WO2017215558A1 - Voiceprint recognition method and device - Google Patents

Abstract

A voiceprint recognition method and device. The method comprises: obtaining authentication voice information generated by reading a first character string by an authenticated user (S201); performing voice recognition on the authentication voice information to obtain video segments included in the authentication voice information and respectively corresponding to a plurality of characters in the first character string (S202); extracting a voiceprint feature of the video segment corresponding to each character (S203); obtaining, according to the voiceprint feature of the video segment corresponding to each character, a feature vector corresponding to each character in the authentication voice information on the basis of preset general background model training corresponding to a corresponding character (S204); and calculating a similarity score between the feature vector corresponding to each character in the authentication voice information and the feature vector corresponding to a corresponding character in preset registration voice information, and if the similarity score reaches a preset authentication threshold, determining the authenticated user to be a registered user corresponding to the registration voice information (S205). The voiceprint recognition accuracy can be effectively improved.

Description

Voiceprint recognition method and device

The present invention claims priority to the priority of the application No. 201610416650.3, entitled "A Voiceprint Identification Method and Apparatus", filed on June 12, 2016, the contents of which are incorporated herein by reference. .

Technical field

The present invention relates to the field of voice recognition technology, and in particular, to a voiceprint recognition method and apparatus.

Background technique

Voiceprint recognition is a method of biometric information recognition, including user registration and user identification. The registration phase maps speech through a series of processes to a user model. In the identification stage, for a voice with unknown identity, the similarity is matched with the model, and then the identity of the unknown voice is consistent with the identity of the registered voice. The existing voiceprint modeling methods usually model from the text-independent level to realize the description of the speaker's identity characteristics, but the text-independent modeling method is difficult to meet when the user reads different content. Claim.

Summary of the invention

In view of this, the embodiment of the invention provides a voiceprint recognition method and device, which can effectively improve the accuracy of voiceprint recognition.

In order to solve the above technical problem, a first aspect of the embodiments of the present invention provides a voiceprint recognition method, where the method includes:

Obtaining verification voice information generated by the verification user reading the first character string;

Calculating a similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset voice information in the voice information of the verification voice information;

When the similarity reaches a preset threshold, the user verification is verified as the registered user corresponding to the registered voice information.

In an implementation manner, the calculating, before verifying, the similarity between the voice segment corresponding to each character in the voice information and the voiceprint feature of the voice segment corresponding to the corresponding character in the preset registration voice information include:

Performing voice recognition on the verification voice information, and obtaining a voice segment respectively corresponding to each character in the first character string included in the verification voice information;

The voiceprint features of the voice segments corresponding to the respective characters are extracted.

In an implementation manner, the similarity between the voiceprint feature of the voice segment corresponding to the corresponding character in the preset registration voice information is verified by the calculation:

And performing, according to the voiceprint feature of the voice segment corresponding to each character, in combination with a common background model corresponding to the preset corresponding character, obtaining a feature vector corresponding to each character in the verification voice information;

Calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, as a voice segment corresponding to each character in the verification voice information and a preset registration The similarity of the voiceprint features of the voice segments corresponding to the corresponding characters in the voice information.

In an implementation manner, before the obtaining the verification voice information generated by the verification user to read the first character string, the method further includes:

Obtaining registration voice information generated by the registered user reading the second character string, the second character string having at least one of the same characters as the first character string;

The feature vector corresponding to each character in the registered voice information is obtained according to the voiceprint feature corresponding to each character in the registered voice information, combined with the common background model corresponding to the preset corresponding character.

In an implementation manner, the voiceprint feature corresponding to the voice segment corresponding to the respective characters is combined with the common background model corresponding to the preset corresponding character to obtain a feature corresponding to each character in the verification voice information. Vectors include:

Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character, and obtain the The feature vector corresponding to each character in the voice information is verified.

In an implementation manner, the voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as training sample data, and the maximum a posteriori probability algorithm is used to correspond to the preset corresponding character. The mean supervector of the background model is adjusted, so that the feature vector corresponding to each character in the verification voice information is obtained:

Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as a training sample In the data, the maximum a posteriori probability algorithm is used to adjust the mean supervector of the universal background model corresponding to the preset corresponding character, and combined with the preset supervector subspace matrix, each of the verified voice information is obtained. The feature vector corresponding to the character.

In an implementation manner, the voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the maximum a posteriori probability algorithm is used to correspond to the preset corresponding character. Adjusting the mean supervector of the universal background model, and combining the preset supervector subspace matrix, obtaining the feature vector corresponding to each character in the verification voice information includes:

Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the mean supervector of the common background model corresponding to the preset corresponding character is adjusted by using the following formula, so that the adjusted The general background model corresponding to the corresponding character has the largest posterior probability:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector The spatial matrix, ω is a feature vector corresponding to the corresponding character in the verification speech information.

In an implementation manner, the preset super-vector sub-space matrix is determined according to a correlation between weights of respective Gaussian modules in the universal background model.

In an implementation manner, the calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the corresponding character in the preset registration voice information includes:

Calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and determining the cosine distance value as the similarity score.

In an implementation manner, the performing voice recognition on the verification voice information to obtain the voice segments respectively included in the verification voice information and corresponding to the plurality of characters in the first character string includes:

Identifying valid voice segments and invalid voice segments in the verification voice information;

Performing speech recognition on the valid speech segment to obtain a speech segment respectively corresponding to a plurality of characters in the first character string.

In an implementation manner, before the determining, by the verification user, the registered user corresponding to the registration voice information, the method further includes:

Determining whether the ordering of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string;

And if the similarity reaches a preset threshold value, and the order of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string, The verification user determines the registered user corresponding to the registration voice information.

In an implementation, the method further includes: before the obtaining the verification voice information generated by the verification user to read the first character string, the method further includes:

The first character string is randomly generated, and the first character string is displayed.

Correspondingly, a second aspect of the embodiments of the present invention provides a voiceprint recognition apparatus, where the apparatus includes:

a voice acquiring module, configured to obtain verification voice information generated by the user to read the first character string;

The similarity judging module is configured to calculate a similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset voice information in the voice information corresponding to each character in the verification voice information;

The user identification module is configured to verify the verification user as the registered user corresponding to the registered voice information when the similarity reaches a preset threshold.

In an implementation manner, the voiceprint recognition apparatus further includes:

a voice segment identification module, configured to perform voice recognition on the verification voice information, and obtain a voice segment respectively included in the verification voice information corresponding to multiple characters in the first character string;

The voiceprint feature extraction module is configured to extract a voiceprint feature of the voice segment corresponding to each character in the verification voice information.

In an implementation manner, the voiceprint recognition apparatus further includes:

a feature model training module, configured to perform, according to the voiceprint feature of the voice segment corresponding to each character, and the common background model corresponding to the corresponding corresponding character, to obtain a feature vector corresponding to each character in the verification voice information;

The similarity judging module is configured to calculate a similarity score of the feature vector corresponding to each character in the verification voice information and the corresponding character in the preset registration voice information, as the verification corresponding to each character in the voice information The similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset registered voice information.

In an implementation manner, the voice acquiring module is further configured to obtain registration voice information generated by a registered user reading a second character string, where the second character string has at least one identical character with the first character string. ;

The feature model training module is further configured to: according to each character in the registered voice information The voiceprint feature is combined with a common background model corresponding to the preset corresponding character to obtain a feature vector corresponding to each character in the registered voice information.

In one implementation, the feature model training module is configured to:

Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character, and obtain the The feature vector corresponding to each character in the voice information is verified.

In one implementation, the feature model training module is configured to:

The voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character. Combined with the preset super-vector subspace matrix, the feature vector corresponding to each character in the voice information is obtained.

In an implementation manner, the feature model training module is specifically configured to:

The voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the mean supervector of the common background model corresponding to the preset corresponding character is adjusted by using the following formula, so that the adjusted corresponding The general background model corresponding to the character has the largest posterior probability:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector The spatial matrix, ω is a feature vector corresponding to the corresponding character in the verification speech information.

In an implementation manner, the preset super-vector sub-space matrix is determined according to a correlation between respective dimension vectors in the mean super-vector of the Gaussian mixture model.

In an implementation manner, the similarity determination module is configured to:

Calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and determining the cosine distance value as the similarity score .

