CN104765996B - Voiceprint password authentication method and system - Google Patents
Voiceprint password authentication method and system Download PDFInfo
- Publication number
- CN104765996B CN104765996B CN201410005651.XA CN201410005651A CN104765996B CN 104765996 B CN104765996 B CN 104765996B CN 201410005651 A CN201410005651 A CN 201410005651A CN 104765996 B CN104765996 B CN 104765996B
- Authority
- CN
- China
- Prior art keywords
- model
- cryptogram
- vocal print
- user
- unit
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 49
- 230000001755 vocal effect Effects 0.000 claims abstract description 139
- 238000012549 training Methods 0.000 claims abstract description 69
- 238000012546 transfer Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 abstract description 4
- 230000003044 adaptive effect Effects 0.000 description 7
- 230000008569 process Effects 0.000 description 7
- 230000000694 effects Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Landscapes
- Collating Specific Patterns (AREA)
Abstract
The invention discloses a kind of voiceprint password authentication method and system, belongs to cipher authentication technique field.This method includes：Receive voice signal input by user；Speech recognition is carried out to the voice signal, obtains cryptogram；Determine whether there is the corresponding background model of the cryptogram；If it is, obtain the background model；If it is not, then being extended according to the cryptogram to the pronunciation unit model that training obtains in advance, the corresponding background model of the cryptogram is obtained；The user is authenticated using the vocal print feature sequence in the voice signal, the background model and the vocal print cryptogram-modle of the user.Using this method and system, can meet the needs of User Defined password and frequently change password.
Description
Technical field
The present invention relates to cipher authentication technique field, more particularly to a kind of voiceprint password authentication method and system.
Background technology
Voiceprint, i.e., according to speak human physiological characteristics and the vocal print of behavioural characteristic reflected in user's input speech signal
Information, automatically confirms that the technology of speaker's identity.Compared to other biological authentication techniques, voiceprint have it is easier, accurate,
Numerous advantages such as economic and good autgmentability, can be widely applied to each side such as safety verification, control.
Vocal print cipher authentication, is the voiceprint authentication technology of text related (Text-Dependent) a kind of, the technical requirements
User speech input determines cryptogram, and confirms speaker's identity according to vocal print feature.In vocal print cipher authentication, user's note
It is required to input the relevant voice signal of definite cryptogram when volume and authentication, thus voiceprint is more consistent, accordingly
Obtain and compare the voiceprint more preferable certification effect unrelated with text, it is close in access control system, startup password, bank paying
Positive effect is played under the application environments such as code.
In the prior art, vocal print cipher authentication need offline collection in advance largely with the relevant voice number of cryptogram
Background model is obtained according to training.In practice, often require that user unifies text password, and cannot arbitrarily change password, ability
Meet the requirements.
Since the mode of user's one password of unified setting is easy to cause correct option or the cryptogram leakage of setting, into
And can cause largely to emit the person's of recognizing voice can be misidentified as target speaker, so that effective safety guarantee can not be provided.Therefore,
In order to improve security and meet that user individual sets demand, it is allowed to which different user is replaced using self-defined password or often
Password is just highly desirable.And training of the prior art to background model often requires that collection is largely relevant with cryptogram
Voice data, it is clear that user-defined vocal print cipher authentication system can not be suitable for, and be unfavorable for vocal print cipher authentication system
Password update.
The content of the invention
An embodiment of the present invention provides a kind of voiceprint password authentication method and system, to meet User Defined password and frequency
The application demand of numerous change password.
Technical solution provided in an embodiment of the present invention is as follows：
On the one hand, there is provided a kind of voiceprint password authentication method, including：
Receive voice signal input by user；
Speech recognition is carried out to the voice signal, obtains cryptogram；
The cryptogram trained beforehand through the password voice data gathered offline is determined whether there is to correspond to
Background model；
If it is, obtain the background model；If it is not, then according to the cryptogram to training obtained hair in advance
Sound model of element is spliced, and obtains the corresponding background model of the cryptogram；
Utilize the vocal print feature sequence in the voice signal, the background model and the vocal print cryptogram-modle of the user
The user is authenticated.
Preferably, training obtains pronunciation unit model in advance in the following way：
Obtain training voice data；
Pronunciation unit is determined according to the trained voice data；
Determine the topological structure of the acoustic model of the pronunciation unit；
Parameter training is carried out to the acoustic model, obtains pronunciation unit model.
Preferably, the acoustic model is GMM model；
It is described that the pronunciation unit model that training obtains in advance is spliced according to the cryptogram, obtain the password
The corresponding background model of text includes：
The corresponding GMM model of each pronunciation unit in the cryptogram is obtained, obtains GMM model set；
To the model unit in the GMM model set, spliced using equal weight, obtain new combination GMM model；
The Gauss weight of the combination GMM model is updated so that the sum of Gauss weight of the combination GMM model
For 1, the corresponding background model of the cryptogram is obtained.
Preferably, the acoustic model is GMM model；
It is described that the pronunciation unit model that training obtains in advance is spliced according to the cryptogram, obtain the password
The corresponding background model of text includes：
The corresponding GMM model of each pronunciation unit in the cryptogram is obtained, obtains GMM model sequence；
After splicing successively at least two model units in the GMM model sequence, probability is redirected certainly using default
Redirect probability with outer and carry out redirecting transfer, obtain the corresponding background model of the cryptogram, wherein, it is described from redirect probability with
Outer the sum of the probability that redirects is 1.
