CN108564955A - Electronic device, auth method and computer readable storage medium - Google Patents
Electronic device, auth method and computer readable storage medium Download PDFInfo
- Publication number
- CN108564955A CN108564955A CN201810225887.2A CN201810225887A CN108564955A CN 108564955 A CN108564955 A CN 108564955A CN 201810225887 A CN201810225887 A CN 201810225887A CN 108564955 A CN108564955 A CN 108564955A
- Authority
- CN
- China
- Prior art keywords
- voice data
- preset
- layer
- neural network
- frame group
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification
- G10L17/02—Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
- G10L25/24—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being the cepstrum
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification
- G10L17/04—Training, enrolment or model building
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification
- G10L17/18—Artificial neural networks; Connectionist approaches
Abstract
The present invention discloses a kind of electronic device, auth method and storage medium, this method：After receiving the current speech data of target user of pending authentication, the corresponding standard voice data of identity to be verified is obtained, the two standard voice datas are distinguished into sub-frame processing, to obtain current speech frame group and received pronunciation frame group；Extract the preset kind acoustic feature of each speech frame in two speech frame groups respectively using Predetermined filter；The preset kind acoustic feature extracted is inputted into trained preset structure deep neural network model in advance respectively, to obtain the characteristic vector of current speech data and the corresponding preset length of standard voice data；The cosine similarity for two characteristic vectors being calculated, and authentication result is determined according to calculated cosine similarity size.Technical solution of the present invention improves the accuracy of speaker's identity verification.
Description
Technical field
The present invention relates to sound groove recognition technology in e field, more particularly to a kind of electronic device, auth method and computer
Readable storage medium storing program for executing.
Background technology
Speaker Identification is commonly referred to as Application on Voiceprint Recognition, is one kind of biological identification technology, is often used to confirm that certain section of voice
Whether it is described in specified someone, is " one-to-one differentiation " problem.Speaker Identification is widely used in numerous areas, for example,
The demand of being widely applied is suffered from fields such as finance, security, social security, public security, army and other civil safety certifications.
Speaker Identification includes text Classical correlation and the unrelated identification two ways of text, in recent years the unrelated speaker of text
Identification technology is constantly broken through, and accuracy has great promotion as compared with the past.However under certain limited situations, for example acquire
In the case of the speaker's efficient voice shorter (duration is less than 5 seconds voices) arrived, the existing unrelated Speaker Identification skill of text
The accuracy of art is not high, it is easy to malfunction.
Invention content
The main object of the present invention is to provide a kind of electronic device, auth method and computer readable storage medium,
It is intended to promote the accuracy of speaker's identity verification.
To achieve the above object, electronic device proposed by the present invention, including memory and processor are deposited on the memory
The authentication system that can be run on the processor is contained, the authentication system is realized when being executed by the processor
Following steps：
After receiving the current speech data of target user of pending authentication, obtained from database to be verified
The corresponding standard voice data of identity, by the current speech data and standard voice data respectively according to preset framing parameter
Sub-frame processing is carried out, it is corresponding to obtain the corresponding current speech frame group of the current speech data and the standard voice data
Received pronunciation frame group；
Extracted respectively using Predetermined filter in current speech frame group the preset kind acoustic feature of each speech frame and
The preset kind acoustic feature of each speech frame in received pronunciation frame group；
It is respectively that the corresponding preset kind acoustic feature of current speech frame group and received pronunciation frame group that extract is corresponding
The advance trained preset structure deep neural network model of preset kind acoustic feature input, to obtain the current speech number
According to the characteristic vector with the corresponding preset length of the standard voice data；
The cosine similarity for two characteristic vectors being calculated, and body is determined according to calculated cosine similarity size
Part verification result, the authentication result include being verified result and authentication failed result.
Preferably, divided respectively according to preset framing parameter by the current speech data and standard voice data
Before the step of frame processing, which is additionally operable to execute the authentication system, to realize following steps：
Active endpoint detection is carried out to the current speech data and standard voice data respectively, by the current speech number
According to the voice deletion with the non-talking people in the standard voice data.
Preferably, the training process of the preset structure deep neural network model is：
S1, preset quantity voice data sample is obtained, each voice data sample is marked respectively and represents corresponding speak
The label of personal part；
S2, active endpoint detection is carried out to each voice data sample respectively, by non-talking people in voice data sample
Voice is deleted, and the standard voice data sample of preset quantity is obtained；
S3, using the first percentage of obtained standard voice data sample as training set, the second percentage is as verification
Collection, first percentage and the second percentage and less than or equal to 100%；
S4, each standard voice data sample that the training set and verification are concentrated is distinguished according to preset framing parameter
Sub-frame processing is carried out, to obtain the corresponding speech frame group of each standard voice data sample, Predetermined filter is recycled to carry respectively
Take out the preset kind acoustic feature of each speech frame in each speech frame group；
S5, the corresponding preset kind acoustic feature of each speech frame group in the training set is divided into M crowd, it is in batches defeated
Enter and be iterated training in the preset structure deep neural network model, and in the preset structure deep neural network model
After the completion of training, verified using the accuracy rate of preset structure deep neural network model described in verification set pair；
If the accuracy rate that S6, verification obtain is more than predetermined threshold value, model training terminates；
If the accuracy rate that S7, verification obtain is less than or equal to predetermined threshold value, increase the voice data sample of acquisition
Quantity, and above-mentioned steps S1-S5 is re-executed based on the voice data sample after increase.
Preferably, the process of the preset structure deep neural network model repetitive exercise includes：
The corresponding preset kind acoustic feature of each speech frame group will be currently inputted according to the parameter current of model to be converted into
The characteristic vector of a corresponding preset length；
It is randomly selected from each characteristic vector to obtain multiple triples, i-th of triple (xi1,xi2,xi3) by
Three different characteristic vector xi1、xi2And xi3Composition, wherein xi1And xi2The corresponding same speaker, xi1And xi3It is corresponding different
Speaker, i is positive integer；
X is calculated using predetermined calculation formulai1And xi2Between cosine similarityAnd calculate xi1And xi3Between
Cosine similarity
According to cosine similarityAnd the parameter of predetermined loss function L more new models, it is described true in advance
The formula of fixed loss function L is：Wherein α is that value range is normal between 0.05~0.2
Amount, N are the numbers of the triple obtained.
Preferably, the network structure of the preset structure deep neural network model is as follows：
First layer：Be several layer heaps it is folded have mutually isostructural neural net layer, wherein every layer of neural network is using side by side
An one forward direction shot and long term memory network LSTM and backward LSTM, the number of plies are 1~3 layer；
The second layer：It is average layer, the effect of this layer was axially averaged to vector sequence along the time, it will be before last layer
The vector sequence exported to LSTM and backward LSTM all averages, and it is backward average to obtain a Forward averaging vector sum one
Vector, and by the two mean vector tandems at a vector；
Third layer：It is the full articulamentums of deep neural network DNN；
4th layer：It is normalization layer, the input of last layer is normalized this layer according to L2 norms, and it is 1 to obtain length
Normalization after characteristic vector；
Layer 5：It is loss layer, the formula of loss function L is：Wherein α is value model
The constant being trapped among between 0.05~0.2,The cosine similarity for belonging to two characteristic vectors of same speaker is represented,Generation
Table is not belonging to the cosine similarity of two characteristic vectors of same speaker.