In an implementation manner, the voice segment identification module includes:

a valid segment identification unit, configured to identify a valid speech segment and an invalid speech segment in the verification speech information;

The voice recognition unit is configured to perform voice recognition on the valid voice segment to obtain a voice segment corresponding to multiple characters in the first character string.

In an implementation manner, the voiceprint recognition apparatus further includes:

a character order determining module, configured to determine whether a sorting of the voice segments of the plurality of characters in the verification voice information is consistent with a ranking of the corresponding characters in the first character string;

The user identification module is further configured to: when the similarity reaches a preset threshold, and verify the order of the voice segments of the plurality of characters in the voice information and the corresponding characters in the first string In the case that the sorting is consistent, the verification user is determined as the registered user corresponding to the registered voice information.

In an implementation manner, the voiceprint recognition apparatus further includes:

a string display module, configured to randomly generate the first string, and display the first string.

A third aspect of the embodiments of the present invention further provides a voiceprint recognition apparatus, including:

a user interface for obtaining voice information;

a memory storing computer executable program code;

a processor for invoking the computer executable program code to perform the following operations:

Acquiring, by the user interface, verification voice information generated by the user to read the first character string;

Calculating a similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset voice information in the voice information of the verification voice information;

When the similarity reaches a preset threshold, the user verification is verified as the registered user corresponding to the registered voice information.

In an implementation manner, before the obtaining the verification voice information generated by the verification user to read the first character string, the processor further calls the computer executable program code to perform the following operations:

Performing voice recognition on the verification voice information, and obtaining a voice segment respectively corresponding to each character in the first character string included in the verification voice information;

The voiceprint features of the voice segments corresponding to the respective characters are extracted.

In an implementation manner, the processor invokes the computer executable program code to perform the following operations to calculate a sound segment of a voice segment corresponding to a corresponding character in the preset registered voice information in the voice information of the verification voice information. Similarity of the pattern features:

And performing, according to the voiceprint feature of the voice segment corresponding to each character, in combination with a common background model corresponding to the preset corresponding character, obtaining a feature vector corresponding to each character in the verification voice information;

Calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, as a voice segment corresponding to each character in the verification voice information and a preset registration The similarity of the voiceprint features of the voice segments corresponding to the corresponding characters in the voice information.

In an implementation manner, before the obtaining the verification voice information generated by the verification user to read the first character string, the processor further calls the computer executable program code to perform the following operations:

Obtaining registration voice information generated by the registered user reading the second character string, the second character string having at least one of the same characters as the first character string;

The feature vector corresponding to each character in the registered voice information is obtained according to the voiceprint feature corresponding to each character in the registered voice information, combined with the common background model corresponding to the preset corresponding character.

In one implementation, the processor invokes the computer executable program code to perform the following operations to combine a common background model corresponding to the preset corresponding character according to a voiceprint feature of the voice segment corresponding to the respective character Training, obtaining a feature vector corresponding to each character in the verification voice information:

Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character, and obtain the The feature vector corresponding to each character in the voice information is verified.

In one implementation, the processor invokes the computer executable program code to perform the following operations to use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as training sample data, using the maximum The probability algorithm adjusts the mean supervector of the universal background model corresponding to the preset corresponding character, thereby obtaining a feature vector corresponding to each character in the verified voice information:

The voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character. And combining the preset super vector subspace matrix to obtain a feature vector corresponding to each character in the verification voice information.

In one implementation, the processor invokes the computer executable program code to perform the following operations to use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as Training the sample data, using the maximum posterior probability algorithm to adjust the mean supervector of the universal background model corresponding to the preset corresponding character, and combining the preset supervector subspace matrix to obtain the verification voice The feature vector corresponding to each character in the message:

Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the mean supervector of the common background model corresponding to the preset corresponding character is adjusted by using the following formula, so that the adjusted The general background model corresponding to the corresponding character has the largest posterior probability:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector The spatial matrix, ω is a feature vector corresponding to the corresponding character in the verification speech information.

In an implementation manner, the preset super-vector sub-space matrix is determined according to a correlation between weights of respective Gaussian modules in the universal background model.

In an implementation manner, the processor calls the computer executable program code to perform the following operations to calculate a feature vector corresponding to a feature vector corresponding to each character in the verification voice information and a corresponding character in the preset registration voice information. The similarity scores include:

Calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and determining the cosine distance value as the similarity score.

In an implementation manner, the processor calls the computer executable program code to perform the following operations to perform voice recognition on the verification voice information to obtain the verification voice information included in the first string and the first character string respectively The voice segments corresponding to multiple characters include:

Identifying valid voice segments and invalid voice segments in the verification voice information;

Performing speech recognition on the valid speech segment to obtain a speech segment respectively corresponding to a plurality of characters in the first character string.

In an implementation manner, before the determining the determined user as the registered user corresponding to the registered voice information, the processor further invokes the computer executable program code to perform the following operations:

Determining whether the ordering of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string;

The similarity reaches a preset threshold, and the plurality of characters in the voice information are verified If the order of the voice segments is consistent with the ordering of the corresponding characters in the first character string, the verification user is determined as the registered user corresponding to the registered voice information.

In an implementation manner, before the obtaining the verification voice information generated by the verification user to read the first character string, the processor further calls the computer executable program code to perform the following operations:

The first character string is randomly generated, and the first character string is displayed through the user interface.

Correspondingly, the fourth aspect of the embodiments of the present invention further provides a storage medium, where the computer program stores a computer program, where the computer program is used to perform any implementation manner according to the first aspect of the embodiments of the present invention. The voiceprint recognition method described.

In this embodiment, the voiceprint feature of the voice segment corresponding to each character in the verification voice information of the user is obtained, and the UTM training corresponding to the preset corresponding character is used to obtain the feature vector corresponding to each character in the voice message, and the voice information is verified. The feature vector corresponding to each character in the registration is compared with the feature vector of the corresponding character in the registered voice information, thereby determining the user identity of the verified user, and the user feature vector used for comparison corresponds to the specific character, fully considering the user reading aloud The voiceprint features of different characters can effectively improve the accuracy of voiceprint recognition.

DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below. Obviously, the drawings in the following description are only It is a certain embodiment of the present invention, and other drawings can be obtained from those skilled in the art without any creative work.

1 is a schematic diagram showing the stages of a voiceprint recognition method in an embodiment of the present invention;

2 is a schematic flow chart of a voiceprint recognition method according to an embodiment of the present invention;

3 is a schematic diagram showing the principle of identifying a voice segment corresponding to a plurality of characters from voice information in the embodiment of the present invention;

4 is a schematic diagram of a principle for acquiring feature vectors corresponding to respective characters from voice information according to an embodiment of the present invention;

5 is a schematic diagram of a voiceprint registration process of a registered user in an embodiment of the present invention;

6 is a schematic flow chart of a voiceprint recognition method in another embodiment of the present invention;

FIG. 7 is a schematic structural diagram of a voiceprint recognition apparatus according to an embodiment of the present invention; FIG.

FIG. 8 is a schematic structural diagram of a voice segment identification module in an embodiment of the present invention.

FIG. 9 is a schematic structural diagram of another voiceprint recognition apparatus in an embodiment of the present invention.

detailed description

The technical solutions in the embodiments of the present invention are clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention. It is obvious that the described embodiments are only a part of the embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts are within the scope of the present invention.

Embodiments of the present invention provide a voiceprint recognition method and apparatus. The voiceprint recognition method and apparatus can be applied to all scenes or devices that need to identify an unknown user. The characters in the string used for voiceprint recognition may be Arabic numerals, English letters or other language characters. In order to simplify the description, the characters in the embodiments of the present invention are exemplified by taking Arabic numerals as an example.

The voiceprint recognition method in the embodiment of the present invention can be divided into two stages, as shown in FIG. 1:

1) Registered user's voiceprint registration stage

In the voiceprint registration phase, the registered user can read a registration string (ie, the second character string appearing later), and the voiceprint recognition device collects the registered voice information when the registered user reads the registration string, and then registers the voice. Performing voice recognition on the information to obtain voice segments respectively corresponding to the plurality of characters in the registration string, and performing voiceprint feature extraction and voiceprint model training on the voice segments corresponding to each character, including And obtaining, according to the voiceprint feature of the voice segment corresponding to each character, a feature vector corresponding to each character in the registered voice information by using a Universal Background Model (UBM, ie, GMM-UBM) corresponding to the preset corresponding character, Then, the voiceprint recognition device can respectively save the feature vectors corresponding to the plurality of characters in the registered voice information that the different registered users read in the voiceprint registration stage in the model library of the voiceprint recognition device.

For example, the registration string is a numeric string 0185851, and includes four numbers “0”, “1”, “5”, “8”, and the voiceprint recognition device performs sound according to the voice segment corresponding to each character in the registered voice information. The pattern feature extraction and the voiceprint model training obtain the voiceprint features of the voice segments corresponding to “0”, “1”, “5”, and “8”, and then obtain the registered voice information in combination with the UBM training corresponding to the preset corresponding characters. Individual characters Corresponding feature vectors include a feature vector corresponding to the number “0”, a feature vector corresponding to the number “1”, a feature vector corresponding to the number “5”, and a feature vector corresponding to the number “8”.

2) Verify the user's identification phase

In the identification phase, the user who authenticates the user, that is, the unknown identity, reads a verification string (ie, the first character string that appears later, and the second character string has at least one of the same characters as the first character string). The pattern recognition device collects the verification voice information when the verification user reads the verification character string, and then performs voice recognition on the verification voice information to obtain corresponding to the plurality of characters in the verification string respectively included in the verification voice information. The voice segment, and then the voiceprint feature extraction and the voiceprint model training for the voice segment corresponding to each character, including the voiceprint feature corresponding to the voice segment corresponding to the respective character, and the UBM training corresponding to the corresponding corresponding character is used to obtain the verification voice. The feature vector corresponding to each character in the information, and finally calculating the similarity score of the feature vector corresponding to each character in the verification voice information and the corresponding character in the preset registration voice information, if the similarity score reaches the preset verification a threshold, the verification user is determined as a note corresponding to the registration voice information Users.

For example, if the verification string is a numeric string 85851510, the voiceprint recognition device performs voiceprint feature extraction and voiceprint model training according to the voice segment corresponding to each character in the verification voice information generated when the user reads aloud, and obtains “0” and “ The GMM corresponding to 1", "5", and "8", and then combined with the preset UBM corresponding to the corresponding character, can calculate the feature vector of the verified voice information of the verified user, including the feature vector corresponding to the number "0", and the number a feature vector corresponding to "1", a feature vector corresponding to the number "5", and a feature vector corresponding to the number "8", thereby respectively calculating "0", "1", "5", "8" in the verification voice information a similarity score between the corresponding feature vectors and the feature vectors corresponding to "0", "1", "5", and "8" in the registered voice information, and if the similarity score reaches a preset verification threshold, And determining, by the verification user, a registered user corresponding to the registered voice information.

It should be noted that the voiceprint registration phase of the registered user and the identity phase of the authenticated user may be implemented in the same device or device, or may be implemented in different devices or devices, for example, the voiceprint registration phase of the registered user is The first device is implemented, and the first device sends the feature vector corresponding to the multiple characters in the registered voice information to the second device, so that the identity recognition phase of the authenticated user can be implemented in the second device.

The above two processes are respectively explained in detail through specific embodiments.

2 is a schematic flow chart of a method for identifying a voiceprint according to an embodiment of the present invention. As shown in the figure, the flow of the voiceprint recognition method in the embodiment may include:

S201. Acquire verification voice information generated by verifying that the user reads the first character string.

The user who authenticates the user, that is, the unknown identity, needs to verify the identity of the user through the voiceprint recognition device. The first character string is a character string used for verifying the user's identity verification, and may be randomly generated, or may be a preset fixed string, for example, a second string corresponding to the pre-generated registration voice information. Part of the same string. Specifically, the first character string may include m characters, wherein n characters are different from each other, m and n are positive integers, and m≥n.

For example, the first string is "12358948", a total of 8 characters, including 7 different characters "1", "2", "3", "4", "5", "8", " 9".

In an alternative embodiment, the voiceprint recognition device may generate and display the first character string for the verification user to read aloud according to the displayed first character string.

S202. Perform speech recognition on the verification voice information to obtain a voice segment respectively included in the verification voice information and corresponding to multiple characters in the first character string.

As shown in FIG. 3, the voiceprint recognition device can divide the verification voice information into voice segments corresponding to multiple characters through voice recognition and voice intensity filtering, and optionally remove the invalid voice segments, and do not participate in subsequent Process.

S203. Extract a voiceprint feature of the voice segment corresponding to each character.

Specifically, the voiceprint recognition device may extract MFCC (Mel Frequency Cepstrum Coefficient) or PLP (Perceptual Linear Predictive Coefficient) in the voice segment corresponding to each character as the voice corresponding to each character. The voiceprint features of the clip.

S204, according to the voiceprint feature of the voice segment corresponding to each character, combined with the common background model corresponding to the preset corresponding character, to obtain a feature vector corresponding to each character in the verification voice information;

The universal background model UBM in the embodiment of the present invention is a mixed Gaussian model in which a plurality of speech segments of a plurality of speakers are mixed and trained, and the distribution of the corresponding digital speech in the feature space is represented, and The data comes from a large number of speakers, so it does not represent a specific type of speaker, and has an identity-independent nature, which can be regarded as a general background model. Schematically, a speech sample with more than 1000 speakers and a duration of more than 20 hours can be used, and each character is out. The current frequency is relatively balanced and training is obtained by UBM. The mathematical expression of UBM is:

P(x)=∑ _t=1...C a _i N(x|μ _i ,∑ _i )............(1)

Where P(x) represents the probability distribution of UBM, C represents the total of C Gaussian modules in UBM, a _i represents the weight of the i-th Gaussian module, μ _i represents the mean of the i-th Gaussian module, ∑ _i represents The variance of the i-th Gaussian module, N(x) represents a Gaussian distribution, and x represents the voiceprint characteristics of the input sample.

The voiceprint recognition device may use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, and adopt a maximum a posteriori probability algorithm (MAP) to the common background model corresponding to the preset corresponding character. The parameter is adjusted, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1) as an input sample, the posterior check is made by continuously adjusting the parameters of the common background model corresponding to the corresponding corresponding character. The probability P(x) is the largest, so that the feature vector corresponding to the corresponding character in the verification speech information can be determined according to the parameter that makes the posterior probability P(x) the largest.

Since a large number of experiments and papers have verified that the mean of each Gaussian module in the UBM model can be used to distinguish the identity information of the speaker, the mean supervector of the UBM model is defined as:

Therefore, the voiceprint recognition device can use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, and adopt the maximum posterior probability algorithm (MAP) to the common background corresponding to the preset corresponding character. The mean value supervector of the model is adjusted, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1) as an input sample, the posterior probability P(x) is adjusted by continuously adjusting the mean supervector. The maximum, so that the mean supervector that maximizes the posterior probability P(x) can be used as the feature vector corresponding to the corresponding character in the verification speech information.

In another alternative embodiment, in order to reduce the slow convergence speed caused by the high dimension of the super vector, The profiling of the mean supervector is limited to a subspace by probabilistic principal component analysis (PPCA), and the voiceprint recognition device can verify the voiceprint features of the speech segments corresponding to each character in the speech information. As the training sample data, the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the corresponding corresponding character, and the preset super-vector subspace matrix is used to obtain the corresponding features of each character in the verified voice information. vector. In a specific implementation, the mean supervector of the common background model corresponding to the preset corresponding character may be adjusted by using the following formula, so that the posterior probability of the universal background model corresponding to the adjusted corresponding character is the largest:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector subspace matrix , ω is the feature vector corresponding to the corresponding character in the verification voice information, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1) as an input sample, the adjustment can be realized by continuously adjusting ω. The mean supervector in (1) maximizes the posterior probability P(x), so that ω that maximizes the posterior probability P(x) can be used as the feature vector corresponding to the corresponding character in the verification speech information. The super-vector subspace matrix T is determined according to the correlation between each dimension vector in the mean supervector of the Gaussian mixture model.

S205. Calculate a similarity score of the feature vector corresponding to the corresponding character in the preset registered voice information in the verification voice information, and if the similarity score reaches the preset verification threshold, the verification user is The registered user corresponding to the registered voice information is determined.

Specifically, the voiceprint recognition device can obtain the registered voice information of the registered user in the voiceprint registration stage, and obtain the voiceprint feature extraction and the voiceprint model training similar to the embodiment, so that the characters in the registered voice information can be obtained. The feature vector corresponding to the speech segment. The registration voice information may be that the voiceprint recognition device acquires the registration voice information generated by the registered user reading the second character string, and the second character string has at least one of the same characters as the first character string, that is, the The second character string corresponding to the registration voice information is at least partially identical to the first character string. In an alternative embodiment, the voiceprint recognition device may further acquire the feature vector corresponding to the corresponding character in the registered voice information from the outside, that is, after the registered user inputs the registered voice information through other devices, the other device or the server passes the voiceprint. Feature extraction and voiceprint model training obtain feature vectors corresponding to voice segments of each character in the registered voice information, and the voiceprint recognition device acquires the phase in the registered voice information by using other devices or servers The feature vector corresponding to the character is used to compare the feature vector corresponding to each character in the verification voice information in the identification phase of the verification user.

In a specific implementation, the similarity score is that the voiceprint recognition device compares the feature vector corresponding to each character in the voice information with the feature vector corresponding to the corresponding character in the preset registered voice information, and then measures two features of the same character. The degree of similarity between vectors. In an optional embodiment, a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information may be calculated, and the cosine distance value is used as the The similarity score, that is, the similarity score between the corresponding feature vector in the verification voice information and the feature vector in the registered voice information is calculated by the following formula:

Wherein, the subscript i indicates a character common to the i-th verification voice information and the registration voice information, ω _i (tar) indicates a corresponding feature vector of the character in the verification voice information, and ω _i (test) indicates that the character is in the registered voice. The corresponding feature vector in the message. If the verification voice information and the registration voice information include multiple identical characters, the similarity scores of the characters calculated according to the above formula may be averaged. If the similarity scores of the respective characters reach the corresponding preset verification threshold, then The verification user is determined to be a registered user corresponding to the registration voice information. If there are multiple registered users, such as the registered users A, B, and C shown in FIG. 1, the registered user may be authenticated according to the similarity between the feature vector of a certain character of the user and the corresponding character of each registered user. The feature vector of the corresponding character and the feature vector of the character of the verification voice have the highest similarity score and the similarity reaches the preset verification threshold, and the registered user is used as the identification result of the verification user.

In an optional embodiment, if the same character appears in the verification voice information more than once, for example, if 0, 1, 5, and 8 appear in the verification voice information as shown in FIG. 2, respectively, then The average of the similarity scores of the feature vectors processed by the two-character 0 corresponding to the speech segment and the feature vector of the character 0 in the preset registration speech information is used as the feature vector and preset of the character 0 in the verification speech information. The similarity score of the feature vector of character 0 in the registered voice information, and so on.

It should be noted that there are many ways to measure the similarity between two feature vectors. The above is only one embodiment provided by the present invention, and those skilled in the art based on the solution disclosed by the present invention. It is possible to obtain more ways of calculating the similarity scores of the feature vectors of the characters shared in the voice information and the registered voice information without creative labor, and the present invention need not be exhaustive.

Therefore, in this embodiment, by obtaining the voiceprint feature of the voice segment corresponding to each character in the verification voice information of the verification user, the UTM training corresponding to the preset corresponding character is used to obtain the feature vector corresponding to each character in the verification voice information, and the verification is performed. The feature vector corresponding to each character in the voice information is compared with the feature vector of the corresponding character in the registered voice information, thereby determining the user identity of the verified user. The user feature vector used for comparison corresponds to the specific character, and the user is fully considered. The voiceprint features of different characters are read aloud, so that the accuracy of voiceprint recognition can be effectively improved.

5 is a schematic diagram of a voiceprint registration process of a registered user in the embodiment of the present invention. As shown in the figure, the voiceprint registration process in this embodiment may include:

S501. Acquire a registration voice information generated by a registered user reading a second character string, where the second character string has at least one identical character with the first character string.

The registered user is a user who determines a legal identity. The second character string is a string used to collect the voiceprint feature vector of the registered user, and may be randomly generated or a fixed string. Specifically, the second character string may also include m characters, wherein there are n characters different from each other, m and n are positive integers, and m≥n.

In an alternative embodiment, the voiceprint recognition device may generate and display the second character string for the registered user to read aloud according to the displayed second character string.

S502. Perform speech recognition on the registration voice information to obtain a voice segment respectively included in the registration voice information and corresponding to multiple characters in the second character string.

The voiceprint recognition device can divide the verification voice information into voice segments corresponding to multiple characters through voice recognition and voice intensity filtering, and optionally remove the invalid voice segments, and do not participate in subsequent processing.

S503. Extract a voiceprint feature of the voice segment corresponding to each character in the registered voice information.

Specifically, the voiceprint recognition device may extract MFCC (Mel Frequency Cepstrum Coefficient) or PLP (Perceptual Linear Predictive Coefficient) in the voice segment corresponding to each character as the voice corresponding to each character. The voiceprint features of the clip.

S504, according to the voiceprint feature of the voice segment corresponding to each character in the registered voice information, combined with the pre- The general background model corresponding to the corresponding character is trained to obtain the feature vector corresponding to each character in the registered voice information.

The expression of UBM can refer to the previous embodiment. This step of the voiceprint registration process is similar to S204 of the voiceprint recognition process. The voiceprint recognition device can use the voiceprint feature of the voice segment corresponding to each character in the registered voice information as the training sample data, and adopt the maximum posterior probability algorithm (Maximum A). Posteriori, MAP) adjusts the parameters of the common background model corresponding to the corresponding corresponding characters, that is, after the voiceprint feature of the voice segment corresponding to each character in the registered voice information is substituted into the formula (1) as an input sample, The parameters of the common background model corresponding to the preset corresponding characters are such that the posterior probability P(x) is the largest, so that the feature vector corresponding to the corresponding character in the registered voice information can be determined according to the parameter that makes the posterior probability P(x) the largest.

Since the mean value of each Gaussian module in the UBM model can be used to distinguish the identity information of the speaker, the voiceprint recognition device can use the voiceprint feature of the voice segment corresponding to each character in the registered voice information as the training sample data, using the maximum posterior The Maximum A Posteriori (MAP) adjusts the mean supervector of the universal background model corresponding to the preset corresponding characters, that is, the voiceprint feature of the voice segment corresponding to each character in the registered voice information is used as an input sample substitution type ( 1) After, by continuously adjusting the mean supervector, the posterior probability P(x) is maximized, so that the mean supervector that maximizes the posterior probability P(x) can be used as the feature vector corresponding to the corresponding character in the registered speech information.

In another optional embodiment, the mean supervector of the universal background model corresponding to the preset corresponding character may be adjusted by using the following formula, so that the posterior probability of the universal background model corresponding to the adjusted corresponding character is the largest:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector subspace matrix , ω is the feature vector corresponding to the corresponding character in the registered voice information, that is, after the voiceprint feature of the voice segment corresponding to each character in the registered voice information is substituted into the formula (1) as an input sample, the adjustment can be realized by continuously adjusting ω. The mean supervector in (1) maximizes the posterior probability P(x), so that ω that maximizes the posterior probability P(x) can be used as the feature vector corresponding to the corresponding character in the registered speech information.

6 is a schematic flow chart of a voiceprint recognition method according to another embodiment of the present invention, as shown in the figure. The voiceprint recognition method in the embodiment may include the following processes:

S601. Randomly generate a first character string and display it.

S602. Acquire verification voice information generated by verifying that the user reads the first character string.

S603. Identify valid voice segments and invalid voice segments in the verification voice information.

Specifically, the verification voice may be divided according to the sound intensity, and the voice segment with less sound intensity is regarded as an invalid voice segment (for example, including a silent segment and impulse noise).

S604. Perform speech recognition on the valid speech segment to obtain a speech segment respectively corresponding to the plurality of characters in the first character string.

A voice segment respectively corresponding to a plurality of characters in the first character string can be obtained by voice recognition.

S605. Determine that the order of the voice segments of the plurality of characters in the verification voice information is consistent with the order of the corresponding characters in the first character string.

In order to effectively prevent the voice information of the registered user from being pirated or illegally copied for voiceprint recognition, a different first character string may be randomly generated each time, and in this step, the voice segment of the plurality of characters in the voice information is determined. Whether the sorting is consistent with the sorting of the corresponding characters in the first string, if not, it can be judged that the voiceprint recognition fails, and if the sorting of the corresponding characters in the first string is consistent, the subsequent process is performed.

S606. Extract voiceprint features of the voice segments corresponding to the respective characters.

Specifically, the voiceprint recognition device may extract MFCC (Mel Frequency Cepstrum Coefficient) or PLP (Perceptual Linear Predictive Coefficient) in the voice segment corresponding to each character as the voice corresponding to each character. The voiceprint features of the clip.

S607, the voiceprint feature of the voice segment corresponding to each character in the voice information is used as the training sample data, and the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the corresponding corresponding character, so that the estimation is verified. A feature vector corresponding to each character in the voice message.

Since a large number of experiments and papers have verified that the mean of each Gaussian module in the UBM model can be used to distinguish the identity information of the speaker, the voiceprint recognition device can verify the voiceprint feature of the voice segment corresponding to each character in the voice information as a training sample. Data, using Maximum A Posteriori (MAP) to adjust the mean supervector of the common background model corresponding to the corresponding corresponding character, that is, the voiceprint feature of the voice segment corresponding to each character in the verification voice information As an input sample After substituting into equation (1), by continuously adjusting the mean supervector, the posterior probability P(x) is maximized, so that the mean supervector with the largest posterior probability P(x) can be used as the corresponding feature of the corresponding character in the verification speech information. vector.

In another optional embodiment, in order to reduce the slow convergence speed caused by the high dimension of the super vector, the voiceprint recognition apparatus may adjust the mean super vector of the common background model corresponding to the preset corresponding character by using the following formula: The posterior probability of the universal background model corresponding to the adjusted corresponding characters is the largest:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector subspace matrix , ω is the feature vector corresponding to the corresponding character in the verification voice information, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1) as an input sample, the adjustment can be realized by continuously adjusting ω. The mean supervector in (1) maximizes the posterior probability P(x), so that ω that maximizes the posterior probability P(x) can be used as the feature vector corresponding to the corresponding character in the verification speech information.

S608. Calculate a similarity score of the feature vector corresponding to each character in the verification voice information and the corresponding character in the preset registration voice information. If the similarity score reaches a preset verification threshold, the verification user is determined to be the registered voice. The registered user corresponding to the information.

In this embodiment, the voiceprint recognition apparatus may calculate a cosine distance value between the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, and use the cosine distance value as The similarity score, that is, the similarity score between the corresponding feature vector in the verification voice information and the feature vector in the registered voice information is calculated by the following formula:

Wherein, the subscript i indicates a character common to the i-th verification voice information and the registration voice information, ω _i (tar) indicates a corresponding feature vector of the character in the verification voice information, and ω _i (test) indicates that the character is in the registered voice. The corresponding feature vector in the message. If the verification voice information and the registration voice information include multiple identical characters, the similarity scores of the characters calculated according to the above formula may be averaged. If the similarity scores of the respective characters reach the corresponding preset verification threshold, then The verification user is determined to be a registered user corresponding to the registration voice information. If there are multiple registered users, such as the registered users A, B, and C shown in FIG. 1, the registered user may be authenticated according to the similarity between the feature vector of a certain character of the user and the corresponding character of each registered user. The feature vector of the corresponding character and the feature vector of the character of the verification voice have the highest similarity score and the similarity reaches the preset verification threshold, and the registered user is used as the identification result of the verification user.

Therefore, in this embodiment, by verifying the similarity between the feature vector corresponding to each character in the verification voice information and the feature vector of the corresponding character in the registered voice information, and combining the timing judgment of the voice segment, the user identity of the user can be further verified. accuracy.

FIG. 7 is a schematic structural diagram of a voiceprint identifying apparatus according to an embodiment of the present invention. As shown in the figure, the voiceprint identifying apparatus in this embodiment may include:

The voice obtaining module 710 is configured to obtain verification voice information generated by the user to read the first character string.

The user who authenticates the user, that is, the unknown identity, needs to verify the identity of the user through the voiceprint recognition device. The first character string is a character string used for verifying the user's identity verification, and may be randomly generated, or may be a preset fixed string, for example, a second string corresponding to the pre-generated registration voice information. Part of the same string. Specifically, the first character string may include m characters, wherein n characters are different from each other, m and n are positive integers, and m≥n.

For example, the first string is "12358948", a total of 8 characters, including 7 different characters "1", "2", "3", "4", "5", "8", " 9".

The voice segment identification module 720 is configured to perform voice recognition on the verification voice information to obtain voice segments respectively corresponding to the plurality of characters in the first character string included in the verification voice information.

As shown in FIG. 3, the voice segment identification module 720 can divide the verification voice information into voice segments corresponding to multiple characters through voice recognition and voice intensity filtering, and optionally remove the invalid voice segments without participating. Subsequent processing.

In an optional embodiment, the voice segment identification module may further include:

The valid segment identification unit 721 is configured to identify a valid speech segment and an invalid speech segment in the verification speech information.

Specifically, the effective segment recognizing unit 721 can divide the verification speech according to the sound intensity, and treat the speech segment with a small sound intensity as an invalid speech segment (for example, including a silent segment and impulse noise).

a voice recognition unit 722, configured to perform voice recognition on the valid voice segment to obtain a separate voice A speech segment corresponding to a plurality of characters in the first character string.

The voiceprint feature extraction module 730 is configured to extract a voiceprint feature of the voice segment corresponding to each character in the verification voice information.

Specifically, the voiceprint feature extraction module 730 can extract MFCC (Mel Frequency Cepstrum Coefficient) or PLP (Perceptual Linear Predictive Coefficient) in the voice segment corresponding to each character, as corresponding to each character. The voiceprint features of the speech segment.

The feature model training module 740 is configured to train, according to the voiceprint feature of the voice segment corresponding to the respective characters, a feature vector corresponding to each character in the verification voice information according to the common background model corresponding to the preset corresponding character.

The feature model training module 740 may use the voiceprint feature of the voice segment corresponding to each character in the voice information as the training sample data, and adopt a maximum posterior probability algorithm (MAP) to the common background model corresponding to the preset corresponding character. The parameters are adjusted, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1) as an input sample, the parameters of the common background model corresponding to the corresponding corresponding characters are continuously adjusted, so that The probability P(x) is the largest, so that the feature model training module 740 can determine the feature vector corresponding to the corresponding character in the verification voice information according to the parameter that maximizes the posterior probability P(x).

Since a large number of experiments and papers have verified that the mean of each Gaussian module in the UBM model can be used to distinguish the identity information of the speaker, we define the mean supervector of the UBM model as:

Therefore, the feature model training module 740 can use the voiceprint feature of the voice segment corresponding to each character in the voice information as the training sample data, and use the Maximum A Posteriori (MAP) algorithm to correspond to the preset corresponding character. The mean supervector of the background model is adjusted In the whole, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1), the posterior probability P(x) is maximized by continuously adjusting the mean supervector, and the feature model training module 740 may use the mean supervector that maximizes the posterior probability P(x) as the feature vector corresponding to the corresponding character in the verification speech information.

In another alternative embodiment, in order to reduce the slow convergence rate caused by the high dimension of the supervector, we limit the variation range of the mean supervector by probabilistic principal component analysis (PPCA). In a subspace, the feature model training module 740 may use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, and adopt the maximum posterior probability algorithm to the common background model corresponding to the preset corresponding character. The mean super vector is adjusted, and the preset super vector subspace matrix is combined to obtain a feature vector corresponding to each character in the verified speech information. In a specific implementation, the feature model training module 740 may adjust the mean super vector of the common background model corresponding to the preset corresponding character by using the following formula, so that the posterior probability of the common background model corresponding to the adjusted corresponding character is the largest:

M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector subspace matrix , ω is the feature vector corresponding to the corresponding character in the verification voice information, that is, after the voiceprint feature of the voice segment corresponding to each character in the verification voice information is substituted into the formula (1) as an input sample, the adjustment can be realized by continuously adjusting ω. The mean supervector in (1) maximizes the posterior probability P(x), so that ω that maximizes the posterior probability P(x) can be used as the feature vector corresponding to the corresponding character in the verification speech information. The super-vector subspace matrix T is determined according to the correlation between each dimension vector in the mean supervector of the Gaussian mixture model.

The similarity determining module 750 is configured to calculate a similarity score of the feature vector corresponding to the corresponding character in the preset registered voice information in the verification voice information.

Specifically, the voiceprint recognition device can obtain the registered voice information of the registered user in the voiceprint registration stage, and can obtain the registered voice information through the voice segment identification module 720, the voiceprint feature extraction module 730, and the feature model training module 740. The feature vector corresponding to the speech segment of each character. The registration voice information may be that the voiceprint recognition device acquires the registration voice information generated by the registered user reading the second character string, and the second character string has at least one of the same characters as the first character string, that is, the The second character string corresponding to the registration voice information is at least partially identical to the first character string. and then In an optional embodiment, the voiceprint recognition device may further acquire a feature vector corresponding to the corresponding character in the registered voice information, that is, after the registered user inputs the registered voice information through other devices, the other device or the server passes the voiceprint feature. The extraction and voiceprint model training obtains a feature vector corresponding to the voice segment of each character in the registered voice information, and the voiceprint recognition device obtains the feature vector corresponding to the corresponding character in the registered voice information from other devices or servers, thereby verifying the user The identification stage similarity judgment module 750 is configured to compare the feature vectors corresponding to the respective characters in the verification voice information.

In a specific implementation, the similarity score is that the voiceprint recognition device compares the feature vector corresponding to each character in the voice information with the feature vector corresponding to the corresponding character in the preset registered voice information, and then measures two features of the same character. The degree of similarity between vectors. In an optional embodiment, the similarity determination module 750 may calculate a cosine distance value between the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, and the cosine distance value is The distance value is used as the similarity score, that is, the similarity score between the corresponding feature vector in the verification voice information and the feature vector in the registered voice information is calculated by the following formula:

Wherein, the subscript i indicates a character common to the i-th verification voice information and the registration voice information, ω _i (tar) indicates a corresponding feature vector of the character in the verification voice information, and ω _i (test) indicates that the character is in the registered voice. The corresponding feature vector in the message. In an optional embodiment, if the same character appears in the verification voice information more than once, for example, if 0, 1, 5, and 8 appear in the verification voice information as shown in FIG. 2, respectively, then The average of the similarity scores of the feature vectors processed by the two-character 0 corresponding to the speech segment and the feature vector of the character 0 in the preset registration speech information is used as the feature vector and preset of the character 0 in the verification speech information. The similarity score of the feature vector of character 0 in the registered voice information, and so on.

It should be noted that there are many ways to measure the similarity between two feature vectors. The above is only one embodiment provided by the present invention, and those skilled in the art can perform no creative work on the basis of the solution disclosed by the present invention. In order to obtain more ways of calculating the similarity scores of the feature vectors of the characters shared in the verification voice information and the registration voice information, the present invention need not be exhaustive.

a user identification module 760, configured to: if the similarity score reaches a preset verification threshold, The verification user determines the registered user corresponding to the registered voice information.

If the verification voice information and the registration voice information include a plurality of identical characters, the user identification module 760 may take the average value of the similarity scores of the respective characters calculated by the similarity determination module 750, if the average value of the similarity scores of the respective characters reaches the corresponding value. The preset verification threshold determines the authenticated user as the registered user corresponding to the registered voice information. If there are multiple registered users, such as the registered users A, B, and C shown in FIG. 1, the user identification module 760 may perform the similarity between the feature vector of a certain character of the user and the feature vector of the corresponding character of each registered user. If the feature vector of the corresponding character of a registered user and the feature vector of the character of the verification voice have the highest similarity score and the similarity reaches the preset verification threshold, the registered user is used as the identification result of the verification user.

In an optional embodiment, the voice acquiring module 710 is further configured to obtain registration voice information generated by the registered user reading the second character string, where the second character string has at least one of the same as the first character string. character of;

The voice segment identification module 720 is further configured to perform voice recognition on the registration voice information to obtain a voice segment respectively included in the registration voice information and corresponding to multiple characters in the second character string;

The voiceprint feature extraction module 730 is further configured to extract a voiceprint feature of a voice segment corresponding to each character in the registered voice information;

The feature model training module 740 is further configured to: according to the voiceprint feature of the voice segment corresponding to each character in the registered voice information, and the common background model corresponding to the preset corresponding character, the corresponding characters in the registered voice information are trained. Feature vector.

In an alternative embodiment, the voiceprint recognition apparatus may further include:

The character order determining module 770 is configured to determine whether the order of the voice segments of the plurality of characters in the verification voice information is consistent with the order of the corresponding characters in the first character string.

In order to effectively prevent the voice information of the registered user from being pirated or illegally copied for voiceprint recognition, a different first character string may be randomly generated each time, and in this step, the voice segment of the plurality of characters in the voice information is determined. Whether the sorting is consistent with the sorting of the corresponding characters in the first string, if not, it may be determined that the voiceprint recognition fails, and if the sorting of the corresponding characters in the first character string is consistent, the voiceprint feature extraction module 730 may be notified. Or feature model training module 740 performs feature extraction and voiceprint training for the verification voice information.

In an alternative embodiment, the voiceprint recognition apparatus may further include:

The string display module 700 is configured to randomly generate the first string and display it.

Therefore, in this embodiment, by obtaining the voiceprint feature of the voice segment corresponding to each character in the verification voice information of the verification user, the UTM training corresponding to the preset corresponding character is used to obtain the feature vector corresponding to each character in the verification voice information, and the verification is performed. The feature vector corresponding to each character in the voice information is compared with the feature vector of the corresponding character in the registered voice information, thereby determining the user identity of the verified user. The user feature vector used for comparison corresponds to the specific character, and the user is fully considered. The voiceprint features of different characters are read aloud, so that the accuracy of voiceprint recognition can be effectively improved.

In the actual test case, in the 1000-person training sample, 290,000 tests (in which the identity matching test is about 10,000 times, the mismatch test is about 280,000 times), it is able to achieve 79.8% of the one-thousandth error rate. The recall rate, EER (Equal Error Rate) is 3.39%. Compared with the traditional text-independent modeling method, the voiceprint recognition performance is improved by more than 40%.

FIG. 9 is a schematic structural diagram of another voiceprint recognition apparatus according to an embodiment of the present invention. As shown in FIG. 9, the voiceprint recognition apparatus 1000 may include at least one processor 1001 (eg, a CPU), a user interface 1003, a memory 1005, and at least one communication bus 1002. Among them, the communication bus 1002 is used to implement connection communication between these components. The user interface 1003 may include a display, a keyboard, a microphone, and the like. Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The memory 1005 may be a high speed RAM memory or a non-volatile memory such as at least one disk memory. As shown in FIG. 9, an operating system, a user interface module, and computer executable program code (eg, a voiceprint recognition program) may be included in the memory 1005 as a computer storage medium.

In the voiceprint recognition apparatus 1000 shown in FIG. 9, the processor 1001 can be used to call the computer executable program code stored in the memory 1005, and specifically perform the following steps:

Acquiring, by the user interface, verification voice information generated by the user to read the first character string;

Calculating a similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset voice information in the voice information of the verification voice information;

When the similarity reaches a preset threshold, the user verification is verified as the registered user corresponding to the registered voice information.

In an implementation manner, the obtaining, by the verification user, reading the first character string Before the verifying the voice information, the processor 1001 also calls the computer executable program code to perform the following operations:

Performing voice recognition on the verification voice information, and obtaining a voice segment respectively corresponding to each character in the first character string included in the verification voice information;

The voiceprint features of the voice segments corresponding to the respective characters are extracted.

In an implementation manner, the processor 1001 invokes the computer executable program code to perform the following operations to calculate a voice segment corresponding to a corresponding one of the preset voice information in the voice message corresponding to each character in the verification voice information. Similarity of voiceprint features:

And performing, according to the voiceprint feature of the voice segment corresponding to each character, in combination with a common background model corresponding to the preset corresponding character, obtaining a feature vector corresponding to each character in the verification voice information;

Calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, as a voice segment corresponding to each character in the verification voice information and a preset registration The similarity of the voiceprint features of the voice segments corresponding to the corresponding characters in the voice information.

In an implementation manner, before the obtaining the verification voice information generated by the verification user to read the first character string, the processor 1001 further invokes the computer executable program code to perform the following operations: acquiring Registering a user to read the registered voice information generated by the second character string, the second character string having at least one of the same characters as the first character string; performing voice recognition on the registered voice information to obtain the registered voice information a voice segment respectively corresponding to the plurality of characters in the second character string; extracting a voiceprint feature of the voice segment corresponding to each character in the registered voice information; and corresponding to each character in the registered voice information The voiceprint feature of the voice segment is trained in conjunction with a common background model corresponding to the corresponding corresponding character to obtain a feature vector corresponding to each character in the registered voice information.

In an implementation manner, the processor 1001 invokes the computer executable program code to perform the following operations, according to the voiceprint feature of the voice segment corresponding to the respective characters, and the common background corresponding to the preset corresponding character Model training, obtaining a feature vector corresponding to each character in the verification voice information: using a voiceprint feature of a voice segment corresponding to each character in the verification voice information as training sample data, and using a maximum posterior probability algorithm on the preset The mean supervector of the universal background model corresponding to the corresponding character is adjusted to obtain a feature vector corresponding to each character in the verified voice information.

In one implementation, the processor 1001 invokes the computer executable program code to perform the following operations to use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as training sample data, using the maximum The posterior probability algorithm adjusts the mean supervector of the universal background model corresponding to the preset corresponding character, thereby obtaining a feature vector corresponding to each character in the verification voice information: corresponding to each character in the verification voice information The voiceprint feature of the voice segment is used as the training sample data, and the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character, and combined with the preset supervector subspace matrix. Obtaining a feature vector corresponding to each character in the verification voice information.

In an implementation manner, the processor 1001 invokes the computer executable program code to perform the following operations to use the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, The maximum posterior probability algorithm adjusts the mean supervector of the universal background model corresponding to the preset corresponding character, and combines the preset supervector subspace matrix to obtain corresponding characters of the verified voice information. Feature vector: using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, and adjusting the mean super vector of the common background model corresponding to the preset corresponding character by using the following formula: The posterior probability of the universal background model corresponding to the adjusted corresponding character is maximized: M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, and m represents the generality of the corresponding character before the adjustment. Mean supervector of the background model, T is the preset supervector subspace matrix, and ω is the verification speech information Eigenvectors corresponding to the respective characters.

In an implementation manner, the preset super-vector sub-space matrix is determined according to a correlation between weights of respective Gaussian modules in the universal background model.

In an implementation manner, the processor 1001 invokes the computer executable program code to perform the following operations to calculate that a feature vector corresponding to each character in the verification voice information corresponds to a corresponding character in the preset registration voice information. The similarity score of the feature vector: calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and The cosine distance value is determined as the similarity score.

In an implementation manner, the processor 1001 invokes the computer executable program code to perform the following operations to perform voice recognition on the verification voice information to obtain the verification voice information package. a speech segment respectively corresponding to the plurality of characters in the first character string: identifying a valid speech segment and an invalid speech segment in the verification speech information; performing speech recognition on the valid speech segment respectively A speech segment corresponding to a plurality of characters in the first character string.

In an implementation manner, before the verifying the user is determined as the registered user corresponding to the registered voice information, the processor 1001 further invokes the computer executable program code to perform the following operations: determining the Verifying whether the ordering of the voice segments of the plurality of characters in the voice information is consistent with the ordering of the corresponding characters in the first character string; and when the similarity reaches a preset threshold value, and the verification voice information is In the case that the ordering of the speech segments of the plurality of characters is consistent with the ordering of the corresponding characters in the first character string, the verification user is determined as the registered user corresponding to the registration voice information.

In an implementation manner, before the obtaining the verification voice information generated by the verification user to read the first character string, the processor 1001 further invokes the computer executable program code to perform the following operations: random Generating the first character string and displaying the first character string.

One of ordinary skill in the art can understand that all or part of the process of implementing the foregoing embodiments can be completed by a computer program to instruct related hardware, and the program can be stored in a computer readable storage medium. When executed, the flow of an embodiment of the methods as described above may be included. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

The above is only the preferred embodiment of the present invention, and the scope of the present invention is not limited thereto, and thus equivalent changes made in the claims of the present invention are still within the scope of the present invention.

Claims (27)

A voiceprint recognition method, characterized in that the method comprises: Obtaining verification voice information generated by the verification user reading the first character string; Calculating a similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset voice information in the voice information of the verification voice information; When the similarity reaches a preset threshold, the user verification is verified as the registered user corresponding to the registered voice information. The voiceprint recognition method according to claim 1, wherein the calculating and verifying the similarity of the voiceprint features of the voice segments corresponding to the respective characters in the preset registered voice information in the voice information is verified. Previously included: Performing voice recognition on the verification voice information, and obtaining a voice segment respectively corresponding to each character in the first character string included in the verification voice information; The voiceprint features of the voice segments corresponding to the respective characters are extracted. The voiceprint recognition method according to claim 1, wherein the calculating and verifying the similarity of the voiceprint features of the voice segments corresponding to the respective characters in the preset registered voice information in the voice information is verified. include: And performing, according to the voiceprint feature of the voice segment corresponding to each character, in combination with a common background model corresponding to the preset corresponding character, obtaining a feature vector corresponding to each character in the verification voice information; Calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, as a voice segment corresponding to each character in the verification voice information and a preset registration The similarity of the voiceprint features of the voice segments corresponding to the corresponding characters in the voice information. The voiceprint recognition method according to claim 3, wherein the method further comprises: before the obtaining the verification voice information generated by the verification user reading the first character string, the method further comprising: Obtaining registration voice information generated by the registered user reading the second character string, the second string and The first string has at least one of the same characters; The feature vector corresponding to each character in the registered voice information is obtained according to the voiceprint feature corresponding to each character in the registered voice information, combined with the common background model corresponding to the preset corresponding character. The voiceprint recognition method according to claim 3, wherein the voiceprint feature according to the voice segment corresponding to each character is combined with the common background model corresponding to the preset corresponding character to obtain the Verifying the feature vectors corresponding to each character in the voice information includes: Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character, and obtain the The feature vector corresponding to each character in the voice information is verified. The voiceprint recognition method according to claim 5, wherein the voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as training sample data, and the maximum posterior probability algorithm is used Adjusting the mean supervector of the common background model corresponding to the preset corresponding character, so as to obtain the feature vector corresponding to each character in the verification voice information, including: The voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the maximum a posteriori probability algorithm is used to adjust the mean supervector of the common background model corresponding to the preset corresponding character. And combining the preset super vector subspace matrix to obtain a feature vector corresponding to each character in the verification voice information. The voiceprint recognition method according to claim 6, wherein the voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the maximum posterior probability algorithm is used. Adjusting the mean supervector of the common background model corresponding to the preset corresponding character, and combining the preset supervector subspace matrix, obtaining the feature vector corresponding to each character in the verification voice information includes: Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the mean supervector of the common background model corresponding to the preset corresponding character is adjusted by using the following formula, so that the adjusted The general background model corresponding to the corresponding character has the largest posterior probability: M=m+Tω, where M represents the mean supervector of the general background model of the adjusted character. m represents the mean supervector of the universal background model of the corresponding character before the adjustment, T is the preset supervector subspace matrix, and ω is the feature vector corresponding to the corresponding character in the verification speech information. The voiceprint recognition method according to claim 6, wherein the preset super-vector sub-space matrix is determined according to a correlation between weights of respective Gaussian modules in the universal background model. The voiceprint recognition method according to claim 5, wherein the mathematical expression of the universal background model is: P(x)=∑ _t=1...C a _i N(x|μ _i ,∑ _i ) Where P(x) represents the probability distribution of UBM, C represents the total of C Gaussian modules in UBM, a _i represents the weight of the i-th Gaussian module, μ _i represents the mean of the i-th Gaussian module, ∑ _i represents The variance of the i-th Gaussian module, N(x) represents a Gaussian distribution, and x represents the voiceprint characteristics of the input sample. The voiceprint recognition method according to claim 3, wherein the calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the corresponding character in the preset registration voice information includes: : Calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and determining the cosine distance value as the similarity score. The voiceprint recognition method according to claim 10, wherein the cosine distance value between the feature vector corresponding to each character in the voice information and the feature vector corresponding to the corresponding character in the preset registered voice information is calculated by the following formula :

Wherein, the subscript i indicates a character common to the i-th verification voice information and the registration voice information, ω _i (tar) indicates a corresponding feature vector of the character in the verification voice information, and ω _i (test) indicates that the character is in the registered voice. The corresponding feature vector in the message. The voiceprint recognition method according to claim 2, wherein the voice recognition of the verification voice information is performed, and the verification voice information is respectively included corresponding to a plurality of characters in the first character string. The voice clips include: Identifying valid voice segments and invalid voice segments in the verification voice information; Performing speech recognition on the valid speech segment to obtain a speech segment respectively corresponding to a plurality of characters in the first character string. The voiceprint recognition method according to claim 1, wherein the method further comprises: before the authenticating user is determined as the registered user corresponding to the registered voice information, the method further comprises: Determining whether the ordering of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string; And if the similarity reaches a preset threshold value, and the order of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string, The verification user determines the registered user corresponding to the registration voice information. The voiceprint recognition method according to any one of claims 1 to 11, wherein the method further comprises: before acquiring the verification voice information generated by the verification user reading the first character string include: The first character string is randomly generated, and the first character string is displayed. A voiceprint recognition device, characterized in that the device comprises: a voice acquiring module, configured to obtain verification voice information generated by the user to read the first character string; a voice segment identification module, configured to perform voice recognition on the verification voice information, and obtain a voice segment respectively included in the verification voice information corresponding to multiple characters in the first character string; a voiceprint feature extraction module, configured to extract a voiceprint feature of a voice segment corresponding to each character in the verification voice information; a feature model training module, configured to perform training according to a voice pattern feature of the voice segment corresponding to each character, and a common background model corresponding to the corresponding corresponding character, to obtain each of the verification voice information a feature vector corresponding to the character; The similarity judging module is configured to calculate a similarity score of the feature vector corresponding to each character in the verification voice information and the corresponding character in the preset registration voice information, as a voice segment corresponding to each character in the verification voice information The similarity of the voiceprint features of the voice segment corresponding to the corresponding character in the preset registration voice information; The user identification module is configured to verify the verification user as the registered user corresponding to the registered voice information when the similarity reaches a preset threshold. A voiceprint recognition device according to claim 15, wherein: The voice acquiring module is further configured to obtain registration voice information generated by the registered user reading the second character string, where the second character string and the first character string have at least one of the same characters; The feature model training module is further configured to: according to the voiceprint feature corresponding to each character in the registered voice information, and the common background model corresponding to the corresponding corresponding character, to obtain the feature corresponding to each character in the registered voice information. vector. The voiceprint recognition device according to claim 15, wherein said feature model training module is configured to: The voiceprint feature of the voice segment corresponding to each character in the verification voice information is used as the training sample data, and the mean supervector of the common background model corresponding to the preset corresponding character is adjusted by using the following formula, so that the adjusted corresponding The general background model corresponding to the character has the largest posterior probability: M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector a spatial matrix, ω is a feature vector corresponding to a corresponding character in the verification voice information, and the preset super-vector sub-space matrix is determined according to a correlation between each dimension vector in the mean super-vector of the Gaussian mixture model. of; The similarity judgment module is used to: Calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and determining the cosine distance value as the similarity score . The voiceprint recognition device according to claim 15, wherein the voice segment recognition module comprises: a valid segment identification unit, configured to identify a valid speech segment and an invalid speech segment in the verification speech information; The voice recognition unit is configured to perform voice recognition on the valid voice segment to obtain a voice segment corresponding to multiple characters in the first character string. The voiceprint recognition device according to claim 15, further comprising: a character order determining module, configured to determine whether a sorting of the voice segments of the plurality of characters in the verification voice information is consistent with a ranking of the corresponding characters in the first character string; The user identification module is further configured to: when the similarity reaches a preset threshold, and verify the order of the voice segments of the plurality of characters in the voice information and the corresponding characters in the first string In the case that the sorting is consistent, the verification user is determined as the registered user corresponding to the registered voice information. The voiceprint recognition device according to any one of claims 15 to 19, wherein the voiceprint recognition device further comprises: a string display module, configured to randomly generate the first string, and display the first string. A voiceprint recognition device includes: a user interface for obtaining voice information; a memory storing computer executable program code; a processor for invoking the computer executable program code to perform the following operations: Acquiring, by the user interface, verification voice information generated by the user to read the first character string; Performing voice recognition on the verification voice information, and obtaining a voice segment respectively corresponding to each character in the first character string included in the verification voice information; Extracting a voiceprint feature of the voice segment corresponding to each character; Corresponding to the preset corresponding characters according to the voiceprint features of the voice segments corresponding to the respective characters The universal background model is trained to obtain a feature vector corresponding to each character in the verification voice information; Calculating a similarity score of the feature vector corresponding to each character in the verification voice information and the feature vector corresponding to the corresponding character in the preset registration voice information, as a voice segment corresponding to each character in the verification voice information and a preset registration The similarity of the voiceprint features of the voice segments corresponding to the corresponding characters in the voice information; When the similarity reaches a preset threshold, the user verification is verified as the registered user corresponding to the registered voice information. The voiceprint recognition apparatus according to claim 21, wherein said processor further calls said processor before said verifying said verification voice information generated by said user reading said first character string The computer executable code performs the following actions: Obtaining registration voice information generated by the registered user reading the second character string, the second character string having at least one of the same characters as the first character string; The feature vector corresponding to each character in the registered voice information is obtained according to the voiceprint feature corresponding to each character in the registered voice information, combined with the common background model corresponding to the preset corresponding character. A voiceprint recognition apparatus according to claim 21, wherein said processor calls said computer executable program code to perform an operation of combining said sound according to a voiceprint feature of said voice segment corresponding to said respective character The general background model corresponding to the corresponding character is trained to obtain the feature vector corresponding to each character in the verification voice information: Using the voiceprint feature of the voice segment corresponding to each character in the verification voice information as the training sample data, the mean supervector of the common background model corresponding to the preset corresponding character is adjusted by using the following formula, so that the adjusted The general background model corresponding to the corresponding character has the largest posterior probability: M=m+Tω, where M represents the mean supervector of the universal background model of the adjusted character, m represents the mean supervector of the universal background model of the corresponding character before adjustment, and T is the preset supervector a spatial matrix, ω is a feature vector corresponding to a corresponding character in the verification voice information, and the preset super-vector sub-space matrix is determined according to a correlation between weights of respective Gaussian modules in the universal background model; The processor invoking the computer executable program code to perform the following operations to calculate the test The similarity score of the feature vector corresponding to each character in the voice information and the corresponding character in the preset registered voice information: Calculating a cosine distance value between a feature vector corresponding to each character in the verification voice information and a feature vector corresponding to a corresponding character in the preset registration voice information, and determining the cosine distance value as the similarity score. A voiceprint recognition apparatus according to claim 21, wherein said processor calls said computer executable program code to perform the following operations to perform voice recognition on said verification voice information to obtain said voice information included in said verification voice information a speech segment corresponding to a plurality of characters in the first character string: Identifying valid voice segments and invalid voice segments in the verification voice information; Performing speech recognition on the valid speech segment to obtain a speech segment respectively corresponding to a plurality of characters in the first character string. The voiceprint identifying apparatus according to claim 21, wherein said processor further calls said computer executable program before said verifying user determines said registered user corresponding to said registered voice information The code does the following: Determining whether the ordering of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string; And if the similarity reaches a preset threshold value, and the order of the voice segments of the plurality of characters in the verification voice information is consistent with the ordering of the corresponding characters in the first character string, The verification user determines the registered user corresponding to the registration voice information. A voiceprint recognition apparatus according to any one of claims 21 to 25, wherein said processor is obtained before said verification voice information generated by said verification user reading said first character string The computer executable program code is also invoked to perform the following operations: The first character string is randomly generated, and the first character string is displayed through the user interface. A storage medium, characterized in that a computer program is stored in the storage medium The computer program is for performing the voiceprint recognition method according to any one of claims 1-14.

Images