Preferably, the acoustic model is HMM model；
It is described that the pronunciation unit model that training obtains in advance is spliced according to the cryptogram, obtain the password
The corresponding background model of text includes：
Obtain the corresponding HMM model sequence of each pronunciation unit in the cryptogram；
After splicing successively to the model unit in the HMM model sequence, the corresponding background mould of the cryptogram is obtained
Type.
Preferably, the vocal print feature sequence using in the voice signal, the background model and the user
Vocal print cryptogram-modle is authenticated including to the user：
First likelihood score of the vocal print feature sequence relative to the vocal print cryptogram-modle, and the sound are calculated respectively
Line characteristic sequence relative to the background model the second likelihood score；
According to first likelihood score and the ratio and predetermined threshold value of second likelihood score, determine whether user is conjunction
Method user.
Preferably, the method further includes：Before speech recognition is carried out to the voice signal or described in acquisition
After the corresponding background model of cryptogram, the vocal print feature sequence in the voice signal is extracted.
Preferably, the method further includes：Before speech recognition is carried out to the voice signal or described in acquisition
After the corresponding background model of cryptogram, the vocal print cryptogram-modle of the user is obtained.
Preferably, the method further includes：Before the vocal print cryptogram-modle of the user is obtained, whether judgement currently deposits
In the vocal print cryptogram-modle of user；
If it does not exist, then according to the registration voice signal of user and the vocal print password mould of background model structure user
Type.
On the other hand, there is provided a kind of vocal print cipher authentication system, including：
Receiving module, for receiving voice signal input by user；
Identification module, for carrying out speech recognition to the voice signal, obtains cryptogram；
Determining module, be used to determine whether to exist and train beforehand through the password voice data gathered offline described in
The corresponding background model of cryptogram；
Background model acquisition module, for determining that there are the corresponding background model of the cryptogram in the determining module
Afterwards, the background model is obtained, after the determining module determines that the corresponding background model of the cryptogram is not present, according to
The cryptogram splices the pronunciation unit model that training obtains in advance, obtains the corresponding background mould of the cryptogram
Type；
Authentication module, for utilizing vocal print feature sequence, the background model and the user in the voice signal
Vocal print cryptogram-modle the user is authenticated.
Preferably, the system also includes：Training module, pronunciation unit model is obtained for training in advance；The training
Module includes：
Voice data acquiring unit, for obtaining trained voice data；
First determination unit, for determining pronunciation unit according to the trained voice data；
Second determination unit, the topological structure of the acoustic model for determining the pronunciation unit；
Parameter training unit, for carrying out parameter training to the acoustic model, obtains pronunciation unit model.
Preferably, the acoustic model is GMM model, and the background model acquisition module includes：
GMM model acquiring unit, for obtaining the corresponding GMM model of each pronunciation unit in the cryptogram, obtains GMM
Model set；
First concatenation unit, for the model unit in the GMM model set, being spliced using equal weight, being obtained new
Combination GMM model；
Weight updating block, for being updated to the Gauss weight of the combination GMM model so that the combination GMM
The sum of Gauss weight of model is 1, obtains the corresponding background model of the cryptogram.
Preferably, the acoustic model is GMM model, and the background model acquisition module includes：
GMM model acquiring unit, for obtaining the corresponding GMM model of each pronunciation unit in the cryptogram, obtains GMM
Model sequence；
Second concatenation unit, after splicing successively at least two model units in the GMM model sequence, uses
It is default from redirect probability and it is outer redirect probability and carry out redirecting transfer, obtain the corresponding background model of the cryptogram, wherein,
Described probability and outer the sum of probability that redirects of redirecting certainly is 1.
Preferably, the acoustic model is HMM model, and the background model acquisition module includes：
HMM model retrieval unit, for obtaining the corresponding HMM model sequence of each pronunciation unit in the cryptogram
Row；
3rd concatenation unit, after splicing successively to the model unit in the HMM model sequence, obtains the password
The corresponding background model of text.
Preferably, the authentication module includes：
Computing unit, for calculating first likelihood of the vocal print feature sequence relative to the vocal print cryptogram-modle respectively
Degree, and the vocal print feature sequence is relative to the second likelihood score of the background model；
Determination unit, for the ratio and predetermined threshold value according to first likelihood score and second likelihood score, really
Determine whether user is validated user.
Preferably, the system also includes：Vocal print feature sequential extraction procedures module, for carrying out language to the voice signal
Before sound identification or after the corresponding background model of the cryptogram is obtained, the vocal print extracted in the voice signal is special
Levy sequence.
Preferably, the system also includes：Vocal print cryptogram-modle acquisition module, for carrying out language to the voice signal
Before sound identification or after the corresponding background model of the cryptogram is obtained, the vocal print password mould of the user is obtained
Type.
Preferably, the system also includes：Judgment module, for before the vocal print cryptogram-modle of the user is obtained,
Judge currently with the presence or absence of the vocal print cryptogram-modle of user；
Vocal print cryptogram-modle build module, for the judgment module judge there is currently no vocal print cryptogram-modle after,
According to the registration voice signal of user and the vocal print cryptogram-modle of background model structure user.
Voiceprint password authentication method provided in an embodiment of the present invention and system, utilize the vocal print feature sequence in voice signal
Row, background model and the vocal print cryptogram-modle of user are authenticated user.If there is no the corresponding background mould of cryptogram
Type, then splice the pronunciation unit model that training obtains in advance according to cryptogram, obtain the corresponding background of cryptogram
Model., can be easily since background model is spliced by the pronunciation unit model for obtaining advance training
Real-time online generates, and disclosure satisfy that the demand that User Defined password and password are frequently changed.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, drawings in the following description are only one described in the present invention
A little embodiments, for those of ordinary skill in the art, can also obtain other attached drawings according to these attached drawings.