The present invention also proposes a kind of auth method, which includes：
After receiving the current speech data of target user of pending authentication, obtained from database to be verified
The corresponding standard voice data of identity, by the current speech data and standard voice data respectively according to preset framing parameter
Sub-frame processing is carried out, it is corresponding to obtain the corresponding current speech frame group of the current speech data and the standard voice data
Received pronunciation frame group；
Extracted respectively using Predetermined filter in current speech frame group the preset kind acoustic feature of each speech frame and
The preset kind acoustic feature of each speech frame in received pronunciation frame group；
It is respectively that the corresponding preset kind acoustic feature of current speech frame group and received pronunciation frame group that extract is corresponding
The advance trained preset structure deep neural network model of preset kind acoustic feature input, to obtain the current speech number
According to the characteristic vector with the corresponding preset length of the standard voice data；
The cosine similarity for two characteristic vectors being calculated, and body is determined according to calculated cosine similarity size
Part verification result, the authentication result include being verified result and authentication failed result.
Preferably, divided respectively according to preset framing parameter by the current speech data and standard voice data
Before the step of frame processing, the auth method further includes step：
Active endpoint detection is carried out to the current speech data and standard voice data respectively, by the current speech number
According to the voice deletion with the non-talking people in the standard voice data.
Preferably, the training process of the preset structure deep neural network model is：
S1, preset quantity voice data sample is obtained, each voice data sample is marked respectively and represents corresponding speak
The label of personal part；
S2, active endpoint detection is carried out to each voice data sample respectively, by non-talking people in voice data sample
Voice is deleted, and the standard voice data sample of preset quantity is obtained；
S3, using the first percentage of obtained standard voice data sample as training set, the second percentage is as verification
Collection, first percentage and the second percentage and less than or equal to 100%；
S4, each standard voice data sample that the training set and verification are concentrated is distinguished according to preset framing parameter
Sub-frame processing is carried out, to obtain the corresponding speech frame group of each standard voice data sample, Predetermined filter is recycled to carry respectively
Take out the preset kind acoustic feature of each speech frame in each speech frame group；
S5, the corresponding preset kind acoustic feature of each speech frame group in the training set is divided into M crowd, it is in batches defeated
Enter and be iterated training in the preset structure deep neural network model, and in the preset structure deep neural network model
After the completion of training, verified using the accuracy rate of preset structure deep neural network model described in verification set pair；
If the accuracy rate that S6, verification obtain is more than predetermined threshold value, model training terminates；
If the accuracy rate that S7, verification obtain is less than or equal to predetermined threshold value, increase the voice data sample of acquisition
Quantity, and above-mentioned steps S1-S5 is re-executed based on the voice data sample after increase.
Preferably, the network structure of the preset structure deep neural network model is as follows：
First layer：Be several layer heaps it is folded have mutually isostructural neural net layer, wherein every layer of neural network is using side by side
An one forward direction shot and long term memory network LSTM and backward LSTM, the number of plies are 1~3 layer；
The second layer：It is average layer, the effect of this layer was axially averaged to vector sequence along the time, it will be before last layer
The vector sequence exported to LSTM and backward LSTM all averages, and it is backward average to obtain a Forward averaging vector sum one
Vector, and by the two mean vector tandems at a vector；
Third layer：It is the full articulamentums of deep neural network DNN；
4th layer：It is normalization layer, the input of last layer is normalized this layer according to L2 norms, and it is 1 to obtain length
Normalization after characteristic vector；
Layer 5：It is loss layer, the formula of loss function L is：Wherein α is value model
The constant being trapped among between 0.05~0.2,The cosine similarity for belonging to two characteristic vectors of same speaker is represented,Generation
Table is not belonging to the cosine similarity of two characteristic vectors of same speaker.
The present invention also proposes that a kind of computer readable storage medium, the computer-readable recording medium storage have identity to test
Card system, the authentication system can be executed by least one processor, so that at least one processor execution is as follows
Step：
After receiving the current speech data of target user of pending authentication, obtained from database to be verified
The corresponding standard voice data of identity, by the current speech data and standard voice data respectively according to preset framing parameter
Sub-frame processing is carried out, it is corresponding to obtain the corresponding current speech frame group of the current speech data and the standard voice data
Received pronunciation frame group；
Extracted respectively using Predetermined filter in current speech frame group the preset kind acoustic feature of each speech frame and
The preset kind acoustic feature of each speech frame in received pronunciation frame group；
It is respectively that the corresponding preset kind acoustic feature of current speech frame group and received pronunciation frame group that extract is corresponding
The advance trained preset structure deep neural network model of preset kind acoustic feature input, to obtain the current speech number
According to the characteristic vector with the corresponding preset length of the standard voice data, the cosine for two characteristic vectors being calculated
Similarity, and authentication is determined as a result, the authentication result includes verification according to calculated cosine similarity size
Pass through result and authentication failed result.
Technical solution of the present invention will be by that will receive the current speech data of the target user of identity to be verified and to be verified
The standard voice data of identity first carries out sub-frame processing, and each speech frame that sub-frame processing obtains is extracted using Predetermined filter
Preset kind acoustic feature is extracted, then the preset kind acoustic feature extracted is input to advance trained preset structure
Deep neural network model, preset structure deep neural network model is respectively by the corresponding preset kind acoustics of current speech data
After feature and the corresponding preset kind acoustic feature of standard voice data are converted into corresponding feature vector, calculate two features to
The cosine similarity of amount confirms verification result according to cosine similarity size.The present embodiment technical solution, by by voice data
First sub-frame processing is multiple speech frames and extracts preset kind acoustic feature according to speech frame so that even if collected effective
Voice data very in short-term, can also extract and be extracted to obtain enough acoustic features according to collected voice data, then pass through instruction
The deep neural network model perfected is handled according to extracting to obtain acoustic feature, to export verification result, compared to existing
For having technology, accuracy and reliability higher that this programme verifies speaker's identity.
Description of the drawings
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
The structure shown according to these attached drawings obtains other attached drawings.
Fig. 1 is the flow diagram of one embodiment of auth method of the present invention；
Fig. 2 is the flow diagram of preset structure deep neural network model training process of the present invention；
Fig. 3 is the running environment schematic diagram of one embodiment of authentication system of the present invention；
Fig. 4 is the Program modual graph of one embodiment of authentication system of the present invention；
Fig. 5 is the Program modual graph of two embodiment of authentication system of the present invention.
The embodiments will be further described with reference to the accompanying drawings for the realization, the function and the advantages of the object of the present invention.
Specific implementation mode
The principle and features of the present invention will be described below with reference to the accompanying drawings, and the given examples are served only to explain the present invention, and
It is non-to be used to limit the scope of the present invention.
As shown in FIG. 1, FIG. 1 is the flow diagrams of one embodiment of auth method of the present invention.
In the present embodiment, which includes：
Step S10 is obtained after receiving the current speech data of target user of pending authentication from database
The corresponding standard voice data of identity to be verified, by the current speech data and standard voice data respectively according to preset
Framing parameter carries out sub-frame processing, to obtain the corresponding current speech frame group of the current speech data and the received pronunciation number
According to corresponding received pronunciation frame group；
It is previously stored with the standard voice data of each identity in the database of authentication system, is receiving pending body
After the current speech data of the target user of part verification, the identity (identity to be verified) of verification, body are required according to target user
Part verification system obtains the corresponding standard voice data of identity to be verified in the database, then again respectively to receiving
Current speech data and the standard voice data got carry out sub-frame processing according to preset framing parameter, to obtain described work as
The corresponding current speech frame group of preceding voice data (including current speech data obtained through framing multiple speech frames) and the mark
The corresponding received pronunciation frame group of quasi- voice data (including standard voice data obtained through framing multiple speech frames).Wherein, institute
Preset framing parameter is stated for example, every 25 milliseconds of framings, frame moves 10 milliseconds.
Step S20 extracts the preset kind sound of each speech frame in current speech frame group using Predetermined filter respectively
Learn the preset kind acoustic feature of each speech frame in feature and received pronunciation frame group；
After obtaining current speech frame group and received pronunciation frame group, authentication system is right respectively using Predetermined filter
Each speech frame in current speech frame group and received pronunciation frame group carries out feature extraction, to extract in current speech frame group
The corresponding preset kind sound of each speech frame in the corresponding preset kind acoustic feature of each speech frame and received pronunciation frame group
Learn feature.For example, the Predetermined filter is Meier (Mel) filter, the preset kind acoustic feature extracted is 36 dimension MFCC
(Mel Frequency Cepstrum Coefficient, mel-frequency cepstrum coefficient) spectrum signature.
Step S30, the corresponding preset kind acoustic feature of current speech frame group and received pronunciation frame that will be extracted respectively
The corresponding advance trained preset structure deep neural network model of preset kind acoustic feature input of group, to obtain described work as
The characteristic vector of preceding voice data and the corresponding preset length of the standard voice data；
Step S40, the cosine similarity for two characteristic vectors being calculated, and it is big according to calculated cosine similarity
Small determining authentication is as a result, the authentication result includes being verified result and authentication failed result.
It is using sample to have advance trained preset structure deep neural network model, the model in authentication system
The model of the corresponding preset kind acoustic feature repetitive exercise of this voice data；Authentication system is to current speech frame group
After carrying out feature extraction with the speech frame in received pronunciation frame group, by the corresponding preset kind acoustic feature of current speech frame group and
The corresponding preset kind acoustic feature of received pronunciation frame group inputs the trained preset structure deep neural network model in advance
In, model is special by the corresponding preset kind acoustic feature of current speech frame group and the corresponding preset kind acoustics of received pronunciation frame group
Sign is separately converted to the characteristic vector (for example, characteristic vector that length is 1) of a preset length, then two spies being calculated
The cosine similarity for levying vector, determines authentication as a result, will the cosine phase according to the size of calculated cosine similarity
Like degree compared with predetermined threshold value (such as 0.95), if the cosine similarity is more than predetermined threshold value, it is determined that authentication passes through, instead
It, it is determined that authentication fails.Wherein, cosine similarity calculation formula is：cos(xi,xj)=xi Txj, xiAnd xjRepresent two
A characteristic vector, T are to predefine value.
The present embodiment technical solution, by the way that the current speech data of the target user of identity to be verified and to be tested will be received
The standard voice data of card identity first carries out sub-frame processing, each speech frame obtained using Predetermined filter extraction sub-frame processing
Extract preset kind acoustic feature, then the preset kind acoustic feature extracted is input to advance trained default knot
Structure deep neural network model, preset structure deep neural network model is respectively by the corresponding preset kind sound of current speech data
After feature and the corresponding preset kind acoustic feature of standard voice data are converted into corresponding feature vector, two features are calculated
The cosine similarity of vector confirms verification result according to cosine similarity size.The present embodiment technical solution, by by voice number
It is multiple speech frames according to first sub-frame processing and preset kind acoustic feature is extracted according to speech frame so that even if has collected
It imitates voice data very in short-term, can also extract and be extracted to obtain enough acoustic features according to collected voice data, then pass through
Trained deep neural network model is handled according to extracting to obtain acoustic feature, to export verification result, compared to
For the prior art, accuracy and reliability higher that this programme verifies speaker's identity.
Further, the present embodiment by the current speech data and standard voice data respectively according to preset framing
Before parameter carries out the step of sub-frame processing, the auth method further includes step：
Active endpoint detection is carried out to the current speech data and standard voice data respectively, by the current speech number
According to the voice deletion with the non-talking people in the standard voice data.
All include some non-talking human speech sounds in the current speech data and pre-stored standard voice data of acquisition
Partly (for example, mute or noise), if these parts are not deleted, to current speech data or standard to voice data into
In the speech frame group obtained after row sub-frame processing, it may appear that speech frame (or even the individual voice comprising non-talking people's phonological component
Non-talking human speech sound is all in frame), in this way, including the speech frame of non-talking people's phonological component according to these using Predetermined filter
The preset kind acoustic feature extracted belongs to impurity characteristics, can reduce preset structure deep neural network model and obtain a result
Accuracy；Therefore the present embodiment first detects in current speech data and standard voice data before to voice data sub-frame processing
Non-talking people's phonological component, and by the non-talking people's phonological component detected delete, the present embodiment use non-talking human speech
The detection mode of line point is that active endpoint detects (Voice Activity Detection, VAD).
As shown in Fig. 2, in the present embodiment, the training process of the preset structure deep neural network model is：
S1, preset quantity voice data sample is obtained, each voice data sample is marked respectively and represents corresponding speak
The label of personal part；
Preset quantity (for example, 10000) voice data sample is first got out, each voice data sample is all known theory
The voice data of the personal part of words；In these voice data samples, each speaker's identity or partial speaker's identity correspond to
There are multiple voice data samples, the label of corresponding speaker's identity will be represented on each voice data sample mark.
S2, active endpoint detection is carried out to each voice data sample respectively, by non-talking people in voice data sample
Voice is deleted, and the standard voice data sample of preset quantity is obtained；
Active endpoint detection is carried out to voice data sample, to detect the non-talking people's in each voice data sample
Voice (for example, mute or noise) is simultaneously deleted, and avoids existing in voice data sample special with the vocal print of corresponding speaker's identity
Unrelated voice data is levied, and influences the training effect to model.
S3, using the first percentage of obtained standard voice data sample as training set, the second percentage is as verification
Collection, first percentage and the second percentage and less than or equal to 100%；
For example, being used as training set by the 70% of obtained standard voice data sample, 30% as verification collection.
S4, each standard voice data sample that the training set and verification are concentrated is distinguished according to preset framing parameter
Sub-frame processing is carried out, to obtain the corresponding speech frame group of each standard voice data sample, Predetermined filter is recycled to carry respectively
Take out the preset kind acoustic feature of each speech frame in each speech frame group；
Wherein, for preset framing parameter for example, every 25 milliseconds of framings, frame moves 10 milliseconds；The Predetermined filter is, for example,
Meier filter is MFCC (Mel Frequency by the preset kind acoustic feature that Meier filter extracts
Cepstrum Coefficient, mel-frequency cepstrum coefficient) spectrum signature, for example, 36 dimension MFCC spectrum signatures.
S5, the corresponding preset kind acoustic feature of each speech frame group in the training set is divided into M crowd, it is in batches defeated
Enter and be iterated training in the preset structure deep neural network model, and in the preset structure deep neural network model
After the completion of training, verified using the accuracy rate of preset structure deep neural network model described in verification set pair；
Batch processing is carried out to the preset kind acoustic feature in training set, is divided into M (such as 30) batches, batch mode can
It is allocation unit according to speech frame group, the corresponding preset kind acoustics of the speech frame group of distribution equivalent or inequality is special in every a batch
Sign；The corresponding preset kind acoustic feature of each speech frame group in training set is preset according to the input of the batch being divided into one by one
It is iterated training in constructional depth neural network model, the preset structure victory is made to read god per a batch preset kind acoustic feature
Primary through network model iteration, each iteration can all update to obtain new model parameter, should after the completion of being trained by successive ignition
Preset structure deep neural network model has been updated to preferable model parameter；After the completion of repetitive exercise, then verification collection is utilized
The accuracy rate of the preset structure deep neural network model is verified, i.e., is divided the standard voice data that verification is concentrated two-by-two
Group, each corresponding preset kind acoustic feature of standard voice data sample inputted in a grouping to the preset structure depth
Neural network model confirms whether the verification structure of output is correct according to the identity label of the two of input standard voice datas,
After completing to the verification of each grouping, accuracy rate is calculated according to the correct number of verification result, such as test 100 groupings
Card, finally obtaining verification result correctly has 99 groups, then accuracy rate is just 99%.
If the accuracy rate that S6, verification obtain is more than predetermined threshold value, model training terminates；
Verification threshold (the i.e. described predetermined threshold value, such as 98.5%), for institute of accuracy rate is pre-set in system
The training effect for stating preset structure deep neural network model is tested；If deep by preset structure described in the verification set pair
The accuracy rate that degree neural network model is verified is more than the predetermined threshold value, then illustrating the preset structure deep neural network
The training of model has reached standard, then terminates model training at this time.
If the accuracy rate that S7, verification obtain is less than or equal to predetermined threshold value, increase the voice data sample of acquisition
Quantity, and above-mentioned steps S1-S5 is re-executed based on the voice data sample after increase.
If being less than by the accuracy rate that preset structure deep neural network model is verified described in the verification set pair
Or it is equal to the predetermined threshold value, then illustrating that the training of the preset structure deep neural network model has reached expected mark not yet
It is accurate, it may be possible to training set quantity is not enough or verification collection quantity is inadequate, so, when this is the case, then increasing the voice number of acquisition
According to the quantity (for example, increase fixed quantity every time or increase random amount every time) of sample, then on the basis of this, re-execute
Above-mentioned steps S1-S5, so cycle are executed, until having reached the requirement of step S6, then terminate model training.
In the present embodiment, the process of the preset structure deep neural network model repetitive exercise includes：
The corresponding preset kind acoustic feature of each speech frame group will be currently inputted according to the parameter current of model to be converted into
The characteristic vector of a corresponding preset length；
It is randomly selected from each characteristic vector to obtain multiple triples, i-th of triple (xi1,xi2,xi3) by
Three different characteristic vector xi1、xi2And xi3Composition, wherein xi1And xi2The corresponding same speaker, xi1And xi3It is corresponding different
Speaker, i is positive integer；
X is calculated using predetermined calculation formulai1And xi2Between cosine similarityAnd calculate xi1And xi3Between
Cosine similarity
According to cosine similarityAnd the parameter of predetermined loss function L more new models, it is described true in advance
The formula of fixed loss function L is：Wherein α is that value range is normal between 0.05~0.2
Amount, N are the numbers of the triple obtained.
Wherein, model parameter update step is：1. calculating the preset structure deep neural network using back-propagation algorithm
Gradient；2. updating the preset structure deep neural network using mini-batch-SGD (small lot stochastic gradient descent) method
Parameter.
Further, the network structure of the preset structure deep neural network model of the present embodiment is as follows：
First layer：Be several layer heaps it is folded have mutually isostructural neural net layer, wherein every layer of neural network is using side by side
An one forward direction shot and long term memory network LSTM and backward LSTM, the number of plies are 1~3 layer；Forward direction LSTM and backward LSTM difference
Export a vector sequence；
The second layer：It is average layer, the effect of this layer was axially averaged to vector sequence along the time, it will be before last layer
The vector sequence exported to LSTM and backward LSTM all averages, and it is backward average to obtain a Forward averaging vector sum one
Vector, and by the two mean vector tandems at a vector；
Third layer：It is the full articulamentums of deep neural network DNN；
4th layer：It is normalization layer, the input of last layer is normalized this layer according to L2 norms, and it is 1 to obtain length
Normalization after characteristic vector；
Layer 5：It is loss layer, the formula of loss function L is：Wherein α is value
Constant of the range between 0.05~0.2,The cosine similarity for belonging to two characteristic vectors of same speaker is represented,
Represent the cosine similarity for being not belonging to two characteristic vectors of same speaker.
In addition, the present invention also proposes a kind of authentication system.
Referring to Fig. 3, being the running environment schematic diagram of 10 preferred embodiment of authentication system of the present invention.
In the present embodiment, authentication system 10 is installed and is run in electronic device 1.Electronic device 1 can be table
The computing devices such as laptop computer, notebook, palm PC and server.The electronic device 1 may include, but be not limited only to, and deposit
Reservoir 11, processor 12 and display 13.Fig. 3 illustrates only the electronic device 1 with component 11-13, it should be understood that
It is not required for implementing all components shown, the implementation that can be substituted is more or less component.
Memory 11 can be the internal storage unit of electronic device 1 in some embodiments, such as the electronic device 1
Hard disk or memory.Memory 11 can also be the External memory equipment of electronic device 1, such as electronics dress in further embodiments
Set the plug-in type hard disk being equipped on 1, intelligent memory card (Smart Media Card, SMC), secure digital (Secure
Digital, SD) card, flash card (Flash Card) etc..Further, memory 11 can also both include the interior of electronic device 1
Portion's storage unit also includes External memory equipment.Memory 11 is for storing the application software for being installed on electronic device 1 and all kinds of
Data, for example, authentication system 10 program code etc..Memory 11 can be also used for temporarily storing exported or
The data that will be exported.
Processor 12 can be in some embodiments a central processing unit (Central Processing Unit,
CPU), microprocessor or other data processing chips, the program code for being stored in run memory 11 or processing data, example
Such as execute authentication system 10.
Display 13 can be in some embodiments light-emitting diode display, liquid crystal display, touch-control liquid crystal display and
OLED (Organic Light-Emitting Diode, Organic Light Emitting Diode) touches device etc..Display 13 is for being shown in
The information that is handled in electronic device 1 and for showing visual user interface, such as business customizing interface etc..Electronic device
1 component 11-13 is in communication with each other by system bus.
Referring to Fig. 4, being the Program modual graph of 10 preferred embodiment of authentication system of the present invention.In the present embodiment,
Authentication system 10 can be divided into one or more modules, one or more module is stored in memory 11,
And it is performed by one or more processors (the present embodiment is processor 12), to complete the present invention.For example, in Fig. 4, identity
Verification system 10 can be divided into framing module 101, extraction module 102, computing module 103 and result determining module 104.This
It is the series of computation machine program instruction section for referring to complete specific function to invent so-called module, than program more suitable for description
The implementation procedure of authentication system 10 in the electronic apparatus 1, wherein：
Framing module 101, for after receiving the current speech data of target user of pending authentication, from data
The corresponding standard voice data of identity to be verified is obtained in library, and the current speech data and standard voice data are pressed respectively
Sub-frame processing is carried out according to preset framing parameter, to obtain the corresponding current speech frame group of the current speech data and the mark
The corresponding received pronunciation frame group of quasi- voice data；
It is previously stored with the standard voice data of each identity in the database of authentication system, is receiving pending body
After the current speech data of the target user of part verification, the identity (identity to be verified) of verification, body are required according to target user
Part verification system obtains the corresponding standard voice data of identity to be verified in the database, then again respectively to receiving
Current speech data and the standard voice data got carry out sub-frame processing according to preset framing parameter, to obtain described work as
The corresponding current speech frame group of preceding voice data (including current speech data obtained through framing multiple speech frames) and the mark
The corresponding received pronunciation frame group of quasi- voice data (including standard voice data obtained through framing multiple speech frames).Wherein, institute
Preset framing parameter is stated for example, every 25 milliseconds of framings, frame moves 10 milliseconds.
Extraction module 102, for extracting the pre- of each speech frame in current speech frame group respectively using Predetermined filter
If the preset kind acoustic feature of each speech frame in type acoustic feature and received pronunciation frame group；
After obtaining current speech frame group and received pronunciation frame group, authentication system is right respectively using Predetermined filter
Each speech frame in current speech frame group and received pronunciation frame group carries out feature extraction, to extract in current speech frame group
The corresponding preset kind sound of each speech frame in the corresponding preset kind acoustic feature of each speech frame and received pronunciation frame group
Learn feature.For example, the Predetermined filter is Meier (Mel) filter, the preset kind acoustic feature extracted is 36 dimension MFCC
(Mel Frequency Cepstrum Coefficient, mel-frequency cepstrum coefficient) spectrum signature.
Computing module 103, the corresponding preset kind acoustic feature of current speech frame group and mark for will extract respectively
The corresponding preset kind acoustic feature input of quasi- speech frame group trained preset structure deep neural network model in advance, with
To the characteristic vector of the current speech data and the corresponding preset length of the standard voice data；
As a result determining module 104, the cosine similarity of two characteristic vectors for being calculated, and according to calculated
Cosine similarity size determines authentication as a result, the authentication result includes being verified result and authentication failed knot
Fruit.
It is using sample to have advance trained preset structure deep neural network model, the model in authentication system
The model of the corresponding preset kind acoustic feature repetitive exercise of this voice data；Authentication system is to current speech frame group
After carrying out feature extraction with the speech frame in received pronunciation frame group, by the corresponding preset kind acoustic feature of current speech frame group and
The corresponding preset kind acoustic feature of received pronunciation frame group inputs the trained preset structure deep neural network model in advance
In, model is special by the corresponding preset kind acoustic feature of current speech frame group and the corresponding preset kind acoustics of received pronunciation frame group
Sign is separately converted to the characteristic vector (for example, characteristic vector that length is 1) of a preset length, then two spies being calculated
The cosine similarity for levying vector, determines authentication as a result, will the cosine phase according to the size of calculated cosine similarity
Like degree compared with predetermined threshold value (such as 0.95), if the cosine similarity is more than predetermined threshold value, it is determined that authentication passes through, instead
It, it is determined that authentication fails.Wherein, cosine similarity calculation formula is：cos(xi,xj)=xi Txj, xiAnd xjRepresent two
A characteristic vector, T are to predefine value.
The present embodiment technical solution, by the way that the current speech data of the target user of identity to be verified and to be tested will be received
The standard voice data of card identity first carries out sub-frame processing, each speech frame obtained using Predetermined filter extraction sub-frame processing
Extract preset kind acoustic feature, then the preset kind acoustic feature extracted is input to advance trained default knot
Structure deep neural network model, preset structure deep neural network model is respectively by the corresponding preset kind sound of current speech data
After feature and the corresponding preset kind acoustic feature of standard voice data are converted into corresponding feature vector, two features are calculated
The cosine similarity of vector confirms verification result according to cosine similarity size.The present embodiment technical solution, by by voice number
It is multiple speech frames according to first sub-frame processing and preset kind acoustic feature is extracted according to speech frame so that even if has collected
It imitates voice data very in short-term, can also extract and be extracted to obtain enough acoustic features according to collected voice data, then pass through
Trained deep neural network model is handled according to extracting to obtain acoustic feature, to export verification result, compared to
For the prior art, accuracy and reliability higher that this programme verifies speaker's identity.
As shown in figure 5, Fig. 5 is the Program modual graph of two embodiment of authentication system of the present invention.
In the present embodiment, authentication system further includes：
Detection module 105, for by current speech data and standard voice data respectively according to preset framing parameter
Before carrying out sub-frame processing, active endpoint detection is carried out to the current speech data and standard voice data respectively, it will be described
The voice of non-talking people in current speech data and the standard voice data is deleted.
All include some non-talking human speech sounds in the current speech data and pre-stored standard voice data of acquisition
Partly (for example, mute or noise), if these parts are not deleted, to current speech data or standard to voice data into
In the speech frame group obtained after row sub-frame processing, it may appear that speech frame (or even the individual voice comprising non-talking people's phonological component
Non-talking human speech sound is all in frame), in this way, including the speech frame of non-talking people's phonological component according to these using Predetermined filter
The preset kind acoustic feature extracted belongs to impurity characteristics, can reduce preset structure deep neural network model and obtain a result
Accuracy；Therefore the present embodiment first detects in current speech data and standard voice data before to voice data sub-frame processing
Non-talking people's phonological component, and by the non-talking people's phonological component detected delete, the present embodiment use non-talking human speech
The detection mode of line point is that active endpoint detects (Voice Activity Detection, VAD).
In the present embodiment, the training process of the preset structure deep neural network model is (can refer to Fig. 2)：
S1, preset quantity voice data sample is obtained, each voice data sample is marked respectively and represents corresponding speak
The label of personal part；
Preset quantity (for example, 10000) voice data sample is first got out, each voice data sample is all known theory
The voice data of the personal part of words；In these voice data samples, each speaker's identity or partial speaker's identity correspond to
There are multiple voice data samples, the label of corresponding speaker's identity will be represented on each voice data sample mark.
S2, active endpoint detection is carried out to each voice data sample respectively, by non-talking people in voice data sample
Voice is deleted, and the standard voice data sample of preset quantity is obtained；
Active endpoint detection is carried out to voice data sample, to detect the non-talking people's in each voice data sample
Voice (for example, mute or noise) is simultaneously deleted, and avoids existing in voice data sample special with the vocal print of corresponding speaker's identity
Unrelated voice data is levied, and influences the training effect to model.
S3, using the first percentage of obtained standard voice data sample as training set, the second percentage is as verification
Collection, first percentage and the second percentage and less than or equal to 100%；
For example, being used as training set by the 70% of obtained standard voice data sample, 30% as verification collection.
S4, each standard voice data sample that the training set and verification are concentrated is distinguished according to preset framing parameter
Sub-frame processing is carried out, to obtain the corresponding speech frame group of each standard voice data sample, Predetermined filter is recycled to carry respectively
Take out the preset kind acoustic feature of each speech frame in each speech frame group；
Wherein, for preset framing parameter for example, every 25 milliseconds of framings, frame moves 10 milliseconds；The Predetermined filter is, for example,
Meier filter is MFCC (Mel Frequency by the preset kind acoustic feature that Meier filter extracts
Cepstrum Coefficient, mel-frequency cepstrum coefficient) spectrum signature, for example, 36 dimension MFCC spectrum signatures.
S5, the corresponding preset kind acoustic feature of each speech frame group in the training set is divided into M crowd, it is in batches defeated
Enter and be iterated training in the preset structure deep neural network model, and in the preset structure deep neural network model
After the completion of training, verified using the accuracy rate of preset structure deep neural network model described in verification set pair；
Batch processing is carried out to the preset kind acoustic feature in training set, is divided into M (such as 30) batches, batch mode can
It is allocation unit according to speech frame group, the corresponding preset kind acoustics of the speech frame group of distribution equivalent or inequality is special in every a batch
Sign；The corresponding preset kind acoustic feature of each speech frame group in training set is preset according to the input of the batch being divided into one by one
It is iterated training in constructional depth neural network model, the preset structure victory is made to read god per a batch preset kind acoustic feature
Primary through network model iteration, each iteration can all update to obtain new model parameter, should after the completion of being trained by successive ignition
Preset structure deep neural network model has been updated to preferable model parameter；After the completion of repetitive exercise, then verification collection is utilized
The accuracy rate of the preset structure deep neural network model is verified, i.e., is divided the standard voice data that verification is concentrated two-by-two
Group, each corresponding preset kind acoustic feature of standard voice data sample inputted in a grouping to the preset structure depth
Neural network model confirms whether the verification structure of output is correct according to the identity label of the two of input standard voice datas,
After completing to the verification of each grouping, accuracy rate is calculated according to the correct number of verification result, such as test 100 groupings
Card, finally obtaining verification result correctly has 99 groups, then accuracy rate is just 99%.
If the accuracy rate that S6, verification obtain is more than predetermined threshold value, model training terminates；
Verification threshold (the i.e. described predetermined threshold value, such as 98.5%), for institute of accuracy rate is pre-set in system
The training effect for stating preset structure deep neural network model is tested；If deep by preset structure described in the verification set pair
The accuracy rate that degree neural network model is verified is more than the predetermined threshold value, then illustrating the preset structure deep neural network
The training of model has reached standard, then terminates model training at this time.
If the accuracy rate that S7, verification obtain is less than or equal to predetermined threshold value, increase the voice data sample of acquisition
Quantity, and above-mentioned steps S1-S5 is re-executed based on the voice data sample after increase.
If being less than by the accuracy rate that preset structure deep neural network model is verified described in the verification set pair
Or it is equal to the predetermined threshold value, then illustrating that the training of the preset structure deep neural network model has reached expected mark not yet
It is accurate, it may be possible to training set quantity is not enough or verification collection quantity is inadequate, so, when this is the case, then increasing the voice number of acquisition
According to the quantity (for example, increase fixed quantity every time or increase random amount every time) of sample, then on the basis of this, re-execute
Above-mentioned steps S1-S5, so cycle are executed, until having reached the requirement of step S6, then terminate model training.
In the present embodiment, the process of the preset structure deep neural network model repetitive exercise includes：
The corresponding preset kind acoustic feature of each speech frame group will be currently inputted according to the parameter current of model to be converted into
The characteristic vector of a corresponding preset length；
It is randomly selected from each characteristic vector to obtain multiple triples, i-th of triple (xi1,xi2,xi3) by
Three different characteristic vector xi1、xi2And xi3Composition, wherein xi1And xi2The corresponding same speaker, xi1And xi3It is corresponding different
Speaker, i is positive integer；
X is calculated using predetermined calculation formulai1And xi2Between cosine similarityAnd calculate xi1And xi3Between
Cosine similarity
According to cosine similarityAnd the parameter of predetermined loss function L more new models, it is described true in advance
The formula of fixed loss function L is：Wherein α is that value range is normal between 0.05~0.2
Amount, N are the numbers of the triple obtained.
Wherein, model parameter update step is：1. calculating the preset structure deep neural network using back-propagation algorithm
Gradient；2. updating the preset structure deep neural network using mini-batch-SGD (small lot stochastic gradient descent) method
Parameter.
Further, the network structure of the preset structure deep neural network model of the present embodiment is as follows：
First layer：Be several layer heaps it is folded have mutually isostructural neural net layer, wherein every layer of neural network is using side by side
An one forward direction shot and long term memory network LSTM and backward LSTM, the number of plies are 1~3 layer；Forward direction LSTM and backward LSTM difference
Export a vector sequence；
The second layer：It is average layer, the effect of this layer was axially averaged to vector sequence along the time, it will be before last layer
The vector sequence exported to LSTM and backward LSTM all averages, and it is backward average to obtain a Forward averaging vector sum one
Vector, and by the two mean vector tandems at a vector；
Third layer：It is the full articulamentums of deep neural network DNN；
4th layer：It is normalization layer, the input of last layer is normalized this layer according to L2 norms, and it is 1 to obtain length
Normalization after characteristic vector；
Layer 5：It is loss layer, the formula of loss function L is：Wherein α is value
Constant of the range between 0.05~0.2,The cosine similarity for belonging to two characteristic vectors of same speaker is represented,
Represent the cosine similarity for being not belonging to two characteristic vectors of same speaker.
Further, the present invention also proposes that a kind of computer readable storage medium, the computer readable storage medium are deposited
Authentication system is contained, the authentication system can be executed by least one processor, so that at least one processing
Device executes the auth method in any of the above-described embodiment.
The foregoing is merely the preferred embodiment of the present invention, are not intended to limit the scope of the invention, every at this
Under the inventive concept of invention, using equivalent structure transformation made by description of the invention and accompanying drawing content, or directly/use indirectly
In the scope of patent protection that other related technical areas are included in the present invention.
Claims (10)
1. a kind of electronic device, which is characterized in that the electronic device includes memory and processor, is stored on the memory
There is the authentication system that can be run on the processor, is realized such as when the authentication system is executed by the processor
Lower step：
After receiving the current speech data of target user of pending authentication, identity to be verified is obtained from database
Corresponding standard voice data carries out the current speech data and standard voice data according to preset framing parameter respectively
Sub-frame processing, to obtain the corresponding current speech frame group of the current speech data and the corresponding standard of the standard voice data
Speech frame group；
Extract the preset kind acoustic feature and standard of each speech frame in current speech frame group respectively using Predetermined filter
The preset kind acoustic feature of each speech frame in speech frame group；
It is respectively that the corresponding preset kind acoustic feature of current speech frame group and received pronunciation frame group that extract is corresponding default
Type acoustic feature input in advance trained preset structure deep neural network model, with obtain the current speech data and
The characteristic vector of the corresponding preset length of standard voice data；
The cosine similarity for two characteristic vectors being calculated, and determine that identity is tested according to calculated cosine similarity size
Card is as a result, the authentication result includes being verified result and authentication failed result.
2. electronic device as described in claim 1, which is characterized in that by the current speech data and standard voice data
Before the step of carrying out sub-frame processing according to preset framing parameter respectively, which is additionally operable to execute the authentication system
System, to realize following steps：
Active endpoint detection is carried out to the current speech data and standard voice data respectively, by the current speech data and
The voice of non-talking people in the standard voice data is deleted.
3. electronic device as described in claim 1, which is characterized in that the training of the preset structure deep neural network model
Process is：
S1, preset quantity voice data sample is obtained, each voice data sample is marked respectively and represents the corresponding person of speaking
The label of part；
S2, active endpoint detection is carried out to each voice data sample respectively, by the voice of non-talking people in voice data sample
It deletes, obtains the standard voice data sample of preset quantity；
S3, using the first percentage of obtained standard voice data sample as training set, the second percentage collects as verification, institute
State the first percentage and the second percentage and it is less than or equal to 100%；
S4, each standard voice data sample that the training set and verification are concentrated is carried out respectively according to preset framing parameter
Sub-frame processing recycles Predetermined filter to extract respectively to obtain the corresponding speech frame group of each standard voice data sample
The preset kind acoustic feature of each speech frame in each speech frame group；
S5, the corresponding preset kind acoustic feature of each speech frame group in the training set is divided into M batches, inputs institute in batches
It states and is iterated training in preset structure deep neural network model, and trained in the preset structure deep neural network model
After the completion, it is verified using the accuracy rate of preset structure deep neural network model described in verification set pair；
If the accuracy rate that S6, verification obtain is more than predetermined threshold value, model training terminates；
If the accuracy rate that S7, verification obtain is less than or equal to predetermined threshold value, increase the quantity of the voice data sample of acquisition,
And above-mentioned steps S1-S5 is re-executed based on the voice data sample after increase.
4. electronic device as claimed in claim 3, which is characterized in that the preset structure deep neural network model iteration instruction
Experienced process includes：
The corresponding preset kind acoustic feature of each speech frame group, which will be currently inputted, according to the parameter current of model is converted into correspondence
A preset length characteristic vector；
It is randomly selected from each characteristic vector to obtain multiple triples, i-th of triple (xi1,xi2,xi3) by three
Different characteristic vector xi1、xi2And xi3Composition, wherein xi1And xi2The corresponding same speaker, xi1And xi3Correspond to different say
People is talked about, i is positive integer；
X is calculated using predetermined calculation formulai1And xi2Between cosine similarityAnd calculate xi1And xi3Between it is remaining
String similarity
According to cosine similarityAnd the parameter of predetermined loss function L more new models, the predetermined damage
The formula for losing function L is：Wherein α is value range constant between 0.05~0.2, and N is
The number of the triple of acquisition.
5. the electronic device as described in any one of claim 1-4, which is characterized in that the preset structure depth nerve net
The network structure of network model is as follows：
First layer：Be several layer heaps it is folded have mutually isostructural neural net layer, wherein every layer of neural network is using one side by side
A forward direction shot and long term memory network LSTM and backward LSTM, the number of plies are 1~3 layer；
The second layer：Average layer, the effect of this layer was axially averaged to vector sequence along the time, it by before last layer to
The vector sequence of LSTM and backward LSTM outputs all averages, and obtains the backward average arrow of a Forward averaging vector sum one
Amount, and by the two mean vector tandems at a vector；
Third layer：It is the full articulamentums of deep neural network DNN；
4th layer：It is normalization layer, the input of last layer is normalized this layer according to L2 norms, and it is 1 to return to obtain length
Characteristic vector after one change；
Layer 5：It is loss layer, the formula of loss function L is：Wherein α is that value range exists
Constant between 0.05~0.2,The cosine similarity for belonging to two characteristic vectors of same speaker is represented,It represents not
Belong to the cosine similarity of two characteristic vectors of same speaker.
6. a kind of auth method, which is characterized in that the auth method includes：
After receiving the current speech data of target user of pending authentication, identity to be verified is obtained from database
Corresponding standard voice data carries out the current speech data and standard voice data according to preset framing parameter respectively
Sub-frame processing, to obtain the corresponding current speech frame group of the current speech data and the corresponding standard of the standard voice data
Speech frame group；
Extract the preset kind acoustic feature and standard of each speech frame in current speech frame group respectively using Predetermined filter
The preset kind acoustic feature of each speech frame in speech frame group；
It is respectively that the corresponding preset kind acoustic feature of current speech frame group and received pronunciation frame group that extract is corresponding default
Type acoustic feature input in advance trained preset structure deep neural network model, with obtain the current speech data and
The characteristic vector of the corresponding preset length of standard voice data；
The cosine similarity for two characteristic vectors being calculated, and determine that identity is tested according to calculated cosine similarity size
Card is as a result, the authentication result includes being verified result and authentication failed result.
7. auth method as claimed in claim 6, which is characterized in that by the current speech data and received pronunciation
Before the step of data carry out sub-frame processing according to preset framing parameter respectively, the auth method further includes step：
Active endpoint detection is carried out to the current speech data and standard voice data respectively, by the current speech data and
The voice of non-talking people in the standard voice data is deleted.
8. auth method as claimed in claim 6, which is characterized in that the preset structure deep neural network model
Training process is：
S1, preset quantity voice data sample is obtained, each voice data sample is marked respectively and represents the corresponding person of speaking
The label of part；
S2, active endpoint detection is carried out to each voice data sample respectively, by the voice of non-talking people in voice data sample
It deletes, obtains the standard voice data sample of preset quantity；
S3, using the first percentage of obtained standard voice data sample as training set, the second percentage collects as verification, institute
State the first percentage and the second percentage and it is less than or equal to 100%；
S4, each standard voice data sample that the training set and verification are concentrated is carried out respectively according to preset framing parameter
Sub-frame processing recycles Predetermined filter to extract respectively to obtain the corresponding speech frame group of each standard voice data sample
The preset kind acoustic feature of each speech frame in each speech frame group；
S5, the corresponding preset kind acoustic feature of each speech frame group in the training set is divided into M batches, inputs institute in batches
It states and is iterated training in preset structure deep neural network model, and trained in the preset structure deep neural network model
After the completion, it is verified using the accuracy rate of preset structure deep neural network model described in verification set pair；
If the accuracy rate that S6, verification obtain is more than predetermined threshold value, model training terminates；
If the accuracy rate that S7, verification obtain is less than or equal to predetermined threshold value, increase the quantity of the voice data sample of acquisition,
And above-mentioned steps S1-S5 is re-executed based on the voice data sample after increase.
9. the auth method as described in any one of claim 6 to 8, which is characterized in that the preset structure depth
The network structure of neural network model is as follows：
First layer：Be several layer heaps it is folded have mutually isostructural neural net layer, wherein every layer of neural network is using one side by side
A forward direction shot and long term memory network LSTM and backward LSTM, the number of plies are 1~3 layer；
The second layer：Average layer, the effect of this layer was axially averaged to vector sequence along the time, it by before last layer to
The vector sequence of LSTM and backward LSTM outputs all averages, and obtains the backward average arrow of a Forward averaging vector sum one
Amount, and by the two mean vector tandems at a vector；
Third layer：It is the full articulamentums of deep neural network DNN；
4th layer：It is normalization layer, the input of last layer is normalized this layer according to L2 norms, and it is 1 to return to obtain length
Characteristic vector after one change；
Layer 5：It is loss layer, the formula of loss function L is：Wherein α is that value range exists
Constant between 0.05~0.2,The cosine similarity for belonging to two characteristic vectors of same speaker is represented,It represents not
Belong to the cosine similarity of two characteristic vectors of same speaker.
10. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has authentication
System, the authentication system can be executed by least one processor, so that at least one processor executes such as right
It is required that the auth method described in any one of 6-9.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810225887.2A CN108564955B (en) | 2018-03-19 | 2018-03-19 | Electronic device, auth method and computer readable storage medium |
| PCT/CN2018/102105 WO2019179029A1 (en) | 2018-03-19 | 2018-08-24 | Electronic device, identity verification method and computer-readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810225887.2A CN108564955B (en) | 2018-03-19 | 2018-03-19 | Electronic device, auth method and computer readable storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108564955A true CN108564955A (en) | 2018-09-21 |
| CN108564955B CN108564955B (en) | 2019-09-03 |
Family
ID=63532742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810225887.2A Active CN108564955B (en) | 2018-03-19 | 2018-03-19 | Electronic device, auth method and computer readable storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN108564955B (en) |
| WO (1) | WO2019179029A1 (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108564954A (en) * | 2018-03-19 | 2018-09-21 | 平安科技（深圳）有限公司 | Deep neural network model, electronic device, auth method and storage medium |
| CN109346086A (en) * | 2018-10-26 | 2019-02-15 | 平安科技（深圳）有限公司 | Method for recognizing sound-groove, device, computer equipment and computer readable storage medium |
| CN110148402A (en) * | 2019-05-07 | 2019-08-20 | 平安科技（深圳）有限公司 | Method of speech processing, device, computer equipment and storage medium |
| CN110289003A (en) * | 2018-10-10 | 2019-09-27 | 腾讯科技（深圳）有限公司 | A kind of method of Application on Voiceprint Recognition, the method for model training and server |
| CN110570871A (en) * | 2019-09-20 | 2019-12-13 | 平安科技（深圳）有限公司 | TristouNet-based voiceprint recognition method, device and equipment |
| CN111241506A (en) * | 2018-11-28 | 2020-06-05 | Sap欧洲公司 | Progressive authentication security adapter |
| CN111933153A (en) * | 2020-07-07 | 2020-11-13 | 北京捷通华声科技股份有限公司 | Method and device for determining voice segmentation points |
| CN112016673A (en) * | 2020-07-24 | 2020-12-01 | 浙江工业大学 | Mobile equipment user authentication method and device based on optimized LSTM |
| CN112309365A (en) * | 2020-10-21 | 2021-02-02 | 北京大米科技有限公司 | Training method and device of speech synthesis model, storage medium and electronic equipment |
| CN112347788A (en) * | 2020-11-06 | 2021-02-09 | 平安消费金融有限公司 | Corpus processing method, apparatus and storage medium |
| WO2022227223A1 (en) * | 2021-04-27 | 2022-11-03 | 平安科技（深圳）有限公司 | Voice verification model training method and apparatus, and computer device |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112712792A (en) * | 2019-10-25 | 2021-04-27 | Tcl集团股份有限公司 | Dialect recognition model training method, readable storage medium and terminal device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060025995A1 (en) * | 2004-07-29 | 2006-02-02 | Erhart George W | Method and apparatus for natural language call routing using confidence scores |
| CN105139857A (en) * | 2015-09-02 | 2015-12-09 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Countercheck method for automatically identifying speaker aiming to voice deception |
| CN106205624A (en) * | 2016-07-15 | 2016-12-07 | 河海大学 | A kind of method for recognizing sound-groove based on DBSCAN algorithm |
| CN106782564A (en) * | 2016-11-18 | 2017-05-31 | 百度在线网络技术（北京）有限公司 | Method and apparatus for processing speech data |
| CN107610707A (en) * | 2016-12-15 | 2018-01-19 | 平安科技（深圳）有限公司 | A kind of method for recognizing sound-groove and device |
-
2018
- 2018-03-19 CN CN201810225887.2A patent/CN108564955B/en active Active
- 2018-08-24 WO PCT/CN2018/102105 patent/WO2019179029A1/en active Application Filing
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060025995A1 (en) * | 2004-07-29 | 2006-02-02 | Erhart George W | Method and apparatus for natural language call routing using confidence scores |
| CN105139857A (en) * | 2015-09-02 | 2015-12-09 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Countercheck method for automatically identifying speaker aiming to voice deception |
| CN106205624A (en) * | 2016-07-15 | 2016-12-07 | 河海大学 | A kind of method for recognizing sound-groove based on DBSCAN algorithm |
| CN106782564A (en) * | 2016-11-18 | 2017-05-31 | 百度在线网络技术（北京）有限公司 | Method and apparatus for processing speech data |
| CN107610707A (en) * | 2016-12-15 | 2018-01-19 | 平安科技（深圳）有限公司 | A kind of method for recognizing sound-groove and device |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108564954B (en) * | 2018-03-19 | 2020-01-10 | 平安科技（深圳）有限公司 | Deep neural network model, electronic device, identity verification method, and storage medium |
| CN108564954A (en) * | 2018-03-19 | 2018-09-21 | 平安科技（深圳）有限公司 | Deep neural network model, electronic device, auth method and storage medium |
| CN110289003B (en) * | 2018-10-10 | 2021-10-29 | 腾讯科技（深圳）有限公司 | Voiceprint recognition method, model training method and server |
| CN110289003A (en) * | 2018-10-10 | 2019-09-27 | 腾讯科技（深圳）有限公司 | A kind of method of Application on Voiceprint Recognition, the method for model training and server |
| CN109346086A (en) * | 2018-10-26 | 2019-02-15 | 平安科技（深圳）有限公司 | Method for recognizing sound-groove, device, computer equipment and computer readable storage medium |
| CN111241506A (en) * | 2018-11-28 | 2020-06-05 | Sap欧洲公司 | Progressive authentication security adapter |
| CN111241506B (en) * | 2018-11-28 | 2023-09-08 | Sap欧洲公司 | Progressive authentication security adapter |
| CN110148402A (en) * | 2019-05-07 | 2019-08-20 | 平安科技（深圳）有限公司 | Method of speech processing, device, computer equipment and storage medium |
| CN110570871A (en) * | 2019-09-20 | 2019-12-13 | 平安科技（深圳）有限公司 | TristouNet-based voiceprint recognition method, device and equipment |
| CN111933153A (en) * | 2020-07-07 | 2020-11-13 | 北京捷通华声科技股份有限公司 | Method and device for determining voice segmentation points |
| CN111933153B (en) * | 2020-07-07 | 2024-03-08 | 北京捷通华声科技股份有限公司 | Voice segmentation point determining method and device |
| CN112016673A (en) * | 2020-07-24 | 2020-12-01 | 浙江工业大学 | Mobile equipment user authentication method and device based on optimized LSTM |
| CN112309365A (en) * | 2020-10-21 | 2021-02-02 | 北京大米科技有限公司 | Training method and device of speech synthesis model, storage medium and electronic equipment |
| CN112347788A (en) * | 2020-11-06 | 2021-02-09 | 平安消费金融有限公司 | Corpus processing method, apparatus and storage medium |
| WO2022227223A1 (en) * | 2021-04-27 | 2022-11-03 | 平安科技（深圳）有限公司 | Voice verification model training method and apparatus, and computer device |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2019179029A1 (en) | 2019-09-26 |
| CN108564955B (en) | 2019-09-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108564955B (en) | Electronic device, auth method and computer readable storage medium | |
| CN108564954A (en) | Deep neural network model, electronic device, auth method and storage medium | |
| TWI641965B (en) | Method and system of authentication based on voiceprint recognition | |
| CN109817246B (en) | Emotion recognition model training method, emotion recognition device, emotion recognition equipment and storage medium | |
| CN107610709B (en) | Method and system for training voiceprint recognition model | |
| CN107862292B (en) | Personage's mood analysis method, device and storage medium | |
| CN107610707A (en) | A kind of method for recognizing sound-groove and device | |
| CN107564511A (en) | Electronic installation, phoneme synthesizing method and computer-readable recording medium | |
| KR102401194B1 (en) | Method and apparatus for authenticating a user using an electrocardiogram signal | |
| CN111898550B (en) | Expression recognition model building method and device, computer equipment and storage medium | |
| CN110473552A (en) | Speech recognition authentication method and system | |
| CN110211571A (en) | Wrong sentence detection method, device and computer readable storage medium | |
| CN110544468B (en) | Application awakening method and device, storage medium and electronic equipment | |
| CN106991312B (en) | Internet anti-fraud authentication method based on voiceprint recognition | |
| CN108345612A (en) | A kind of question processing method and device, a kind of device for issue handling | |
| CN108831484A (en) | A kind of offline and unrelated with category of language method for recognizing sound-groove and device | |
| CN112418059A (en) | Emotion recognition method and device, computer equipment and storage medium | |
| CN110111798A (en) | A kind of method and terminal identifying speaker | |
| CN113807103B (en) | Recruitment method, device, equipment and storage medium based on artificial intelligence | |
| CN114639152A (en) | Multi-modal voice interaction method, device, equipment and medium based on face recognition | |
| CN110364163A (en) | The identity identifying method that a kind of voice and lip reading blend | |
| CN109545226A (en) | A kind of audio recognition method, equipment and computer readable storage medium | |
| US6499012B1 (en) | Method and apparatus for hierarchical training of speech models for use in speaker verification | |
| CN1522431A (en) | Method and system for non-intrusive speaker verification using behavior model | |
| CN116151965A (en) | Risk feature extraction method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |