the server inputs the acoustic feature vector corresponding to the speech frame into the pre-trained identity verification multi-task model, and outputs the intermediate user feature vector and the first posterior probability set corresponding to the speech frame. Wherein, the first posterior probability set includes the posterior probability corresponding to each preset pronunciation unit. Finally, the server determines the first user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm, and according to the first user feature vector corresponding to the target user and the first user feature vector corresponding to each speech frame A set of posterior probabilities for authenticating the target user. In this way, the server does not need to deploy two sets of voiceprint recognition models and speech recognition models with different structures and parameters, but only needs to deploy a set of authentication multi-tasking models to process the user's voice data, thereby reducing the computational complexity of the server , improve the processing efficiency of the server.

Description of drawings

Fig. 1 is the architectural diagram of a kind of identity verification system provided by the embodiment of the present application;

FIG. 2 is a flow chart of a method for identity verification provided by an embodiment of the present application;

FIG. 3 is a schematic structural diagram of an identity verification multi-task model provided by an embodiment of the present application;

FIG. 4 is a flow chart of an identity verification method provided in an embodiment of the present application;

FIG. 5 is a flow chart of a method for determining a target dynamic verification code score provided by an embodiment of the present application;

FIG. 6 is a flow chart of an account registration method provided in an embodiment of the present application;

FIG. 7 is a flow chart of a training method for an identity verification multi-task model provided in an embodiment of the present application;

FIG. 8 is a flow chart of a verification method for an identity verification multi-task model provided in an embodiment of the present application;

FIG. 9 is a schematic structural diagram of an identity verification device provided by an embodiment of the present application;

FIG. 10 is a schematic structural diagram of a computer device provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, and are not intended to limit the present application.

The embodiment of the present application provides an identity verification method, which can be applied to an identity verification system. FIG. 1 is an architecture diagram of an identity verification system provided by an embodiment of the present application. As shown in Figure 1, the identity verification system includes a user terminal and a server connected through a communication network. Wherein, the communication network can be a wired network, a wireless network, or other types of communication networks, which are not limited in this embodiment of the application; the user terminal can be a portable electronic device with recording and computing functions, such as a mobile phone, Tablets, laptops, etc.

The user terminal is used to receive the user's account registration request and identity verification request, generate a dynamic verification code corresponding to the user, collect voice data input by the user according to the dynamic verification code, and send the user's corresponding account number, dynamic verification code and voice data The data is sent to the server; the server is used to train and verify the identity verification multi-task model to be trained; during the account registration process, the server is also used to divide the voice data into at least one voice frame according to a preset segmentation algorithm , input the acoustic feature vectors corresponding to each speech frame into the identity verification multi-task model, output the intermediate user feature vectors corresponding to each speech frame, and obtain the user The corresponding target user feature vector, and store the account corresponding to the user and the target user feature vector; in the identity verification process, the server is also used to divide the voice data into at least one voice frame according to the preset segmentation algorithm, and divide the voice data into at least one voice frame. The acoustic feature vector corresponding to each speech frame is input to the identity verification multi-task model, and the intermediate user feature vector and the first posterior probability set corresponding to each speech frame are output. According to the intermediate user feature vector corresponding to each speech frame and the preset pooling Algorithm, obtain the target user feature vector corresponding to the user, obtain the target dynamic verification code score according to the first posterior probability set corresponding to each speech frame, and use the target user feature vector and target dynamic verification code score corresponding to the user for the The user is authenticated.

As an optional implementation manner, the server may also send the identity verification multitasking model to the user terminal, and the user terminal processes the voice data input by the user based on the identity verification multitasking model. In the embodiment of the present application, the server processes the voice data input by the user based on the identity verification multi-task model as an example, and other situations are similar.

A user authentication method provided in the embodiment of the present application will be described in detail below in combination with specific implementation manners. As shown in Figure 2, the specific steps are as follows:

Step 201, acquire voice data input by the target user according to the target dynamic verification code.

In implementation, when a certain user (namely the target user) logs in the user terminal using the target account corresponding to the target user, the user terminal may generate a target dynamic verification code corresponding to the target account. Wherein, the user terminal can randomly select a candidate dynamic verification code from the pre-stored candidate dynamic verification code set as the target dynamic verification code corresponding to the target account; Select a preset number of candidate words, and randomly combine the selected preset number of candidate words as the target dynamic verification code corresponding to the target account; the user terminal can also use other methods to generate the target dynamic verification code corresponding to the target account. The verification code is not limited in the embodiment of this application. After the user terminal generates the target dynamic verification code corresponding to the target account, the target dynamic verification code may be displayed on the display interface. As an optional implementation manner, the user terminal may also display prompt information on the display interface for prompting the target user to read the target dynamic verification code aloud. Then, the user terminal can activate a voice collection device (such as a microphone) to collect voice data when the target user reads the target dynamic verification code aloud. After the user terminal obtains the voice data input by the target user according to the target dynamic verification code, it can send an identity verification request to the server. Wherein, the identity verification request carries the target account number, the target dynamic verification code and voice data. After receiving the identity verification request, the server may analyze the identity verification request to obtain the target account number, target dynamic verification code and voice data carried in the identity verification request.

Step 202: Divide the voice data into at least one voice frame according to a preset segmentation algorithm.

A preset segmentation algorithm may be pre-stored in the server. Wherein, the segmentation algorithm may be a frame segmentation and windowing algorithm, or may be another type of segmentation algorithm, which is not limited in this embodiment of the present application. After the server obtains the voice data, it can divide the voice data into at least one voice frame according to a pre-stored segmentation algorithm. Wherein, the speech frame is a short-term speech segment with a duration of about 25 milliseconds.

Step 203, for each speech frame, according to a preset acoustic feature extraction algorithm, extract the corresponding acoustic feature vector of the speech frame.

In implementation, an acoustic feature extraction algorithm may be pre-stored in the server. For each speech frame, the server may extract the corresponding acoustic feature vector of the speech frame according to a preset acoustic feature extraction algorithm. Wherein, the acoustic feature vector may be a Mel cepstral coefficient or the like.