Fig. 1 is vocal print cipher authentication process schematic of the prior art；
Fig. 2 is the flow chart of voiceprint password authentication method provided in an embodiment of the present invention；
Fig. 3 is the structure diagram of vocal print cipher authentication system provided in an embodiment of the present invention.
Embodiment
In order to make those skilled in the art more fully understand the scheme of the embodiment of the present invention, below in conjunction with the accompanying drawings and embodiment party
Formula is described in further detail the embodiment of the present invention.
Vocal print cipher authentication process of the prior art is briefly described first below.
As shown in Figure 1, being vocal print cipher authentication process schematic in the prior art, comprise the following steps：
Step 101：Receive voice signal input by user；
Step 102：Extract the vocal print feature sequence in voice signal；
Step 103：Speech recognition is carried out to voice signal, obtains cryptogram；
Step 104：Obtain the corresponding background model of cryptogram that training obtains in advance；
Step 105：According to the vocal print cryptogram-modle of registration voice signal and background model the structure user of user；
Step 106：User is authenticated using vocal print feature sequence, background model and vocal print cryptogram-modle.
Vocal print cipher authentication of the prior art, it is main to use based on the assumption that the frame examined, i.e., calculate user and survey respectively
Voice is tried relative to background model and the likelihood score of vocal print cryptogram-modle, to carry out user authentication.
Due to the accuracy of the corresponding background model of cryptogram and vocal print cryptogram-modle, directly influence vocal print password and recognize
Demonstrate,prove effect.Under the statistical model setting of data-driven, the more big then modelling effect of amount of training data is better.In the prior art, carry on the back
Scape model can just be obtained often through substantial amounts of password voice data is gathered offline beforehand through data training.Due to adopting offline
Collect a large amount of password voice data and carry out data training, more troublesome, operability is poor.In practice, often through using single
One universal code, and the method that cannot arbitrarily change password is achieved.
Under mainstream GMM-UBM algorithm frames, be respectively adopted mixed Gauss model (Gaussian Mixture Model,
Abbreviation GMM) simulate universal code background model (Universal Background Model, abbreviation UBM) and vocal print password mould
Type.
Since vocal print password voice is usually shorter, for describing the back of the body of the user to the vocal print general character of universal code voicing text
Scape model generally selects Gaussage as 256 or the GMM model of bigger Gaussage.
Training obtains universal code background model by the following method：
(1) vocal print feature sequence is extracted from a large amount of password training voices prestored, obtains vocal print feature vector storehouse；
(2) using for example traditional LBG clustering algorithms of clustering algorithm, the vocal print feature vector of extraction is clustered, is obtained
To the initialization average of K Gauss, wherein K is the pre-set scale of model parameter of system.
(3) by EM algorithm (Expectation-Maximization algorithm, EM algorithm), iteration is more
Average, variance and the corresponding weighting coefficient of each Gauss of new GMM model, obtain new UBM model：
Wherein, T represents the totalframes of word, otRepresent that t frames observe data, cmIt is the weight of m-th of Gauss, meetsμm,ΣmIt is the average and variance of m-th of Gauss respectively, wherein, N () meets normal distribution：
Vocal print cryptogram-modle is obtained according to the registration voice on-line training of user.Since registration speech samples quantity has
The problems such as limit, directly trained complex model is easy to cause Sparse accordingly.Vocal print cryptogram-modle of the prior art is to carry on the back
Scape model is initial model, adjusts what model part parameter obtained by adaptive algorithm.Adaptive algorithm can be according to a small amount of
Personal data of speaking should be current speaker's individual character by user's vocal print general character is adaptive.Currently used adaptive algorithm has based on maximum
The adaptive algorithm of posterior probability (Maximum a Posteriori, abbreviation MAP).
Specific training process using adaptive algorithm training vocal print cryptogram-modle is as follows：
(1) vocal print feature sequence is extracted from user's registration voice, forms training characteristics vector storehouse；
(2) training characteristics vector storehouse is utilized, using the average of adaptive algorithm renewal universal background model mixed Gaussian, is obtained
To the Gaussian mean μ of vocal print cryptogram-modlem；Wherein, vocal print cryptogram-modle Gaussian mean μmFor sample statistic and common background mould
The weighted average of type Gaussian mean：
Wherein, xtRepresent observation vector, i.e. vocal print feature sequence；μmRepresent the average of m-th of Gauss；γm(xt) represent the
T frame vocal print features fall within the probability of m-th of Gauss, and τ is forgetting factor, for balancing history average and sample statistic to sound
The update intensity of line cryptogram-modle Gaussian mean.In general, τ values are larger, then vocal print cryptogram-modle Gaussian mean is mainly by original
Beginning average restricts；τ values are smaller, then vocal print cryptogram-modle Gaussian mean is mainly determined by sample statistic, more embody new sample
The characteristics of this distribution.
(3) background model variance is replicated, as vocal print cryptogram-modle variance.