Step 203: Input the acoustic feature vector corresponding to the speech frame into the pre-trained identity verification multi-task model, and output the intermediate user feature vector and the first posterior probability set corresponding to the speech frame.

Wherein, the first posterior probability set includes the posterior probability corresponding to each preset pronunciation unit.

In implementation, a pre-trained identity verification multi-task model may be pre-stored in the server. For each speech frame in the speech frame set, after the server obtains the acoustic feature vector corresponding to the speech frame, the acoustic feature vector corresponding to the speech frame may be input into the identity verification multi-task model. The identity verification multi-task model can output the intermediate user feature vector and the first posterior probability set corresponding to the voice frame. Wherein, the first posterior probability set includes the posterior probability corresponding to each preset pronunciation unit in the identity verification multi-task model.

FIG. 3 is a schematic structural diagram of an identity verification multi-task model provided by an embodiment of the present application. As shown in Figure 3, the identity verification multi-task model includes multi-task sharing hidden layer, voiceprint recognition network and speech recognition network. Wherein, for each speech frame in the speech frame set, the server inputs the acoustic feature vector corresponding to the speech frame to the pre-trained identity verification multi-task model, and outputs the intermediate user feature vector and the first posterior probability corresponding to the speech frame Collections are processed as follows:

Step 1: Input the acoustic feature vector corresponding to the speech frame to the multi-task sharing hidden layer, and output the intermediate feature vector corresponding to the speech frame.

In an implementation, for each speech frame in the speech frame set, the server may input the acoustic feature vector corresponding to the speech frame into the multi-task sharing hidden layer in the identity verification multi-task model. The multi-task shared hidden layer can output the intermediate feature vector corresponding to the speech frame.

Step 2: Input the intermediate feature vector corresponding to the speech frame to the speech recognition network, and output the pronunciation feature vector and the first posterior probability set corresponding to the speech frame.

In implementation, after the server obtains the intermediate feature vector corresponding to the speech frame, it may input the intermediate feature vector corresponding to the speech frame to the speech recognition network. The speech recognition network can output the pronunciation feature vector and the first posterior probability set corresponding to the speech frame.

Step 3: Input the intermediate feature vector and pronunciation feature vector corresponding to the speech frame to the voiceprint recognition network, and output the intermediate user feature vector corresponding to the speech frame.

In implementation, after the server obtains the intermediate feature vector and the pronunciation feature vector corresponding to the speech frame, the server may input the intermediate feature vector and the pronunciation feature vector corresponding to the speech frame into the voiceprint recognition network. The voiceprint recognition network can output the intermediate user feature vector corresponding to the voice frame.

In the embodiment of the present application, compared with the traditional voiceprint recognition model, the server jointly inputs the intermediate feature vector corresponding to the speech frame output by the multi-task sharing hidden layer and the pronunciation feature vector corresponding to the speech frame output by the speech recognition network into the voiceprint recognition model. The fingerprint recognition network improves the accuracy of the user feature vector corresponding to the voice frame output by the voiceprint recognition network.

Step 205: Determine the first user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm.

In implementation, a pooling algorithm may be pre-stored in the server. After obtaining the intermediate user feature vectors corresponding to each speech frame, the server may perform pooling processing on the intermediate user feature vectors corresponding to each speech frame according to a preset pooling algorithm to obtain the first user feature vector corresponding to the target user.

Step 206: Perform identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame.

In implementation, after the server obtains the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame, it can use the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame A collection of probabilities to authenticate the target user. In this way, the server does not need to deploy two sets of voiceprint recognition models and speech recognition models with different structures and parameters, but only needs to deploy a set of authentication multi-tasking models to process the user's voice data, thereby reducing the computational complexity of the server , improve the processing efficiency of the server.

Wherein, as shown in Figure 4, the server performs identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame as follows:

Step 401: Determine the target dynamic verification code score corresponding to the target user according to the first posterior probability set corresponding to each speech frame.

In implementation, after obtaining the first posterior probability set corresponding to each voice frame, the server may further determine the target dynamic verification code score corresponding to the target user according to the first posterior probability set corresponding to each voice frame. After the server obtains the first user feature vector and the target dynamic verification code score corresponding to the target user, it can further determine the second user feature vector corresponding to the pre-stored target user (that is, the target account number carried in the identity verification request). Whether the user feature vectors are the same, and whether the target dynamic verification code score is greater than or equal to a preset dynamic verification code score threshold. If the similarity between the first user feature vector and the second user feature vector corresponding to the pre-stored target user is greater than or equal to the preset similarity threshold, and the target dynamic verification code score is greater than or equal to the preset dynamic verification code score threshold, Then step 402 is executed. If the similarity between the first user feature vector and the second user feature vector is smaller than the preset similarity threshold, or the target dynamic verification code score is smaller than the preset dynamic verification code score threshold, step 403 is executed. Wherein, the similarity may be Euclidean similarity, cosine similarity, or other types of similarity, which are not limited in this embodiment of the present application.

Figure 5 is a flow chart of a method for determining the target dynamic verification code score provided by the embodiment of the present application. As shown in Figure 5, the server determines the target dynamic verification code corresponding to the target user according to the first posterior probability set corresponding to each voice frame Code fractions are processed as follows:

Step 501, acquire the pronunciation unit sequence corresponding to the target dynamic verification code.

In implementation, after receiving the identity verification request, the server may parse the identity verification request to obtain the target dynamic verification code carried in the identity verification request. Then, the server may further acquire the pronunciation unit sequence corresponding to the target dynamic verification code. Wherein, the process of the server obtaining the pronunciation unit sequence corresponding to the target dynamic verification code is as follows:

Step 1: Determine the word set corresponding to the target dynamic verification code according to the target dynamic verification code and the preset word segmentation algorithm.

In implementation, word segmentation algorithms may be pre-stored in the server. After obtaining the target dynamic verification code, the server may perform word segmentation processing on the target dynamic verification code according to a preset word segmentation algorithm to obtain a word set corresponding to the target dynamic verification code. For example, the target dynamic verification code is "Tsinghua University", and the server performs word segmentation processing on the target dynamic verification code to obtain the word set {"Tsinghua", "University"} corresponding to the target dynamic verification code.

Step 2, for each word in the word set, according to the pre-stored correspondence between the word and the pronunciation unit sequence, determine the pronunciation unit sequence corresponding to the word.

In an implementation, the correspondence relationship between words and pronunciation unit sequences (also called a pronunciation dictionary) may be pre-stored in the server. Table 1 shows the correspondence between words and pronunciation unit sequences pre-stored in the server, as shown in Table 1.

Table I

serial number word articulation unit sequence 1 Beijing b ei3 j ing1 2 Tsinghua University q ing1 h ua2 3 the University d a4 x ve2 4 aerospace h ang2 t ian1

After the server obtains the word set corresponding to the target dynamic verification code, for each word in the word set, the server can determine the pronunciation unit sequence corresponding to the word according to the pre-stored correspondence between the word and the pronunciation unit sequence.

Step 3: sort the pronunciation unit sequence corresponding to each word according to the order of each word in the target dynamic verification code, and obtain the pronunciation unit sequence corresponding to the target dynamic verification code.

In implementation, after the server obtains the pronunciation unit sequence corresponding to each word, it can sort the pronunciation unit sequence corresponding to each word according to the order of each word in the target dynamic verification code, and obtain the pronunciation unit sequence corresponding to the target dynamic verification code. For example, the target dynamic verification code is "Tsinghua University", and the pronunciation unit sequence corresponding to the target dynamic verification code is "q ing1 h ua2 d a4 x ve2".

Step 502: Determine the target pronunciation unit corresponding to each speech frame according to the first posterior probability set corresponding to each speech frame, the pronunciation unit sequence and the preset forced alignment algorithm.

In implementation, the mandatory alignment algorithm may be pre-stored in the server. The forced alignment algorithm may be a Viterbi algorithm, or other types of forced alignment algorithms, which are not limited in this embodiment of the present application. After the server obtains the first posterior probability set corresponding to each speech frame and the pronunciation unit sequence corresponding to the target dynamic verification code, it can forcibly align each speech frame with the pronunciation unit sequence, that is, to obtain the position of each speech frame in the pronunciation unit sequence The corresponding start time and end time, so as to obtain the target pronunciation unit corresponding to each speech frame.

Step 503, for each speech frame, in the first posterior probability set corresponding to the speech frame, determine the posterior probability of the target pronunciation unit corresponding to the speech frame, and determine the posterior probability of the target pronunciation unit and the pre-stored The product of the prior probability of the target pronunciation unit is taken as the likelihood value of the target pronunciation unit.

In an implementation, prior probabilities of each preset pronunciation unit may be pre-stored in the server. After the server obtains the target pronunciation unit corresponding to each speech frame, for each speech frame, the server may determine the posterior probability of the target pronunciation unit corresponding to the speech frame in the first posterior probability set corresponding to the speech frame. Then, the server may determine the prior probability of the target pronunciation unit from the pre-stored prior probability of each preset pronunciation unit. Afterwards, the server may calculate the product of the posterior probability of the target pronunciation unit and the prior probability of the target pronunciation unit to obtain the likelihood value of the target pronunciation unit.

Step 504: Determine the target dynamic verification code score corresponding to the target user according to the likelihood value of the target pronunciation unit corresponding to each speech frame.

In implementation, after obtaining the likelihood value of the target pronunciation unit corresponding to each speech frame, the server may determine the target dynamic verification code score corresponding to the target user according to the likelihood value of the target pronunciation unit corresponding to each speech frame. Wherein, the server determines the target dynamic verification code score corresponding to the target user according to the likelihood value of the target pronunciation unit corresponding to each voice frame as follows:

Step 1, for each speech frame, determine the difference between the likelihood value of the target pronunciation unit corresponding to the speech frame and the likelihood value of each preset pronunciation unit corresponding to the speech frame, as the difference of the maximum likelihood value in the speech frame The dynamic verification code score corresponding to the frame.

In implementation, after the server obtains the likelihood value of the target pronunciation unit corresponding to the speech frame, the server may determine the maximum likelihood value among the likelihood values of the preset pronunciation units corresponding to the speech frame. Then, for each speech frame, the server may calculate the difference between the likelihood value of the target pronunciation unit corresponding to the speech frame and the maximum likelihood value, to obtain the dynamic verification code score corresponding to the speech frame.

Step 2: Determine the average value of the dynamic verification code scores corresponding to each voice frame, and use it as the target dynamic verification code score corresponding to the target user.

In implementation, after obtaining the dynamic verification code scores corresponding to each voice frame, the server may calculate an average value of the dynamic verification code scores corresponding to each voice frame, as the target dynamic verification code score corresponding to the target user.

The traditional speech recognition model uses the decoding method to output the text content corresponding to the speech data. Because the decoding method needs to consider the relationship between the acoustic feature vector corresponding to the speech data and the pronunciation unit, the pronunciation unit and the word, and the word and the word. , and it is necessary to perform a decoding operation in a large candidate vocabulary to obtain the most likely text content corresponding to the speech data, which leads to a high computational complexity of the traditional speech recognition model. In the embodiment of the present application, after obtaining the posterior probability of each preset pronunciation unit corresponding to each voice frame, the server determines the target pronunciation unit corresponding to each voice frame according to the pronunciation unit sequence corresponding to the dynamic verification code and the forced alignment algorithm, and determines The likelihood value of the target pronunciation unit corresponding to each speech frame. Then, the server can determine the target dynamic verification code score according to the likelihood value of the target pronunciation unit corresponding to each speech frame, thereby reducing the computational complexity.

Step 402, determining that the target user is a legitimate user.

In implementation, if the similarity between the first user feature vector and the second user feature vector corresponding to the pre-stored target user is greater than or equal to the preset similarity threshold, and the target dynamic verification code score is greater than or equal to the preset dynamic verification code score threshold, the server can determine that the target user is a legitimate user. Then, the server may send an authentication success response to the user terminal. After receiving the identity verification success response, the user terminal allows the target user to use the target account to log in to the user terminal.