After obtaining universal code background model and vocal print cryptogram-modle, it can be carried out according to voice signal input by user
User authentication.Specifically, according to the vocal print feature sequence of voice signal input by user, it is opposite that vocal print feature sequence is calculated respectively
In the first likelihood score of vocal print cryptogram-modle, vocal print feature sequence relative to universal code background model the second likelihood score, so
The ratio of the first likelihood score and the second likelihood score is calculated afterwards, judges whether user is legal use by the ratio and predetermined threshold value
Family.If the ratio of the first likelihood score and the second likelihood score is more than predetermined threshold value, judge user for validated user；Otherwise, sentence
Disconnected user is disabled user.
Vocal print cipher authentication of the prior art, in the case of single password, since sufficiently training number can be obtained
According to the background model and vocal print cryptogram-modle that training obtains are preferable, disclosure satisfy that certification demand.However, in User Defined sound
Under the application environment of line password, due to the use of the reason such as number is limited, it is difficult to obtain substantial amounts of training data, lead to not in advance
Training obtains corresponding background model so that traditional vocal print cipher authentication is needing User Defined password or frequently changing close
In the case of code, it can not be applied.
For this reason, the embodiment of the present invention proposes a kind of voiceprint password authentication method and system, it would be preferable to support User Defined
Password and the application demand for frequently replacing password.
As shown in Fig. 2, being the flow chart of voiceprint password authentication method provided in an embodiment of the present invention, following step can be included
Suddenly：
Step 201：Receive voice signal input by user；
Step 202：Speech recognition is carried out to voice signal, obtains cryptogram；
Step 203：Determine whether there is the corresponding background model of cryptogram；
Step 204：If there is the corresponding background model of cryptogram, then background model is obtained；
Step 205：If there is no the corresponding background model of cryptogram, then training in advance is obtained according to cryptogram
Pronunciation unit model spliced, obtain the corresponding background model of cryptogram；
Step 206：Using the vocal print feature sequence in voice signal, background model and the vocal print of user cryptogram-modle to
Family is authenticated.
In embodiments of the present invention, required vocal print feature sequence, background model and user are authenticated to user
Vocal print cryptogram-modle, can obtain in the following manner：Vocal print feature sequence can be extracted from voice signal input by user,
Specifically, can be before speech recognition be carried out to voice signal, naturally it is also possible to obtaining the corresponding background mould of cryptogram
After type, the extraction of the vocal print feature sequence of voice signal is carried out.On the acquisition of background model, can first judge whether
The corresponding background model of cryptogram, if existed, can directly acquire；If it does not, can be according to password text
This splices the pronunciation unit model that training obtains in advance, obtains the corresponding background model of cryptogram；It is close on vocal print
Code model, can first judge whether the vocal print cryptogram-modle of user, if it does, can directly acquire；If it does not,
Can be according to the vocal print password for the background model structure user that the registration voice signal and step 204 or step 205 of user obtains
Model.
Vocal print feature sequence, background model and the vocal print of user in the voice signal that can be obtained in the above manner is close
Code model is authenticated user, specifically, can calculate vocal print feature sequence respectively relative to the first of vocal print cryptogram-modle
Likelihood score, and vocal print feature sequence is relative to the second likelihood score of background model；According to the first likelihood score and the second likelihood score
Ratio and predetermined threshold value, determine whether user is validated user.If the ratio of the first likelihood score and the second likelihood score is big
In predetermined threshold value, it is determined that user is validated user；Otherwise, it determines user is disabled user.
In embodiments of the present invention, the corresponding background model of cryptogram can be by training obtained pronunciation in advance
What model of element was spliced.The training process of pronunciation unit model can specifically include：Training voice data is first obtained,
Then pronunciation unit is determined according to training voice data, then determines the topological structure of the acoustic model of pronunciation unit, then to acoustics
Model carries out parameter training, obtains pronunciation unit model.Wherein, pronunciation unit for example can be phoneme unit or syllable unit.
For example, " my birthday " corresponding phoneme unit for " w, o, d, e, sh, eng, r, i ", after being determined to phoneme unit,
Acoustic model can be established respectively to each phoneme unit therein, so as to obtain the topology knot of the acoustic model of phoneme unit
Structure, then parameter training is carried out to acoustic model, so as to obtain pronunciation unit model.Wherein, can to the parameter training of acoustic model
To use some training methods of the prior art, this embodiment of the present invention is not limited.
It should be noted that in embodiments of the present invention, above-mentioned acoustic model can use GMM model, can also use
HMM model (Hidden Markov Model, hidden Markov model).
When above-mentioned acoustic model uses GMM model, according to cryptogram to training obtained pronunciation unit model in advance
Spliced, obtaining the corresponding background model of cryptogram can include：Obtain the corresponding GMM of each pronunciation unit in cryptogram
Model, obtains GMM model set；To the model unit in GMM model set, spliced using equal weight, obtain new combination GMM
Model；The Gauss weight for combining GMM model is updated so that it is 1 to combine the sum of Gauss weight of GMM model, is obtained close
The corresponding background model of code text.
The Gauss weight renewal process of above-mentioned each model unit can be as follows：Equal weight splicing is being carried out to each model unit
During, Gauss weight is arranged to wn, n=1:N, N are model unit number；The Gauss weight of each model unit is updated to
w′ni=wnwni, wniFor the weight of each Gauss in master mould unit, and meet wni=1；As can be seen that splicing from above formula
GMM model afterwards remains able to meet after the Gauss weight of each model unit is updated
When above-mentioned acoustic model uses GMM model, according to cryptogram to training obtained pronunciation unit model in advance
Spliced, obtaining the corresponding background model of cryptogram can include：Obtain the corresponding GMM of each pronunciation unit in cryptogram
Model, obtains GMM model sequence；After splicing successively at least two model units in GMM model sequence, using it is default from
Redirect probability and the outer probability that redirects carry out redirecting transfer, obtain the corresponding background model of cryptogram, wherein, redirect certainly probability with
Outer the sum of the probability that redirects is 1.