Step 403, determining that the target user is an illegal user.

In implementation, if the similarity between the first user feature vector and the second user feature vector is less than the preset similarity threshold, or the target dynamic verification code score is less than the preset dynamic verification code score threshold, the server determines that the target user is illegal user. Then, the server may send an authentication failure response to the user terminal. After receiving the identity verification failure response, the user terminal rejects the target user's use of the target account to log in to the user terminal.

The embodiment of this application provides an identity verification method. The server acquires the voice data input by the target user according to the target dynamic verification code, and divides the voice data into at least one voice frame according to a preset segmentation algorithm. Then, for each speech frame, the server determines an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm. Afterwards, the server inputs the acoustic feature vector corresponding to the speech frame into the pre-trained identity verification multi-task model, and outputs the intermediate user feature vector and the first posterior probability set corresponding to the speech frame. Wherein, the first posterior probability set includes the posterior probability corresponding to each preset pronunciation unit. Finally, the server determines the first user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm, and according to the first user feature vector corresponding to the target user and the first user feature vector corresponding to each speech frame A set of posterior probabilities for authenticating the target user. In this way, the server does not need to deploy two sets of voiceprint recognition models and speech recognition models with different structures and parameters, but only needs to deploy a set of authentication multi-tasking models to process the user's voice data, thereby reducing the computational complexity of the server , improve the processing efficiency of the server.

The embodiment of the present application also provides a method for account registration, as shown in Figure 6, the specific processing process is as follows:

Step 601, acquire voice data input by the target user according to the target dynamic verification code.

In implementation, when a certain user (i.e., the target user) creates a target account corresponding to the target user's login user terminal, the target user can input the target account of the target user in the account input box in the account registration interface of the user terminal. account. Then, the user terminal may generate a target dynamic verification code corresponding to the target account. Wherein, the processing process of the user terminal generating the target dynamic verification code corresponding to the target account is similar to the processing process of the user terminal generating the target dynamic verification code corresponding to the target account in step 201, and will not be repeated here. After the user terminal generates the target dynamic verification code corresponding to the target account, it can display the target dynamic verification code on the display interface. As an optional implementation manner, the user terminal may also display prompt information on the display interface for prompting the target user to read the target dynamic verification code aloud. Then, the user terminal can activate a voice collection device (such as a microphone) to collect voice data when the target user reads the target dynamic verification code aloud. After the user terminal obtains the voice data input by the target user according to the target dynamic verification code, it can send an account registration request to the server. Wherein, the account registration request carries the target account, the target dynamic verification code and voice data. After receiving the account registration request, the server may analyze the account registration request to obtain the target account number, target dynamic verification code and voice data carried in the account registration request.

Step 602: Divide the voice data into at least one voice frame according to a preset segmentation algorithm.

Step 603, for each speech frame, extract an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm.

Step 604: Input the acoustic feature vector corresponding to the speech frame into the pre-trained identity verification multi-task model, and output the intermediate user feature vector corresponding to the speech frame.

Step 605: Determine the target user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm.

In practice, the processing from step 602 to step 605 is similar to the processing in which the server determines the first user feature vector corresponding to the target user in the steps of the identity verification method described above, and will not be repeated here.

Step 606, storing the corresponding relationship between the target account and the feature vector of the target user.

In implementation, after the server obtains the target user feature vector corresponding to the target user, it may store the corresponding relationship between the target account corresponding to the target user and the target user feature vector locally, so as to authenticate the target user.

The embodiment of the present application also provides a training method for an identity verification multi-task model, as shown in FIG. 7 , and the specific processing process is as follows:

Step 701, initialize the identity verification multi-task model to be trained.

In implementation, the identity verification multi-task model to be trained may be stored in advance in the server. When the server needs to train the identity verification multi-task model to be trained, the server can randomly initialize parameters in the identity verification multi-task model to be trained.

Step 702A, obtain a pre-stored first training sample set.

Wherein, the first training sample set includes a plurality of sample user identifiers and first sample speech data corresponding to each sample user identifier.

Step 702B, acquiring a pre-stored second training sample set.

Wherein, the second training sample set includes a plurality of second sample speech frames and a sample pronunciation unit corresponding to each second sample speech frame.

In an implementation, multiple first training sample sets and multiple second training sample sets may be pre-stored in the server. Wherein, the first training sample set includes a plurality of sample user IDs and first sample voice data corresponding to each sample user ID; the second training sample set includes a plurality of second sample voice frames and each second sample voice frame corresponds to A sample pronunciation unit of . After the server initializes the parameters in the identity verification multi-task model to be trained, it may acquire a first preset number of first training sample sets and a second preset number of second training sample sets stored in advance.

Step 703: For each first sample speech data in the first training sample set, divide the first sample speech data into at least one first sample speech frame according to a preset segmentation algorithm.

In implementation, the processing process of step 703 is similar to the processing process of step 202, and will not be repeated here.

Step 704A, for each first sample speech frame corresponding to the first sample speech data, extract an acoustic feature vector corresponding to the first sample speech frame according to a preset acoustic feature extraction algorithm.

Step 704B, for each second sample speech frame in the second training sample set, extract an acoustic feature vector corresponding to the second sample speech frame according to a preset acoustic feature extraction algorithm.

In implementation, the processing procedures of step 704A and step 704B are similar to the processing procedure of step 203, and will not be repeated here.

Step 705A, input the acoustic feature vectors of each first sample speech frame corresponding to the first sample speech data into the identity verification multi-task model to be trained, and output the second posterior probability corresponding to the first sample speech data gather.

Wherein, the second posterior probability set includes the posterior probability corresponding to each sample user identifier.

In implementation, after the server obtains the acoustic feature vectors of each first sample speech frame corresponding to the first sample speech data, the acoustic feature vector of each first sample speech frame corresponding to the first sample speech data can be Input to the authentication multitasking model. The identity verification multi-task model can output a second posterior probability set corresponding to the first sample voice data. Wherein, the second posterior probability set includes the posterior probability corresponding to each sample user identifier.

Step 705B: Input the acoustic feature vector corresponding to the second sample speech frame into the identity verification multi-task model to be trained, and output the third posterior probability set corresponding to the second sample speech frame.

Wherein, the third posterior probability set includes the posterior probability corresponding to each sample pronunciation unit.

In an implementation, for each second sample speech frame, after the server obtains the acoustic feature vector corresponding to the second sample speech frame, the acoustic feature vector corresponding to the second sample speech frame may be input into the identity verification multi-task model. The identity verification multi-task model may output a third posterior probability set corresponding to the second sample speech frame. Wherein, the third posterior probability set includes the posterior probability corresponding to each sample pronunciation unit.

Step 706A: Determine the first cost function corresponding to the first training sample set according to the posterior probability of the sample user ID corresponding to each first sample speech data.

In an implementation, after the server obtains the second posterior probability set corresponding to each first sample speech data, for each first sample speech data, the server can calculate the second posterior probability set corresponding to the first sample speech data The posterior probability of the sample user identity corresponding to the first sample voice data is determined in the set. Then, the server may determine the first cost function corresponding to the first training sample set according to the posterior probability of the sample user identification corresponding to each first sample voice data.

Step 706B: Determine a second cost function corresponding to the second training sample set according to the posterior probability of the sample pronunciation unit corresponding to each second sample speech frame.

In implementation, after the server obtains the third posterior probability set corresponding to each second sample speech frame, for each second sample speech frame, the server can determine in the third posterior probability set corresponding to the second sample speech frame The posterior probability of the sample pronunciation unit corresponding to the second sample speech frame. Then, the server may determine the second cost function corresponding to the second training sample set according to the posterior probability of the sample pronunciation unit corresponding to each second sample speech frame.

Step 707A, according to the first cost function and the preset first parameter update algorithm, update the parameters corresponding to the multi-task shared hidden layer in the identity verification multi-task model to be trained, the parameters corresponding to the voiceprint recognition network and the corresponding parameters of the speech recognition network parameters.

In implementation, the first parameter update algorithm may be pre-stored in the server. The first parameter update algorithm may be a stochastic gradient descent method. After the server obtains the first cost function, it can update the algorithm according to the first cost function and the first parameter to update the parameters corresponding to the multi-task shared hidden layer in the identity verification multi-task model to be trained, the parameters corresponding to the voiceprint recognition network and the voice Identify the parameters corresponding to the network.

Step 707B, according to the second cost function and the preset second parameter update algorithm, update the parameters corresponding to the multi-task shared hidden layer and the parameters corresponding to the speech recognition network in the identity verification multi-task model to be trained.

In implementation, the second parameter update algorithm may be pre-stored in the server. The second parameter update algorithm may be a stochastic gradient descent method. After obtaining the second cost function, the server can update parameters corresponding to the multi-task shared hidden layer and parameters corresponding to the speech recognition network in the identity verification multi-task model to be trained according to the second cost function and the second parameter update algorithm.

The embodiment of the present application also provides a verification method for an identity verification multi-task model, as shown in FIG. 8 , and the specific processing process is as follows:

Step 801, obtaining a plurality of verification sample sets stored in advance.

Wherein, each verification sample set includes a plurality of sample user identifiers and each sample user identifier corresponds to the second sample voice data.

In implementation, multiple verification sample sets may be pre-stored in the server. Wherein, each verification sample set includes a plurality of sample user IDs and each sample user ID corresponds to a third sample speech frame, and the multiple sample user IDs are sample user IDs in the first training sample set. When the server needs to verify the identity verification multi-task model to be verified, the server can acquire multiple verification sample sets stored in advance.

Step 802, for each second sample speech data in each verification sample set, divide the second sample speech data into at least one third sample speech frame according to a preset segmentation algorithm.

In implementation, the processing process of step 802 is similar to the processing process of step 703, and will not be repeated here.

Step 803, for each third sample speech frame corresponding to the second sample speech data, extract an acoustic feature vector corresponding to the third sample speech frame according to a preset acoustic feature extraction algorithm.

In implementation, the processing process of step 803 is similar to the processing process of step 704A and step 704B, and will not be repeated here.

Step 804: Input the acoustic feature vectors of the third sample speech frames corresponding to the second sample speech data into the identity verification multi-task model to be verified, and output the fourth posterior probability set corresponding to the second sample speech data.

Wherein, the fourth posterior probability set includes the posterior probability corresponding to each sample user identifier.

In implementation, the processing process of step 804 is similar to the processing process of step 705A, and will not be repeated here.

Step 805, if the posterior probability of the sample user identity corresponding to the second sample voice data is the maximum value in the fourth posterior probability set corresponding to the second sample voice data, then determine that the second sample voice data is the target sample voice data.

In implementation, after the server obtains the fourth posterior probability set corresponding to the second sample voice data, it can determine whether the posterior probability of the sample user identifier corresponding to the second sample voice data is the largest in the fourth posterior probability set value. If the posterior probability of the sample user identifier corresponding to the second sample voice data is the maximum value, the server may determine the second sample voice data as the target sample voice data.

Step 806: Determine the ratio of the number of target sample speech data in the verification sample set to the total number of second sample speech data in the verification sample set as the accuracy rate of the verification sample set.

In implementation, for each verification sample set, after the server determines each target sample speech data in the verification sample set, it can further calculate the number of target sample speech data in the verification sample set and the number of the second sample speech data in the verification sample set. The ratio of the total number of speech data to obtain the accuracy rate of the verification sample set.

Step 807, according to the accuracy rate of each verification sample set, determine the rate of change of the accuracy rate corresponding to each verification sample set, if there are consecutive preset numbers of verification sample sets, the rate of change of the accuracy rate corresponding to the set is less than or equal to the preset rate of change threshold, it is determined that the identity verification multi-task model training to be verified is completed.

In implementation, after the server obtains the accuracy rate of each verification sample set, it can calculate the change rate corresponding to each verification sample set. If there are a preset number of consecutive verification sample sets whose change rates are less than or equal to the preset change rate threshold, the server may determine that the training of the identity verification multi-task model to be verified is completed.

The embodiment of the present application also provides an identity verification device, as shown in Figure 9, the device includes:

The first obtaining module 910 is used to obtain the voice data input by the target user according to the target dynamic verification code;

The first division module 920 is used to divide the voice data into at least one voice frame according to a preset segmentation algorithm;

The first extraction module 930 is used for extracting the acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm for each speech frame;

The first output module 940 is used to input the acoustic feature vector corresponding to the voice frame into the pre-trained identity verification multi-task model, output the intermediate user feature vector and the first posterior probability set corresponding to the voice frame, and the first posterior The probability set includes the posterior probability corresponding to each preset pronunciation unit;

The first determination module 950 is used to determine the first user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm;

The verification module 960 is configured to perform identity verification on the target user according to the first user feature vector corresponding to the target user and the first posterior probability set corresponding to each speech frame.

As an optional implementation, the identity verification multi-task model includes a multi-task shared hidden layer, a voiceprint recognition network and a speech recognition network;

The first output module 940 is specifically used for:

Input the acoustic feature vector corresponding to the speech frame to the multi-task sharing hidden layer, and output the intermediate feature vector corresponding to the speech frame;

The intermediate feature vector corresponding to the speech frame is input to the speech recognition network, and the pronunciation feature vector and the first posterior probability set corresponding to the speech frame are output;

The intermediate feature vector and pronunciation feature vector corresponding to the voice frame are input to the voiceprint recognition network, and the intermediate user feature vector corresponding to the voice frame is output.

As an optional implementation manner, the verification module 960 is specifically used for:

Determine the target dynamic verification code score corresponding to the target user according to the first posterior probability set corresponding to each speech frame;

If the similarity between the first user feature vector and the second user feature vector corresponding to the pre-stored target user is greater than or equal to the preset similarity threshold, and the target dynamic verification code score is greater than or equal to the preset dynamic verification code score threshold, Then determine that the target user is a legitimate user;

If the similarity between the first user feature vector and the second user feature vector is smaller than a preset similarity threshold, or the target dynamic verification code score is smaller than the preset dynamic verification code score threshold, it is determined that the target user is an illegal user.

As an optional implementation manner, the verification module 960 is specifically used for:

Obtain the pronunciation unit sequence corresponding to the target dynamic verification code;

According to the first posterior probability set corresponding to each speech frame, the pronunciation unit sequence and the preset forced alignment algorithm, determine the target pronunciation unit corresponding to each speech frame;

For each speech frame, in the first posterior probability set corresponding to the speech frame, determine the posterior probability of the target pronunciation unit corresponding to the speech frame, and determine the posterior probability of the target pronunciation unit and the pre-stored target The product of the prior probability of the pronunciation unit is used as the likelihood value of the target pronunciation unit;

According to the likelihood value of the target pronunciation unit corresponding to each speech frame, the target dynamic verification code score corresponding to the target user is determined.

As an optional implementation manner, the verification module 960 is specifically used for:

Determine the word set corresponding to the target dynamic verification code according to the target dynamic verification code and the preset word segmentation algorithm;

For each word in the word set, determine the pronunciation unit sequence corresponding to the word according to the correspondence between the pre-stored word and the pronunciation unit sequence;

The pronunciation unit sequence corresponding to each word is sorted according to the order of each word in the target dynamic verification code, and the pronunciation unit sequence corresponding to the target dynamic verification code is obtained.

As an optional implementation manner, the verification module 960 is specifically used for:

For each speech frame, determine the difference between the likelihood value of the target pronunciation unit corresponding to the speech frame and the maximum likelihood value in the likelihood values of each preset pronunciation unit corresponding to the speech frame, as the corresponding dynamic captcha score;

The average value of the dynamic verification code scores corresponding to each speech frame is determined as the target dynamic verification code score corresponding to the target user.

As an optional implementation, the device also includes:

The second obtaining module is used to obtain a pre-stored first training sample set, the first training sample set includes a plurality of sample user identifications and first sample voice data corresponding to each sample user identification;

The second division module is used to divide the first sample speech data into at least one first sample speech frame according to a preset segmentation algorithm for each first sample speech data in the first training sample set ;

The second extraction module is used to extract an acoustic feature vector corresponding to the first sample speech frame according to a preset acoustic feature extraction algorithm for each first sample speech frame corresponding to the first sample speech data;

The second output module is used to input the acoustic feature vector of each first sample speech frame corresponding to the first sample speech data to the identity verification multi-task model to be trained, and output the first sample speech data corresponding to the first acoustic feature vector Two posterior probability sets, the second posterior probability set includes the posterior probability corresponding to each sample user identifier;

The second determination module is used to determine the first cost function corresponding to the first training sample set according to the posterior probability of the sample user identification corresponding to each first sample speech data;

The first update module is used to update the parameters corresponding to the multi-task shared hidden layer in the identity verification multi-task model to be trained according to the first cost function and the preset first parameter update algorithm, the parameters corresponding to the voiceprint recognition network and Parameters corresponding to the speech recognition network.

As an optional implementation, the device also includes:

The third obtaining module is used to obtain a pre-stored second training sample set, the second training sample set includes a plurality of second sample speech frames and sample pronunciation units corresponding to each second sample speech frame;

The third extraction module is used to extract an acoustic feature vector corresponding to the second sample speech frame according to a preset acoustic feature extraction algorithm for each second sample speech frame in the second training sample set;

The third output module is used to input the acoustic feature vector corresponding to the second sample speech frame to the identity verification multi-task model to be trained, and output the third posterior probability set corresponding to the second sample speech frame, the third posterior The probability set includes the posterior probability corresponding to each sample pronunciation unit;

The third determination module is used to determine the second cost function corresponding to the second training sample set according to the posterior probability of the sample pronunciation unit corresponding to each second sample speech frame;

The second update module is used to update the parameters corresponding to the multi-task shared hidden layer and the parameters corresponding to the speech recognition network in the identity verification multi-task model to be trained according to the second cost function and the preset second parameter update algorithm.

As an optional implementation, the device also includes:

A fourth acquisition module, configured to acquire a plurality of pre-stored verification sample sets, each verification sample set including a plurality of sample user identifiers and second sample voice data corresponding to each sample user identifier;

The third dividing module is used to divide the second sample speech data into at least one third sample speech frame according to a preset segmentation algorithm for each second sample speech data in each verification sample set;

The fourth extraction module is used to extract an acoustic feature vector corresponding to the third sample speech frame according to a preset acoustic feature extraction algorithm for each third sample speech frame corresponding to the second sample speech data;

The fourth output module is used to input the acoustic feature vector of each third sample speech frame corresponding to the second sample speech data to the identity verification multi-task model to be verified, and output the fourth posteriori corresponding to the second sample speech data A probability set, the fourth posterior probability set includes the posterior probability corresponding to each sample user identifier;

The fourth determining module is used to determine the second sample voice if the posterior probability of the sample user identification corresponding to the second sample voice data is the maximum value in the fourth posterior probability set corresponding to the second sample voice data The data is target sample voice data;

The fifth determination module is used to determine the ratio of the number of target sample speech data in the verification sample set to the total number of second sample speech data in the verification sample set, as the accuracy rate of the verification sample set;

The sixth determination module is used to determine the rate of change of the accuracy rate corresponding to each verification sample set according to the accuracy rate of each verification sample set. If the set change rate threshold is determined, it is determined that the identity verification multi-task model training to be verified is completed.

An embodiment of the present application provides an identity verification device. The server acquires the voice data input by the target user according to the target dynamic verification code, and divides the voice data into at least one voice frame according to a preset segmentation algorithm. Then, for each speech frame, the server determines an acoustic feature vector corresponding to the speech frame according to a preset acoustic feature extraction algorithm. Afterwards, the server inputs the acoustic feature vector corresponding to the speech frame into the pre-trained identity verification multi-task model, and outputs the intermediate user feature vector and the first posterior probability set corresponding to the speech frame. Wherein, the first posterior probability set includes the posterior probability corresponding to each preset pronunciation unit. Finally, the server determines the first user feature vector corresponding to the target user according to the intermediate user feature vector corresponding to each speech frame and the preset pooling algorithm, and according to the first user feature vector corresponding to the target user and the first user feature vector corresponding to each speech frame A set of posterior probabilities for authenticating the target user. In this way, the server does not need to deploy two sets of voiceprint recognition models and speech recognition models with different structures and parameters, but only needs to deploy a set of authentication multi-tasking models to process the user's voice data, thereby reducing the computational complexity of the server , improve the processing efficiency of the server.

In one embodiment, a computer device, as shown in FIG. 10 , includes a memory and a processor, the memory stores a computer program that can run on the processor, and the processor implements the computer program when executing the computer program. The method steps of identity verification described in any one of the above.

In one embodiment, a computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps of the identity verification method described in any one of the foregoing are implemented.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the computer programs can be stored in a non-volatile computer-readable memory In the medium, when the computer program is executed, it may include the processes of the embodiments of the above-mentioned methods. Wherein, any references to memory, storage, database or other media used in the various embodiments provided in the present application may include non-volatile and/or volatile memory. Nonvolatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Chain Synchlink DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The technical features of the above embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be It is considered to be within the range described in this specification.

The above-mentioned embodiments only represent several implementation modes of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as limiting the scope of the patent for the invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the scope of protection of the patent application should be based on the appended claims.

Claims (10)

1. a kind of method of authentication, which is characterized in that the described method includes:

Obtain the voice data that target user inputs according to target dynamic identifying code；

According to preset segmentation algorithm, the voice data is divided at least one speech frame；

The corresponding acoustic feature vector of the speech frame is extracted according to preset acoustic feature extraction algorithm for each speech frame；

The corresponding acoustic feature vector of the speech frame is input to authentication multi task model trained in advance, exports the voice The corresponding intermediate user feature vector of frame and the first posterior probability set, the first posterior probability set include each default pronunciation The corresponding posterior probability of unit；

According to the corresponding intermediate user feature vector of each speech frame and preset pond algorithm, determine that the target user is corresponding First user characteristics vector；

The corresponding first posterior probability collection of each speech frame according to the corresponding first user characteristics vector sum of the target user It closes, authentication is carried out to the target user.

2. the method according to claim 1, wherein the authentication multi task model includes that multitask is shared Hidden layer, Application on Voiceprint Recognition network and speech recognition network；

Described that the corresponding acoustic feature vector of the speech frame is input to authentication multi task model trained in advance, output should The corresponding intermediate user feature vector of speech frame and the first posterior probability set, comprising:

The corresponding acoustic feature vector of the speech frame is input to the multitask and shares hidden layer, it is corresponding to export the speech frame Median feature vector；

The corresponding median feature vector of the speech frame is input to the speech recognition network, exports the corresponding pronunciation of the speech frame Feature vector and the first posterior probability set；

The corresponding median feature vector of the speech frame and pronunciation character vector are input to the Application on Voiceprint Recognition network, export the language The corresponding intermediate user feature vector of sound frame.

3. the method according to claim 1, wherein described special according to corresponding first user of the target user The corresponding first posterior probability set of each speech frame described in vector sum is levied, authentication is carried out to the target user, comprising:

According to the corresponding first posterior probability set of each speech frame, the corresponding target dynamic verifying of the target user is determined Code score；

If the first user characteristics vector second user feature vector corresponding with the pre-stored target user Similarity is greater than or equal to preset similarity threshold, and the target dynamic identifying code score is greater than or equal to preset dynamic Identifying code score threshold, it is determined that the target user is legitimate user；

If the first user characteristics vector is less than described preset similar to the similarity of the second user feature vector It spends threshold value or the target dynamic identifying code score is less than the preset dynamic verification code score threshold, it is determined that is described Target user is illegal user.

4. according to the method described in claim 3, it is characterized in that, described general according to corresponding first posteriority of each speech frame Rate set determines the corresponding target dynamic identifying code score of the target user, comprising:

Obtain the corresponding pronunciation unit sequence of the target dynamic identifying code；

It is calculated according to the corresponding first posterior probability set of each speech frame, the pronunciation unit sequence and preset pressure alignment Method determines the corresponding target speaker unit of each speech frame；

The corresponding target of the speech frame is determined in the corresponding first posterior probability set of the speech frame for each speech frame The posterior probability of pronunciation unit, and the posterior probability of the determining target speaker unit and the pre-stored target speaker unit The product of prior probability, the likelihood value as the target speaker unit；

According to the likelihood value of the corresponding target speaker unit of each speech frame, the corresponding target dynamic of the target user is determined Identifying code score.

5. according to the method described in claim 4, it is characterized in that, described obtain the corresponding pronunciation of the target dynamic identifying code Unit sequence, comprising:

According to the target dynamic identifying code and preset segmentation methods, the corresponding set of words of the target dynamic identifying code is determined It closes；

For each word in the set of letters, the corresponding relationship of word according to the pre-stored data and pronunciation unit sequence, Determine the corresponding pronunciation unit sequence of the word；

The corresponding pronunciation unit sequence of each word is carried out according to sequence of each word in the target dynamic identifying code Sequence, obtains the corresponding pronunciation unit sequence of the target dynamic identifying code.

6. according to the method described in claim 4, it is characterized in that, described according to the corresponding target speaker list of each speech frame The likelihood value of member, determines the corresponding target dynamic identifying code score of the target user, comprising:

For each speech frame, determine that the likelihood value of the corresponding target speaker unit of the speech frame is corresponding with the speech frame described The difference of maximum likelihood value in the likelihood value of each default pronunciation unit, as the corresponding dynamic verification code score of the speech frame；

The average value for determining the corresponding dynamic verification code score of each speech frame, it is dynamic as the corresponding target of the target user State identifying code score.

7. according to the method described in claim 2, it is characterized in that, the method also includes:

Obtain pre-stored first training sample set, the first training sample set include multiple sample of users marks and Each sample of users identifies corresponding first sample voice data；

For each first sample voice data in the first training sample set, according to preset segmentation algorithm, by this First sample voice data is divided at least one first sample speech frame；

For the corresponding each first sample speech frame of the first sample voice data, is extracted and calculated according to preset acoustic feature Method extracts the corresponding acoustic feature vector of the first sample speech frame；

The acoustic feature vector of the corresponding each first sample speech frame of the first sample voice data is input to body to be trained Part verifying multi task model, exports the corresponding second posterior probability set of the first sample voice data, second posteriority is general Rate set includes the corresponding posterior probability of each sample user identifier；

According to the posterior probability of the corresponding sample of users mark of each first sample voice data, first training sample set is determined Close corresponding first cost function；

According to first cost function and preset first parameter more new algorithm, to be trained authentication more described in update Multitask described in business model is shared the corresponding parameter of hidden layer, the corresponding parameter of the Application on Voiceprint Recognition network and the voice and is known The corresponding parameter of other network.

8. the method according to claim 2 or 7, which is characterized in that the method also includes:

Pre-stored second training sample set is obtained, the second training sample set includes multiple second sample voice frames Sample pronunciation unit corresponding with each second sample voice frame；

For the second sample voice frame of each of described second training sample set, extracted according to the preset acoustic feature Algorithm extracts the corresponding acoustic feature vector of the second sample voice frame；

The corresponding acoustic feature vector of second sample voice frame is input to authentication multi task model to be trained, is exported The corresponding third posterior probability set of second sample voice frame, the third posterior probability set include each sample pronunciation unit Corresponding posterior probability；

According to the posterior probability of the corresponding sample pronunciation unit of each second sample voice frame, the second training sample set is determined Corresponding second cost function；

According to second cost function and preset second parameter more new algorithm, to be trained authentication more described in update The corresponding parameter of hidden layer and the corresponding parameter of the speech recognition network are shared in multitask described in business model.

9. the method according to the description of claim 7 is characterized in that the method also includes:

Obtain pre-stored multiple verifying sample sets, each verifying sample set include multiple sample of users marks and Each sample of users identifies corresponding second sample voice data；

For the second sample voice data of each of each verifying sample set, according to preset segmentation algorithm, by this second Sample voice data are divided at least one third sample voice frame；

For the corresponding each third sample voice frame of the second sample voice data, is extracted and calculated according to preset acoustic feature Method extracts the corresponding acoustic feature vector of third sample voice frame；

The acoustic feature vector of the corresponding each third sample voice frame of the second sample voice data is input to body to be verified Part verifying multi task model, exports the corresponding 4th posterior probability set of the second sample voice data, the 4th posteriority is general Rate set includes the corresponding posterior probability of each sample user identifier；

If the posterior probability of the corresponding sample of users mark of the second sample voice data is the second sample voice data pair The maximum value in the 4th posterior probability set answered, it is determined that the second sample voice data are target sample voice data；

Determine the number of target sample voice data and the second sample voice in the verifying sample set in the verifying sample set The ratio of the total number of data, the accuracy rate as the verifying sample set；

According to the accuracy rate of each verifying sample set, the change rate of the corresponding accuracy rate of each verifying sample set is determined, if deposited It is less than or equal to preset change rate threshold value in the change rate of the corresponding accuracy rate of continuous preset number verifying sample set, then Determine that the authentication multi task model training to be verified is completed.

10. a kind of device of authentication, which is characterized in that described device includes:

First obtains module, the voice data inputted for obtaining target user according to target dynamic identifying code；

First division module, for according to preset segmentation algorithm, the voice data to be divided at least one speech frame；

First extraction module, for extracting the speech frame pair according to preset acoustic feature extraction algorithm for each speech frame The acoustic feature vector answered；

First output module, for the corresponding acoustic feature vector of the speech frame to be input to trained in advance authentication more Business model, exports the corresponding intermediate user feature vector of the speech frame and the first posterior probability set, first posterior probability Set includes the corresponding posterior probability of each default pronunciation unit；

First determining module, for determining according to the corresponding intermediate user feature vector of each speech frame and preset pond algorithm The corresponding first user characteristics vector of the target user；

Authentication module, for each speech frame according to the corresponding first user characteristics vector sum of the target user corresponding the One posterior probability set carries out authentication to the target user.