When above-mentioned acoustic model uses HMM model, according to cryptogram to training obtained pronunciation unit model in advance
Spliced, obtaining the corresponding background model of cryptogram includes：Obtain the corresponding HMM model sequence of cryptogram；To HMM moulds
After at least two model units splicing in type sequence, using it is default from redirect probability and it is outer redirect probability and redirect turn
Move, obtain the corresponding background model of cryptogram, wherein, probability and outer the sum of the probability that redirects are redirected certainly as 1.
In embodiments of the present invention, redirect certainly and redirect can be carried out in units of word outside, therefore, it is necessary to two
Or two or more model unit is spliced, to obtain the GMM model of some word or HMM model, afterwards again in units of word
Set it and redirect transition probability.Wherein, redirecting probability certainly and redirecting probability outside rule of thumb to be set in advance, for example,
It will can certainly redirect probability and be arranged to 0.9, can be by redirecting probability is arranged to 0.1 outside.
Since pronunciation has certain when ductility, vocal print multidate information and password can preferably be simulated using HMM model
The timing information of text, effectively to solve out of order text, wrongly written character, multiword and few word phenomenon.
Voiceprint password authentication method provided in an embodiment of the present invention, utilizes vocal print feature sequence, the background in voice signal
Model and the vocal print cryptogram-modle of user are authenticated user.If there is no the corresponding background model of cryptogram, then root
The pronunciation unit model that training obtains in advance is spliced according to cryptogram, obtains the corresponding background model of cryptogram.By
In background model spliced by the pronunciation unit model for obtaining advance training, can easily real-time online
Generation, disclosure satisfy that the demand that User Defined password and password are frequently changed.
Correspondingly, the embodiment of the present invention also provides a kind of vocal print cipher authentication system, as shown in figure 3, implementing for the present invention
The structure diagram for the vocal print cipher authentication system that example provides.
In this embodiment, the vocal print cipher authentication system can include：
Receiving module 301, for receiving voice signal input by user；
Identification module 302, for carrying out speech recognition to voice signal, obtains cryptogram；
Determining module 303, is used to determine whether that there are the corresponding background model of cryptogram；
Background model acquisition module 304, for determining that there are the corresponding background model of cryptogram in determining module 303
Afterwards, background model is obtained, after determining module 303 determines that the corresponding background model of cryptogram is not present, according to cryptogram
The pronunciation unit model that training obtains in advance is spliced, obtains the corresponding background model of cryptogram；
Authentication module 305, for utilizing vocal print feature sequence, background model and the vocal print of the user password in voice signal
Model is authenticated user.
In embodiments of the present invention, the corresponding background model of cryptogram can be to trained in advance according to cryptogram
To pronunciation unit model spliced, and pronunciation unit model can in advance be trained and obtained.It is for this reason, of the invention
The system of embodiment can also further comprise：Training module (not shown), pronunciation unit model is obtained for training in advance.The instruction
Practicing module can include：
Voice data acquiring unit, for obtaining trained voice data；
First determination unit, for determining pronunciation unit according to training voice data；
Second determination unit, the topological structure of the acoustic model for determining pronunciation unit；
Parameter training unit, for carrying out parameter training to acoustic model, obtains pronunciation unit model.
It should be noted that in embodiments of the present invention, above-mentioned acoustic model can use GMM model, can also use
HMM model.
Above-mentioned acoustic model uses GMM model, and a kind of specific implementation structure of above-mentioned background model acquisition module 304 can be with
Including：
GMM model acquiring unit, for obtaining the corresponding GMM model of each pronunciation unit in cryptogram, obtains GMM model
Set；
First concatenation unit, for the model unit in GMM model set, being spliced using equal weight, obtaining new group
Close GMM model；
Weight updating block, for being updated to the Gauss weight for combining GMM model so that combine the height of GMM model
The sum of this weight is 1, obtains the corresponding background model of cryptogram.
Above-mentioned acoustic model uses GMM model, another specific implementation structure of above-mentioned background model acquisition module 304
It can include：
GMM model acquiring unit, for obtaining the corresponding GMM model of each pronunciation unit in the cryptogram, obtains GMM
Model sequence；
Second concatenation unit, after splicing successively at least two model units in the GMM model sequence, uses
It is default from redirect probability and it is outer redirect probability and carry out redirecting transfer, obtain the corresponding background model of the cryptogram, wherein,
Described probability and outer the sum of probability that redirects of redirecting certainly is 1.
Above-mentioned acoustic model uses HMM model, and a kind of specific implementation structure of above-mentioned background model acquisition module 304 can be with
Including：
HMM model retrieval unit, for obtaining the corresponding HMM model sequence of each pronunciation unit in the cryptogram
Row；
3rd concatenation unit, after splicing successively to the model unit in the HMM model sequence, obtains cryptogram
Corresponding background model.
In embodiments of the present invention, a kind of concrete structure of authentication module 305 can include：
Computing unit, for calculating first likelihood score of the vocal print feature sequence relative to vocal print cryptogram-modle respectively, and
Vocal print feature sequence relative to background model the second likelihood score；
Determination unit, for the ratio and predetermined threshold value according to the first likelihood score and the second likelihood score, determines that user is
No is validated user.
Vocal print cipher authentication system provided in an embodiment of the present invention, can further include：
Vocal print feature sequential extraction procedures module, for before speech recognition is carried out to voice signal or literary obtaining password
After this corresponding background model, the vocal print feature sequence in voice signal is extracted.
Further, said system can also include：
Vocal print cryptogram-modle acquisition module, for before speech recognition is carried out to voice signal or literary obtaining password
After this corresponding background model, the vocal print cryptogram-modle of user is obtained.
Further, said system can also include：
Judgment module, for before the vocal print cryptogram-modle of user is obtained, judging currently with the presence or absence of the vocal print of user
Cryptogram-modle；
Vocal print cryptogram-modle build module, for judgment module judge there is currently no vocal print cryptogram-modle after, according to
The vocal print cryptogram-modle of registration voice signal and background model the structure user of user.
Vocal print cipher authentication system provided in an embodiment of the present invention, utilizes vocal print feature sequence, the background in voice signal
Model and the vocal print cryptogram-modle of user are authenticated user.If there is no the corresponding background model of cryptogram, then root
The pronunciation unit model that training obtains in advance is spliced according to cryptogram, obtains the corresponding background model of cryptogram.By
In background model spliced by the pronunciation unit model for obtaining advance training, can easily real-time online
Generation, disclosure satisfy that the demand that User Defined password and password are frequently changed.
The voiceprint password authentication method and system that above-described embodiment provides belong to same inventive concept, each module in system,
The function of unit and realize that process can refer to the description in embodiment of the method, which is not described herein again.
Each embodiment in this specification is described by the way of progressive, identical similar portion between each embodiment
Divide mutually referring to what each embodiment stressed is the difference with other embodiment.It is real especially for system
For applying example, since it is substantially similar to embodiment of the method, so describing fairly simple, related part is referring to embodiment of the method
Part explanation.System embodiment described above is only schematical, wherein described be used as separating component explanation
Unit may or may not be physically separate, may or may not be as the component that unit is shown
Physical location, you can with positioned at a place, or can also be distributed in multiple network unit.Can be according to the actual needs
Some or all of module therein is selected to realize the purpose of this embodiment scheme.Those of ordinary skill in the art are not paying
In the case of creative work, you can to understand and implement.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all the present invention spirit and
Within principle, any modification, equivalent replacement, improvement and so on, should all be included in the protection scope of the present invention.
Claims (18)
- A kind of 1. voiceprint password authentication method, it is characterised in that including：Receive voice signal input by user；Speech recognition is carried out to the voice signal, obtains cryptogram；Determine whether there is the corresponding back of the body of the cryptogram trained beforehand through the password voice data gathered offline Scape model；If it is, obtain the background model；If it is not, then according to the cryptogram to training obtained pronunciation list in advance Meta-model is spliced, and obtains the corresponding background model of the cryptogram；Using the vocal print feature sequence in the voice signal, the background model and the vocal print cryptogram-modle of the user to institute User is stated to be authenticated.
- 2. according to the method described in claim 1, it is characterized in that, training obtains pronunciation unit mould in advance in the following way Type：Obtain training voice data；Pronunciation unit is determined according to the trained voice data；Determine the topological structure of the acoustic model of the pronunciation unit；Parameter training is carried out to the acoustic model, obtains pronunciation unit model.
- 3. according to the method described in claim 2, it is characterized in that, the acoustic model is GMM model；It is described that the pronunciation unit model that training obtains in advance is spliced according to the cryptogram, obtain the cryptogram Corresponding background model includes：The corresponding GMM model of each pronunciation unit in the cryptogram is obtained, obtains GMM model set；To the model unit in the GMM model set, spliced using equal weight, obtain new combination GMM model；The Gauss weight of the combination GMM model is updated so that the sum of Gauss weight of the combination GMM model is 1, Obtain the corresponding background model of the cryptogram.
- 4. according to the method described in claim 2, it is characterized in that, the acoustic model is GMM model；It is described that the pronunciation unit model that training obtains in advance is spliced according to the cryptogram, obtain the cryptogram Corresponding background model includes：The corresponding GMM model of each pronunciation unit in the cryptogram is obtained, obtains GMM model sequence；After splicing successively at least two model units in the GMM model sequence, using default from redirecting probability and outer Redirect probability to carry out redirecting transfer, obtain the corresponding background model of the cryptogram, wherein, it is described to redirect probability and outer jump certainly Turn the sum of probability as 1.
- 5. according to the method described in claim 2, it is characterized in that, the acoustic model is HMM model；It is described that the pronunciation unit model that training obtains in advance is spliced according to the cryptogram, obtain the cryptogram Corresponding background model includes：Obtain the corresponding HMM model sequence of each pronunciation unit in the cryptogram；After splicing successively to the model unit in the HMM model sequence, the corresponding background model of the cryptogram is obtained.
- 6. the according to the method described in claim 1, it is characterized in that, vocal print feature sequence using in the voice signal Row, the background model and the vocal print cryptogram-modle of the user are authenticated including to the user：First likelihood score of the vocal print feature sequence relative to the vocal print cryptogram-modle, and vocal print spy are calculated respectively Levy second likelihood score of the sequence relative to the background model；According to first likelihood score and the ratio and predetermined threshold value of second likelihood score, determine whether user is legal use Family.
- 7. method according to any one of claims 1 to 6, it is characterised in that the method further includes：Before speech recognition is carried out to the voice signal or after the corresponding background model of the cryptogram is obtained, Extract the vocal print feature sequence in the voice signal.
- 8. method according to any one of claims 1 to 6, it is characterised in that the method further includes：Before speech recognition is carried out to the voice signal or after the corresponding background model of the cryptogram is obtained, Obtain the vocal print cryptogram-modle of the user.
- 9. according to the method described in claim 8, it is characterized in that, the method further includes：Before the vocal print cryptogram-modle of the user is obtained, judge currently to whether there is the vocal print cryptogram-modle of user；If it does not exist, then according to the registration voice signal of user and the vocal print cryptogram-modle of background model structure user.
- A kind of 10. vocal print cipher authentication system, it is characterised in that including：Receiving module, for receiving voice signal input by user；Identification module, for carrying out speech recognition to the voice signal, obtains cryptogram；Determining module, is used to determine whether there is the password trained beforehand through the password voice data gathered offline The corresponding background model of text；Background model acquisition module, for after the determining module determines there are the corresponding background model of the cryptogram, The background model is obtained, after the determining module determines that the corresponding background model of the cryptogram is not present, according to institute State cryptogram to splice the pronunciation unit model that training obtains in advance, obtain the corresponding background mould of the cryptogram Type, pronunciation unit model to training voice data by being trained to obtain；Authentication module, for utilizing vocal print feature sequence, the background model and the sound of the user in the voice signal Line cryptogram-modle is authenticated the user.
- 11. system according to claim 10, it is characterised in that the system also includes：Training module, for instructing in advance Get pronunciation unit model；The training module includes：Voice data acquiring unit, for obtaining trained voice data；First determination unit, for determining pronunciation unit according to the trained voice data；Second determination unit, the topological structure of the acoustic model for determining the pronunciation unit；Parameter training unit, for carrying out parameter training to the acoustic model, obtains pronunciation unit model.
- 12. system according to claim 11, it is characterised in that the acoustic model is GMM model, the background model Acquisition module includes：GMM model acquiring unit, for obtaining the corresponding GMM model of each pronunciation unit in the cryptogram, obtains GMM model Set；First concatenation unit, for the model unit in the GMM model set, being spliced using equal weight, obtaining new group Close GMM model；Weight updating block, for being updated to the Gauss weight of the combination GMM model so that the combination GMM model The sum of Gauss weight be 1, obtain the corresponding background model of the cryptogram.
- 13. system according to claim 11, it is characterised in that the acoustic model is GMM model, the background model Acquisition module includes：GMM model acquiring unit, for obtaining the corresponding GMM model of each pronunciation unit in the cryptogram, obtains GMM model Sequence；Second concatenation unit, after splicing successively at least two model units in the GMM model sequence, using default Redirect certainly probability and it is outer redirect probability and carry out redirecting transfer, obtain the corresponding background model of the cryptogram, wherein, it is described From probability and outer the sum of the probability that redirects is redirected for 1.
- 14. system according to claim 11, it is characterised in that the acoustic model is HMM model, the background model Acquisition module includes：HMM model retrieval unit, for obtaining the corresponding HMM model sequence of each pronunciation unit in the cryptogram；3rd concatenation unit, after splicing successively to the model unit in the HMM model sequence, obtains the cryptogram Corresponding background model.
- 15. system according to claim 10, it is characterised in that the authentication module includes：Computing unit, for calculating first likelihood score of the vocal print feature sequence relative to the vocal print cryptogram-modle respectively, And the vocal print feature sequence is relative to the second likelihood score of the background model；Determination unit, for the ratio and predetermined threshold value according to first likelihood score and second likelihood score, determines to use Whether family is validated user.
- 16. according to claim 10 to 15 any one of them system, it is characterised in that the system also includes：Vocal print feature sequential extraction procedures module, for before speech recognition is carried out to the voice signal or described close obtaining After the corresponding background model of code text, the vocal print feature sequence in the voice signal is extracted.
- 17. according to claim 10 to 15 any one of them system, it is characterised in that the system also includes：Vocal print cryptogram-modle acquisition module, for before speech recognition is carried out to the voice signal or described close obtaining After the corresponding background model of code text, the vocal print cryptogram-modle of the user is obtained.
- 18. system according to claim 17, it is characterised in that the system also includes：Judgment module, for before the vocal print cryptogram-modle of the user is obtained, judging currently with the presence or absence of the vocal print of user Cryptogram-modle；Vocal print cryptogram-modle build module, for the judgment module judge there is currently no vocal print cryptogram-modle after, according to The registration voice signal of user and the vocal print cryptogram-modle of background model structure user.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410005651.XA CN104765996B (en) | 2014-01-06 | 2014-01-06 | Voiceprint password authentication method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410005651.XA CN104765996B (en) | 2014-01-06 | 2014-01-06 | Voiceprint password authentication method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104765996A CN104765996A (en) | 2015-07-08 |
| CN104765996B true CN104765996B (en) | 2018-04-27 |
Family
ID=53647821
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410005651.XA Active CN104765996B (en) | 2014-01-06 | 2014-01-06 | Voiceprint password authentication method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104765996B (en) |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106373575B (en) * | 2015-07-23 | 2020-07-21 | 阿里巴巴集团控股有限公司 | User voiceprint model construction method, device and system |
| CN106782572B (en) * | 2017-01-22 | 2020-04-07 | 清华大学 | Voice password authentication method and system |
| CN107068154A (en) * | 2017-03-13 | 2017-08-18 | 平安科技（深圳）有限公司 | The method and system of authentication based on Application on Voiceprint Recognition |
| CN107274906A (en) * | 2017-06-28 | 2017-10-20 | 百度在线网络技术（北京）有限公司 | Voice information processing method, device, terminal and storage medium |
| GB201801527D0 (en) * | 2017-07-07 | 2018-03-14 | Cirrus Logic Int Semiconductor Ltd | Method, apparatus and systems for biometric processes |
| CN110310647B (en) | 2017-09-29 | 2022-02-25 | 腾讯科技（深圳）有限公司 | Voice identity feature extractor, classifier training method and related equipment |
| CN109872721A (en) * | 2017-12-05 | 2019-06-11 | 富士通株式会社 | Voice authentication method, information processing device, and storage medium |
| CN109727342A (en) * | 2018-07-06 | 2019-05-07 | 平安科技（深圳）有限公司 | Recognition methods, device, access control system and the storage medium of access control system |
| CN111292733A (en) * | 2018-12-06 | 2020-06-16 | 阿里巴巴集团控股有限公司 | Voice interaction method and device |
| CN110880327B (en) * | 2019-10-29 | 2024-07-09 | 平安科技（深圳）有限公司 | Audio signal processing method and device |
| CN114003878A (en) * | 2021-10-25 | 2022-02-01 | 国电南京自动化股份有限公司 | Double-factor authentication method, device and system based on voice recognition |
| TWI876358B (en) * | 2023-05-30 | 2025-03-11 | 兆豐國際商業銀行股份有限公司 | Identity recognition system and method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1412741A (en) * | 2002-12-13 | 2003-04-23 | 郑方 | Chinese speech identification method with dialect background |
| CN1447278A (en) * | 2002-11-15 | 2003-10-08 | 郑方 | Method for recognizing voice print |
| CN1652206A (en) * | 2005-04-01 | 2005-08-10 | 郑方 | Sound veins identifying method |
| CN102238189A (en) * | 2011-08-01 | 2011-11-09 | 安徽科大讯飞信息科技股份有限公司 | Voiceprint password authentication method and system |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101154380B (en) * | 2006-09-29 | 2011-01-26 | 株式会社东芝 | Method and device for registration and validation of speaker's authentication |
-
2014
- 2014-01-06 CN CN201410005651.XA patent/CN104765996B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1447278A (en) * | 2002-11-15 | 2003-10-08 | 郑方 | Method for recognizing voice print |
| CN1412741A (en) * | 2002-12-13 | 2003-04-23 | 郑方 | Chinese speech identification method with dialect background |
| CN1652206A (en) * | 2005-04-01 | 2005-08-10 | 郑方 | Sound veins identifying method |
| CN102238189A (en) * | 2011-08-01 | 2011-11-09 | 安徽科大讯飞信息科技股份有限公司 | Voiceprint password authentication method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104765996A (en) | 2015-07-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104765996B (en) | Voiceprint password authentication method and system | |
| CN108510976B (en) | Multi-language mixed voice recognition method | |
| KR102198835B1 (en) | End-to-end speaker recognition using deep neural network | |
| CN105336322B (en) | Polyphone model training method, and speech synthesis method and device | |
| US10176811B2 (en) | Neural network-based voiceprint information extraction method and apparatus | |
| Diez et al. | Speaker Diarization based on Bayesian HMM with Eigenvoice Priors. | |
| Larcher et al. | Text-dependent speaker verification: Classifiers, databases and RSR2015 | |
| CN102354495B (en) | Test method and system for semi-open oral test questions | |
| CN108766415B (en) | A method of voice assessment | |
| CN110050302A (en) | Speech synthesis | |
| Chen et al. | Pronunciation and silence probability modeling for ASR. | |
| CN109313892A (en) | Steady language identification method and system | |
| WO2006034200A2 (en) | Method and system for the automatic generation of speech features for scoring high entropy speech | |
| CN110415725B (en) | Method and system for evaluating pronunciation quality of second language using first language data | |
| Yin et al. | Automatic cognitive load detection from speech features | |
| Zhong et al. | DNN i-Vector Speaker Verification with Short, Text-Constrained Test Utterances. | |
| Serizel et al. | Deep neural network adaptation for children's and adults' speech recognition | |
| CN103594087A (en) | Method and system for improving oral evaluation performance | |
| CN109300339A (en) | A kind of exercising method and system of Oral English Practice | |
| US20180012602A1 (en) | System and methods for pronunciation analysis-based speaker verification | |
| Jyothi et al. | Lexical access experiments with context-dependent articulatory feature-based models | |
| KR20220071960A (en) | A method and a TTS system for calculating an encoder score of an attention alignment corresponded to a spectrogram | |
| Behravan et al. | Introducing attribute features to foreign accent recognition | |
| TW201937479A (en) | Multilingual mixed speech recognition method | |
| CN112863476B (en) | Personalized speech synthesis model construction, speech synthesis and test methods and devices |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB03 | Change of inventor or designer information |
Inventor after: Yin Bing Inventor after: Jin Xin Inventor after: Wei Si Inventor after: Liu Lin Inventor after: Liu Junhua Inventor after: Zhang Chengfeng Inventor after: Wang Jianshe Inventor after: Liu Haibo Inventor after: Hu Guoping Inventor after: Hu Yu Inventor before: Yin Bing Inventor before: Wei Si Inventor before: Hu Guoping Inventor before: Hu Yu |
|
| CB03 | Change of inventor or designer information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |