CN110428842A - Speech model training method, device, equipment and computer readable storage medium - Google Patents
Speech model training method, device, equipment and computer readable storage medium Download PDFInfo
- Publication number
- CN110428842A CN110428842A CN201910744145.5A CN201910744145A CN110428842A CN 110428842 A CN110428842 A CN 110428842A CN 201910744145 A CN201910744145 A CN 201910744145A CN 110428842 A CN110428842 A CN 110428842A
- Authority
- CN
- China
- Prior art keywords
- vocal print
- voice data
- print feature
- feature vector
- target
- 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.)
- Pending
Links
- 238000012549 training Methods 0.000 title claims abstract description 137
- 238000000034 method Methods 0.000 title claims abstract description 57
- 239000013598 vector Substances 0.000 claims abstract description 176
- 230000001755 vocal effect Effects 0.000 claims abstract description 176
- 238000003062 neural network model Methods 0.000 claims description 31
- 238000012545 processing Methods 0.000 claims description 23
- 230000008569 process Effects 0.000 claims description 15
- 238000009432 framing Methods 0.000 claims description 10
- 238000000605 extraction Methods 0.000 claims description 8
- 238000001514 detection method Methods 0.000 claims description 6
- 230000011218 segmentation Effects 0.000 claims description 4
- 210000004218 nerve net Anatomy 0.000 claims description 3
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 238000012512 characterization method Methods 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 13
- 238000001228 spectrum Methods 0.000 description 12
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 5
- 238000009826 distribution Methods 0.000 description 5
- 239000000284 extract Substances 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000007781 pre-processing Methods 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 210000002569 neuron Anatomy 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000011840 criminal investigation Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000001771 impaired effect Effects 0.000 description 1
- 238000012067 mathematical method Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 210000001260 vocal cord Anatomy 0.000 description 1
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification techniques
- 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
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification techniques
- G10L17/04—Training, enrolment or model building
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification techniques
- G10L17/18—Artificial neural networks; Connectionist approaches
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification techniques
- G10L17/22—Interactive procedures; Man-machine interfaces
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Electrically Operated Instructional Devices (AREA)
Abstract
The invention discloses a kind of speech model training method, device, equipment and computer readable storage mediums, which comprises obtains the voice data of target speaker, and the voice data is divided into multiple voice data sections；Vocal print feature is extracted respectively from the multiple voice data section, obtains multiple vocal print feature vectors；The multiple vocal print feature vector is ranked up according to predetermined order mode, and chooses target vocal print feature vector based on ranking results；Training pattern is treated based on the target vocal print feature vector to be trained, and obtains the speech recognition modeling of the target speaker.The present invention realizes when the voice data of target speaker is not pure, the vocal print feature vector for capableing of accurate Characterization target speaker's vocal print feature can be extracted from unpurified voice data, and the speech recognition modeling of target speaker is obtained based on the training of accurate vocal print feature vector, to improve the accuracy of target speaker's speech recognition modeling.
Description
Technical field
The present invention relates to identity identification technical field more particularly to a kind of speech model training method, device, equipment and meters
Calculation machine readable storage medium storing program for executing.
Background technique
Speaker Identification is divided into two classes, i.e. speaker's identification and speaker verification.If the former is to judge certain section of voice
It is " multiselect one " problem described in which of dry people；And the latter is to confirm whether certain section of voice is specified someone
Described, it is " one-to-one differentiation " problem.Speaker Recognition Technology is in military affairs, national security, criminal investigation field and bank, security etc.
Financial field has a wide range of applications.
The voice data that Speaker Identification needs to acquire a large number of users is trained, and obtains the speech recognition mould of each user
Type is identified by way of pattern match when needing to carry out Speaker Identification.However, carrying out model training at present
During, due to the limitation of acquisition cost, collected user voice data is simultaneously lack of standardization, and user speech is often impure
, for example by the collected user voice data of calling record, often adulterate many noises, it is also possible to other people can be mixed into
Voice therefore directly adopt the training that such voice data carries out speech model, will lead to the speech model that training obtains
Recognition efficiency is low, accuracy is not high.
Summary of the invention
The main purpose of the present invention is to provide a kind of speech model training method, device, equipment and computer-readable deposit
Storage media, it is intended to solve at present due to causing speech model to identify using unpurified user voice data training speech model
The low technical problem of rate.
To achieve the above object, the present invention provides a kind of speech model training method, the speech model training method packet
It includes:
The voice data of target speaker is obtained, and the voice data is divided into multiple voice data sections；
Vocal print feature is extracted respectively from the multiple voice data section, obtains multiple vocal print feature vectors；
The multiple vocal print feature vector is ranked up according to predetermined order mode, and chooses target based on ranking results
Vocal print feature vector；
Training pattern is treated based on the target vocal print feature vector to be trained, and obtains the voice of the target speaker
Identification model.
Optionally, the voice data for obtaining target speaker, and the voice data is divided into multiple voice numbers
Include: according to the step of section
The voice data for obtaining target speaker, pre-processes the voice data；
The pretreated voice data is segmented according to prefixed time interval, obtains multiple voice data sections.
Optionally, described that the multiple vocal print feature vector is ranked up according to predetermined order mode, and based on sequence
As a result the step of selection target vocal print feature vector includes:
Seek the average vector of the multiple vocal print feature vector；
Calculate the distance between the multiple vocal print feature vector and the average vector, and by the multiple vocal print feature
Vector sorts according to the sequence of the average vector distance from big to small；
The vocal print feature vector for falling into preset range is chosen as target vocal print feature vector based on ranking results.
Optionally, it is described to training pattern be deep neural network model, it is described be based on the target vocal print feature vector
The step for the treatment of training pattern to be trained, obtaining the speech recognition modeling of the target speaker include:
By the target vocal print feature input initialization treated the deep neural network model, it is iterated instruction
Practice；
When detect the deep neural network model convergence after, using the deep neural network model after convergence as
The speech recognition modeling of the target speaker.
Optionally, described to extract vocal print feature respectively from the multiple voice data section, obtain multiple vocal print features to
The step of amount includes:
Based on preparatory trained universal background model self-adaptive processing is carried out to the multiple voice data section respectively, obtained
To multiple vocal print feature vectors.
Optionally, described the multiple voice data section to be carried out respectively certainly based on preparatory trained universal background model
Before the step of adaptation is handled, and obtains multiple vocal print feature vectors, further includes:
Trained voice data gathered in advance is pre-processed；
Training phonetic feature is extracted to the pretreated trained voice data；
Universal background model training is carried out by the trained phonetic feature, obtains universal background model；
Optionally, described to include: to the pretreated step of trained voice data progress gathered in advance
Preemphasis, framing, adding window and endpoint detection processing are successively carried out to trained voice data gathered in advance.
In addition, to achieve the above object, the present invention also provides a kind of speech model training device, the speech model training
Device includes:
Segmentation module is divided into multiple voices for obtaining the voice data of target speaker, and by the voice data
Data segment；
Extraction module obtains multiple vocal print features for extracting vocal print feature respectively from the multiple voice data section
Vector；
Order models, for being ranked up according to predetermined order mode to the multiple vocal print feature vector, and based on row
Sequence result chooses target vocal print feature vector；
Training module is trained for treating training pattern based on the target vocal print feature vector, obtains the mesh
Mark the speech recognition modeling of speaker.
In addition, to achieve the above object, the present invention also provides a kind of speech model training equipment, the speech model training
Equipment includes the speech model training that memory, processor and being stored in can be run on the memory and on the processor
Program, the speech model training program realize the step of speech model training method as described above when being executed by the processor
Suddenly.
In addition, to achieve the above object, it is described computer-readable the present invention also provides a kind of computer readable storage medium
Speech model training program is stored on storage medium, the speech model training program realizes institute as above when being executed by processor
The step of speech model training method stated.
In the present invention, by obtaining the voice data of target speaker, and the voice data of target speaker is divided into
Multiple voice data sections；Vocal print feature is extracted respectively from multiple voice data sections, obtains multiple vocal print feature vectors；According to pre-
If sortord is ranked up multiple vocal print feature vectors, and chooses target vocal print feature vector based on ranking results；It is based on
Target vocal print feature vector is treated training pattern and is trained, and the speech recognition modeling of target speaker is obtained, realize even if
When the voice data of target speaker is not pure, can be extracted from unpurified voice data can accurate Characterization target from the perspective of
The vocal print feature vector of people's vocal print feature is talked about, and is known based on the voice that the training of accurate vocal print feature vector obtains target speaker
Other model, to improve the accuracy of target speaker's speech recognition modeling.
Detailed description of the invention
Fig. 1 is the structural schematic diagram for the hardware running environment that the embodiment of the present invention is related to；
Fig. 2 is the flow diagram of speech model training method first embodiment of the present invention；
Fig. 3 is the functional schematic module map of speech model training device preferred embodiment 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 embodiment
It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not intended to limit the present invention.
The present invention provides a kind of speech model training equipment, and referring to Fig.1, Fig. 1 is that the embodiment of the present invention is related to
The structural schematic diagram of hardware running environment.
It should be noted that Fig. 1 can be the structural schematic diagram of the hardware running environment of speech model training equipment.This hair
Bright embodiment speech model training equipment can be PC, be also possible to smart phone, intelligent TV set, tablet computer, portable meter
The terminal device having a display function such as calculation machine.
As shown in Figure 1, speech model training equipment may include: processor 1001, such as CPU, network interface 1004,
User interface 1003, memory 1005, communication bus 1002.Wherein, communication bus 1002 is for realizing between these components
Connection communication.User interface 1003 may include display screen (Display), input unit such as keyboard (Keyboard), optional
User interface 1003 can also include standard wireline interface and wireless interface.Network interface 1004 optionally may include standard
Wireline interface, wireless interface (such as WI-FI interface).Memory 1005 can be high speed RAM memory, be also possible to stable
Memory (non-volatile memory), such as magnetic disk storage.Memory 1005 optionally can also be independently of aforementioned
The storage device of processor 1001.
Optionally, speech model training equipment can also include camera, RF (Radio Frequency, radio frequency) circuit,
Sensor, voicefrequency circuit, WiFi module etc..It will be understood by those skilled in the art that the training of speech model shown in Fig. 1 is set
Standby structure does not constitute the restriction to speech model training equipment, may include than illustrating more or fewer components or group
Close certain components or different component layouts.
As shown in Figure 1, as may include that operating system, network are logical in a kind of memory 1005 of computer storage medium
Believe module, Subscriber Interface Module SIM and speech model training program.
In speech model training equipment shown in Fig. 1, network interface 1004 is mainly used for connecting background server, and rear
Platform server carries out data communication；User interface 1003 is mainly used for connecting client (user terminal), carries out data with client
Communication；And processor 1001 can be used for calling the speech model training program stored in memory 1005, and execute following behaviour
Make:
The voice data of target speaker is obtained, and the voice data is divided into multiple voice data sections；
Vocal print feature is extracted respectively from the multiple voice data section, obtains multiple vocal print feature vectors；
The multiple vocal print feature vector is ranked up according to predetermined order mode, and chooses target based on ranking results
Vocal print feature vector；
Training pattern is treated based on the target vocal print feature vector to be trained, and obtains the voice of the target speaker
Identification model.
Further, the voice data for obtaining target speaker, and the voice data is divided into multiple voices
The step of data segment includes:
The voice data for obtaining target speaker, pre-processes the voice data；
The pretreated voice data is segmented according to prefixed time interval, obtains multiple voice data sections.
Further, described that the multiple vocal print feature vector is ranked up according to predetermined order mode, and based on row
Sequence result choose target vocal print feature vector the step of include:
Seek the average vector of the multiple vocal print feature vector；
Calculate the distance between the multiple vocal print feature vector and the average vector, and by the multiple vocal print feature
Vector sorts according to the sequence of the average vector distance from big to small；
The vocal print feature vector for falling into preset range is chosen as target vocal print feature vector based on ranking results.
Further, it is described to training pattern be deep neural network model, it is described based on the target vocal print feature to
Amount treats training pattern and includes: the step of being trained, obtain the speech recognition modeling of the target speaker
By the target vocal print feature input initialization treated the deep neural network model, it is iterated instruction
Practice；
When detect the deep neural network model convergence after, using the deep neural network model after convergence as
The speech recognition modeling of the target speaker.
Further, described to extract vocal print feature respectively from the multiple voice data section, obtain multiple vocal print features
The step of vector includes:
Based on preparatory trained universal background model self-adaptive processing is carried out to the multiple voice data section respectively, obtained
To multiple vocal print feature vectors.
Further, described that the multiple voice data section is carried out respectively based on preparatory trained universal background model
Self-adaptive processing, before the step of obtaining multiple vocal print feature vectors, processor 1001 can be used for calling in memory 1005
The speech model training program of storage, and execute following operation:
Trained voice data gathered in advance is pre-processed；
Training phonetic feature is extracted to the pretreated trained voice data；
Universal background model training is carried out by the trained phonetic feature, obtains universal background model；
Further, described to include: to the pretreated step of trained voice data progress gathered in advance
Preemphasis, framing, adding window and endpoint detection processing are successively carried out to trained voice data gathered in advance.
Based on above-mentioned hardware configuration, each embodiment of speech model training method of the present invention is proposed.
Referring to Fig. 2, speech model training method first embodiment of the present invention provides a kind of speech model training method, needs
Illustrate, it, in some cases, can be to be different from sequence herein although logical order is shown in flow charts
Execute shown or described step.The executing subject of each embodiment of speech model training method of the present invention can be PC,
The terminal devices such as smart phone, intelligent TV set, tablet computer and portable computer, for ease of description, in following embodiment
Middle omission executing subject is illustrated.The speech model training method includes:
Step S10, obtains the voice data of target speaker, and the voice data is divided into multiple voice data
Section；
Obtain the voice data of collected target speaker in advance, wherein for needing to train speech recognition modeling
People, as target speaker.In advance there are many kinds of the modes of acquisition target speaker voice data, it can such as be recorded by phone
The mode of sound acquires the voice data of target speaker.The voice data of target speaker is divided into multiple voice data sections.
Specifically, can it is prespecified by voice data divide number of segment, such as it is prespecified by one minute voice data, be divided into 20
Section.
Further, step S10 includes:
Step S101 obtains the voice data of target speaker, pre-processes to the voice data；
After the voice data for obtaining target speaker, voice data is pre-processed.Specific preprocessing process can be with
Including preemphasis processing.Wherein, preemphasis processing is a kind of signal compensated in transmitting terminal to input signal high fdrequency component
Processing mode.With the increase of signal rate, signal is damaged very greatly in transmission process, in order to be compared in reception terminal
Good signal waveform, it is necessary to which impaired signal is compensated.The thought of pre-emphasis technique is exactly to increase at the beginning of transmission line
The radio-frequency component of strong signal, to compensate excessive decaying of the high fdrequency component in transmission process.Preemphasis does not have an impact to noise,
Therefore output signal-to-noise ratio can be effectively improved.Vocal cords and lip in speaker's voiced process can be eliminated by being handled using the preemphasis
It is interfered Deng caused by, can be with the effective compensation training pent-up high frequency section of voice data, and voice data height can be highlighted
The formant of frequency reinforces the signal amplitude of voice data, helps to extract phonetic feature.
Step S102 is segmented the pretreated voice data according to prefixed time interval, obtains multiple languages
Sound data segment.
Pretreated voice data is segmented according to prefixed time interval, obtains multiple voice data sections, wherein
Prefixed time interval can be configured as needed, such as be set as 10ms, then be divided into continuous voice data every 10ms
One section, obtain multiple voice data sections.
Step S20 extracts vocal print feature from the multiple voice data section respectively, obtains multiple vocal print feature vectors；
Vocal print feature is extracted respectively from multiple voice data sections, obtains multiple vocal print feature vectors, i.e. a voice number
A vocal print feature vector is extracted according to section.Vocal print feature vector can be the vector of the vocal print spectrum signature of identity user, In
In the present embodiment, there are many kinds of the modes of extraction vocal print feature vector, is such as extracted in voice data by gauss hybrid models
Vocal print feature vector.
Step S30 is ranked up the multiple vocal print feature vector according to predetermined order mode, and is based on ranking results
Choose target vocal print feature vector；
Multiple vocal print feature vectors are ranked up according to predetermined order mode, and choose target vocal print based on ranking results
Feature vector.Specifically, clustering processing can be carried out to multiple vocal print feature vectors by clustering algorithm, similar vocal print is special
Sign vector is classified as one kind, and classification number can be prespecified, is divided into 5 classes Ru prespecified, then passes through clustering algorithm for multiple sound
Line feature vector is classified.It is understood that it is every one kind in vocal print feature vector between most like, the sound in inhomogeneity
Line feature vector similitude is low.It can be ranked up, and be chosen sound according to the population size of the vocal print feature vector of every one kind
A kind of or a few class more than line feature vector, using the vocal print feature vector under the category as target vocal print feature vector, remaining
Vocal print feature vector is then given up.It should be noted that due to may include more noise in the voice data of target speaker, or
It is the sound comprising other people, voice data is divided into segment, and vocal print feature vector is extracted based on segment respectively, by phase
As vocal print feature vector be divided into same class, can make characterize target speaker vocal print feature vocal print feature vector quilt
Be divided into same class, due in voice data based on the voice of target speaker, by vocal print feature quantity it is most one
The vocal print feature vector of class or a few classes as target speaker, and other are low with target speaker's vocal print feature vector similitude
Noise data excluded, to purify the vocal print feature vector of target speaker.
Further, in the present embodiment, another feasible target vocal print feature vector acquisition modes, step are proposed
S30 includes:
Step S301 seeks the average vector of the multiple vocal print feature vector；
After extraction obtains multiple vocal print feature vectors, the average vector of multiple vocal print feature vectors can also be sought, is put down
The mode of seeking of equal vector seeks that the average mode of vector is identical with existing, seeks to each element of each vocal print feature vector
Average value is to get the average vector for arriving multiple vocal print feature vectors.
Step is 302, calculates the distance between the multiple vocal print feature vector and the average vector, and will be described more
A vocal print feature vector sorts according to the sequence of the average vector distance from big to small；
The distance between multiple vocal print feature vectors and average vector are calculated, specifically, calculates the side of distance between vector
There are many formulas, can such as calculate the COS distance between vocal print feature vector and average vector, is also possible to calculate vocal print feature
Euclidean distance between vector and average vector.After the distance between multiple vocal print feature vectors and average vector is calculated,
Multiple vocal print feature vectors are sorted according to the sequence of average vector distance from small to large.
Step is 303, chooses the vocal print feature vector for falling into preset range as target vocal print feature based on ranking results
Vector.
The vocal print feature vector for falling into preset range is chosen as target vocal print feature vector based on ranking results.Wherein,
Preset range, which can be, presets vocal print feature vector using first few items are come as target vocal print feature vector, will such as come
First 5 vocal print feature vectors are as target vocal print feature vector.In the present embodiment, due to the voice data of target speaker
In based on the voice spoken with target, therefore, the average vector of the multiple vocal print feature vectors extracted can most characterize target
The vocal print feature of speaker, using with average vector apart from nearest vocal print feature vector as the vocal print feature of target speaker to
Amount, can purify the vocal print feature vector of target speaker.
Step S40 treats training pattern based on the target vocal print feature vector and is trained, and obtains the target and speaks
The speech recognition modeling of people.
After selecting target vocal print eigen vector, training pattern is treated based on target vocal print feature vector and is trained,
Obtain the speech recognition modeling of target speaker, wherein can be neural network model, such as deep neural network to training pattern
Model, convolutional neural networks model.By using target vocal print feature vector as the output to training pattern, and treat trained mould
Type is trained, and the speech recognition modeling identified to target speaker can be obtained.When training obtains target speaker's
After speech recognition modeling, to identify whether the speaker of one section of voice is target speaker, for the voice of speaker to be identified
Data extract the vocal print feature vector of speaker to be identified, using target speaker speech recognition modeling to vocal print feature vector
It is identified, obtains identification probability value；If the identification probability value is greater than predetermined probabilities value, it is determined that be target speaker sheet
People.
In the present embodiment, by obtaining the voice data of target speaker, and the voice data of target speaker is drawn
It is divided into multiple voice data sections；Vocal print feature is extracted respectively from multiple voice data sections, obtains multiple vocal print feature vectors；It presses
Multiple vocal print feature vectors are ranked up according to predetermined order mode, and choose target vocal print feature vector based on ranking results；
Training pattern is treated based on target vocal print feature vector to be trained, and is obtained the speech recognition modeling of target speaker, is realized
Even if the voice data of target speaker is not pure, can extract from unpurified voice data being capable of accurate Characterization mesh
The vocal print feature vector of speaker's vocal print feature is marked, and the language of target speaker is obtained based on the training of accurate vocal print feature vector
Sound identification model, to improve the accuracy of target speaker's speech recognition modeling.
Further, it is based on above-mentioned first embodiment, speech model training method second embodiment of the present invention provides one kind
Speech model training method.It in the present embodiment, is deep neural network model to training pattern, the step S40 includes:
Step S401, by the target vocal print feature input initialization treated the deep neural network model, into
Row iteration training；
After acquiring target vocal print feature vector, by target vocal print feature input initialization treated depth mind
Training is iterated through network model.Wherein, in deep neural network (Deep Neural Networks, abbreviation DNN) model
Including the input layer, hidden layer and output layer being made of neuron.It include each between each layer in the deep neural network model
The weight of neuron connection and biasing, these weights and biasing determine the property and recognition effect of neural network model.Initially
Change deep neural network model, the initial value of weight and biasing in deep neural network model is set, experience can be directly used
Initial weight and biasing is arranged in value.
Target vocal print feature vector is first divided into the sample of preset group number, then is grouped to be input in neural network model and carry out
Training, i.e., be separately input to neural network model for the sample after grouping and be trained.Propagated forward algorithm is according to nerve net
A series of linear operations that weight, biasing and the input value of each neuron carry out in neural network model are connected in network model
With activation operation, since input layer, operation from level to level, operation always to output layer, until obtaining output valve.According to preceding to biography
Each layer of network in neural network model of output valve can be calculated by broadcasting algorithm, arrive last one layer of output valve until calculating.Pass through
Output valve is compared with the standard output value pre-set, calculate penalty values, and Detectability loss value whether be less than it is default
Penalty values, to judge whether deep neural network restrains, if be less than preset penalty values, it is determined that convergence, if being not less than,
Determine not converged, output valve based on deep neural network model carries out error-duration model at this time, updates deep neural network model
Target vocal print feature vector is substituted into deep neural network model again after update, obtains output valve by the weight of each layer and biasing,
Loop iteration, until detecting deep neural network model convergence.
Step S402, after detecting deep neural network model convergence, by the depth nerve net after convergence
Speech recognition modeling of the network model as the target speaker.
After detecting deep neural network model convergence, the deep neural network model after convergence is spoken as target
The speech recognition modeling of people.That is, after detecting that penalty values are less than preset penalty values, by current depth neural network model
The speech recognition modeling of target speaker is arrived in the weight of each layer and biasing as final weight with biasing.
In the present embodiment, depth nerve is substituted by the pure target vocal print feature vector of the target speaker that will be obtained
It is trained in network model, obtains the speech recognition modeling that can accurately identify target speaker.
Further, it is based on above-mentioned second embodiment, speech model training method 3rd embodiment of the present invention provides one kind
Speech model training method.In the present embodiment, the step S20 includes:
Step S201 carries out the multiple voice data section based on preparatory trained universal background model adaptive respectively
It should handle, obtain multiple vocal print feature vectors.
What a universal background model, universal background model (Universal Background Model, letter are trained in advance
Claim UBM) it is gauss hybrid models (the Gaussian Mixture for indicating a large amount of nonspecific speaker's phonetic feature distributions
Model, abbreviation GMM), since the training of UBM generallys use largely, voice number that channel unrelated unrelated with speaker dependent
According to, therefore generally it can be thought that UBM is the model unrelated with speaker dependent, it is only fitted the phonetic feature distribution of people, and
Some specific speaker is not represented.Gauss hybrid models are accurate with Gaussian probability-density function (i.e. normal distribution curve)
Ground quantifies things, and a things is decomposed into several moulds formed based on Gaussian probability-density function (i.e. normal distribution curve)
Type.
Based on preparatory trained universal background model self-adaptive processing is carried out to multiple voice data sections respectively, obtained more
A vocal print feature vector.It is for each voice data section, therefrom extracts a vocal print feature vector.Wherein, adaptively
Processing refer to using in universal background model with a part of nonspecific speaker's phonetic feature similar in voice data section as mesh
The processing method of speaker's voice data is marked, which can specifically use MAP estimation algorithm (Maximum A
Posteriori, abbreviation MAP) it realizes.MAP estimation is the estimation that rule of thumb data obtain to the amount for being difficult to observe, and is estimated
During meter, posterior probability, objective function (i.e. expression target vocal print feature mould need to be obtained using prior probability and Bayes' theorem
The expression formula of type) for the likelihood function of posterior probability, parameter value when acquiring the likelihood function maximum (can be used gradient decline to calculate
Method is found out like the maximum value for obtaining right function), also just realizing will be close with target speaker's voice data in universal background model
The effect trained together as target speaker's voice data of a part of nonspecific speaker's phonetic feature, according to acquiring seemingly
Parameter value when right function maximum gets target voiceprint feature model corresponding with voice data section.Wherein, target vocal print
Characteristic model is for calculating target vocal print feature vector field homoemorphism type, and target vocal print feature vector refers to through target vocal print feature
What model obtained, represent the feature vector of voice data section.
Further, before step S201, further includes:
Step S50 pre-processes trained voice data gathered in advance；
Trained voice data gathered in advance is pre-processed.Wherein, training voice data is a large amount of non-spies of acquisition
Determine the voice data of user.Specific preprocessing process, which can be, successively carries out preemphasis, framing, adding window to training voice data
And endpoint detection processing.Wherein, preemphasis processing is explained in the above-described embodiments, and in this not go into detail.
Training voice data after preemphasis is subjected to sub-frame processing.Framing refers to that the voice signal by whole section is cut into the language of several segments
Sound processing technique, the size of every frame are moved in the range of 10-30ms, using general 1/2 frame length as frame.Frame shifting refers to adjacent two frame
Between overlapping region, can be avoided adjacent two frame and change excessive problem.Carrying out sub-frame processing to trained voice data can incite somebody to action
Training voice data is divided into the voice data of several segments, can segment trained voice data, convenient for the extraction of training phonetic feature.
Training voice data after framing is subjected to windowing process.After carrying out sub-frame processing to training voice data, each frame is risen
Discontinuous place can all occur in beginning section and end end, so framing is mostly also bigger with the error of original signal.Using adding
Window is able to solve this problem, and the training voice data after can making framing becomes continuously, and each frame is showed
The feature of periodic function out.Windowing process specifically refers to handle training voice data using window function, and window function can be with
Select Hamming window.Windowing process carried out to training voice data, training voice data after enabling to framing is in the time domain
Signal becomes continuously to facilitate to extract the training phonetic feature of trained voice data.
Step S60 extracts training phonetic feature to the pretreated trained voice data；
Training phonetic feature is extracted to pretreated trained voice data.It specifically can be to pretreated training
Voice data makees Fast Fourier Transform (FFT), obtains the frequency spectrum for training voice data, and is obtained training voice data according to frequency spectrum
Power spectrum.Wherein, Fast Fourier Transform (FFT) (Fast Fourier Transformation, abbreviation FFT), refers to and utilizes computer
Calculate efficient, quick calculation method the general designation of discrete Fourier transform.Using this calculation method can make computer calculate from
It dissipates multiplication number required for Fourier transformation to be greatly reduced, the number of sampling points being especially transformed is more, fft algorithm calculation amount
Saving it is more significant.The power spectrum that training voice data is handled using melscale filter group obtains training voice data
Meier power spectrum.It wherein, using the power spectrum that melscale filter group handles training voice data is carried out to power spectrum
Mel-frequency analysis, and mel-frequency analysis be based on human auditory perception analysis.Observation discovery human ear is filtered just as one
Wave device group is the same, only focuses on certain specific frequency components (i.e. the sense of hearing of people is selective frequency), that is to say, that people
Ear only allows the signal of certain frequencies to pass through, and directly ignores the certain frequency signals for being not desired to perception.Specifically, melscale filters
Device group includes multiple filters, these filters are not but univesral distributions on frequency coordinate axis, is had much in low frequency region
Filter, be distributed than comparatively dense, but in high-frequency region, the number of filter just becomes fewer, is distributed very sparse.It can manage
Xie Di, in the high resolution of low frequency part, the auditory properties with human ear are consistent melscale filter group, this is also Meier
Where the physical significance of scale.Cutting is carried out to frequency-region signal by using mel-frequency scale filter group, so that last every
The corresponding energy value of a frequency band, if the number of filter is 22, then the Meier power spectrum that will obtain training voice data
Corresponding 22 energy values.Mel-frequency analysis is carried out by the power spectrum to training voice data, so that the power spectrum is protected
The frequency-portions closely related with human ear characteristic are kept, which can be well reflected out the spy for training voice data
Sign.Cepstral analysis is carried out on Meier power spectrum, obtains the mel-frequency cepstrum coefficient for training voice data, and by mel-frequency
Cepstrum coefficient is determined as training phonetic feature.
Step S70 carries out universal background model training by the trained phonetic feature, obtains universal background model；
After getting trained phonetic feature, universal background model training is carried out by training phonetic feature, is obtained general
Background model.The training phonetic feature can indicate that vector can be read directly in computer equipment in the form of vector (matrix)
The training voice data of form is inputted training phonetic feature by frame, and calculated using EM when carrying out universal background model training
Method iterative calculation obtains the parameter in universal background model expression formula, to obtain universal background model, EM algorithm is to calculate to contain
There is the common mathematical method of the probability density function of hidden variable, herein without repeating.
In the present embodiment, by extracting the voice feature data of multiple voice data sections using universal background model, make
It must not need that the vocal print for capableing of each voice data section vocal print feature of accurate characterization can be obtained by a large amount of voice data section
Feature vector.
Furthermore the embodiment of the present invention also proposes a kind of speech model training device, referring to Fig. 3, the speech model training cartridge
It sets and includes:
Segmentation module 10 is divided into multiple languages for obtaining the voice data of target speaker, and by the voice data
Sound data segment；
It is special to obtain multiple vocal prints for extracting vocal print feature respectively from the multiple voice data section for extraction module 20
Levy vector；
Order models 30 for being ranked up according to predetermined order mode to the multiple vocal print feature vector, and are based on
Ranking results choose target vocal print feature vector；
Training module 40 is trained for treating training pattern based on the target vocal print feature vector, is obtained described
The speech recognition modeling of target speaker.
Further, the segmentation module 10 includes:
Pretreatment unit pre-processes the voice data for obtaining the voice data of target speaker；
Segmenting unit obtains more for being segmented to the pretreated voice data according to prefixed time interval
A voice data section.
Further, the order models 30 include:
Computing unit, for seeking the average vector of the multiple vocal print feature vector；
Sequencing unit, for calculating the distance between the multiple vocal print feature vector and the average vector, and by institute
Multiple vocal print feature vectors are stated to sort according to the sequence of the average vector distance from big to small；
Selection unit, it is special as target vocal print for choosing the vocal print feature vector for falling into preset range based on ranking results
Levy vector.
Further, the training module 40 includes:
Training unit, for by the target vocal print feature input initialization treated the deep neural network mould
Type is iterated training；
Detection unit, for when detect the deep neural network model convergence after, by after convergence the depth mind
Speech recognition modeling through network model as the target speaker.
Further, the extraction module 20 includes:
Processing unit, for being carried out respectively based on preparatory trained universal background model to the multiple voice data section
Self-adaptive processing obtains multiple vocal print feature vectors.
Further, the speech model training device further include:
Preprocessing module, for being pre-processed to trained voice data gathered in advance；
The extraction model 20 is also used to extract training phonetic feature to the pretreated trained voice data；
The training pattern 40 is also used to carry out universal background model training by the trained phonetic feature, obtains general
Background model；
Further, the preprocessing module be also used to successively carry out trained voice data gathered in advance preemphasis,
Framing, adding window and endpoint detection processing.
The expansion content of the specific embodiment of speech model training device of the present invention and above-mentioned speech model training method
Each embodiment is essentially identical, and this will not be repeated here.
In addition, the embodiment of the present invention also proposes a kind of computer readable storage medium, the computer readable storage medium
On be stored with speech model training program, the speech model training program realizes voice mould as described above when being executed by processor
The step of type training method.
The expansion content of speech model of the present invention training equipment and the specific embodiment of computer readable storage medium with
Above-mentioned each embodiment of speech model training method is essentially identical, and this will not be repeated here.
It should be noted that, in this document, the terms "include", "comprise" or its any other variant are intended to non-row
His property includes, so that the process, method, article or the system that include a series of elements not only include those elements, and
And further include other elements that are not explicitly listed, or further include for this process, method, article or system institute it is intrinsic
Element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that including being somebody's turn to do
There is also other identical elements in the process, method of element, article or system.
The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment side
Method can be realized by means of software and necessary general hardware platform, naturally it is also possible to by hardware, but in many cases
The former is more preferably embodiment.Based on this understanding, technical solution of the present invention substantially in other words does the prior art
The part contributed out can be embodied in the form of software products, which is stored in one as described above
In storage medium (such as ROM/RAM, magnetic disk, CD), including some instructions are used so that terminal device (it can be mobile phone,
Computer, server, air conditioner or network equipment etc.) execute method described in each embodiment of the present invention.
The above is only a preferred embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair
Equivalent structure or equivalent flow shift made by bright specification and accompanying drawing content is applied directly or indirectly in other relevant skills
Art field, is included within the scope of the present invention.
Claims (10)
1. a kind of speech model training method, which is characterized in that the speech model training method includes:
The voice data of target speaker is obtained, and the voice data is divided into multiple voice data sections；
Vocal print feature is extracted respectively from the multiple voice data section, obtains multiple vocal print feature vectors；
The multiple vocal print feature vector is ranked up according to predetermined order mode, and chooses target vocal print based on ranking results
Feature vector；
Training pattern is treated based on the target vocal print feature vector to be trained, and obtains the speech recognition of the target speaker
Model.
2. speech model training method as described in claim 1, which is characterized in that the voice number for obtaining target speaker
According to, and the step of voice data is divided into multiple voice data sections includes:
The voice data for obtaining target speaker, pre-processes the voice data；
The pretreated voice data is segmented according to prefixed time interval, obtains multiple voice data sections.
3. speech model training method as described in claim 1, which is characterized in that it is described according to predetermined order mode to described
Multiple vocal print feature vectors are ranked up, and the step of choosing target vocal print feature vector based on ranking results includes:
Seek the average vector of the multiple vocal print feature vector；
Calculate the distance between the multiple vocal print feature vector and the average vector, and by the multiple vocal print feature vector
It sorts according to the sequence of the average vector distance from big to small；
The vocal print feature vector for falling into preset range is chosen as target vocal print feature vector based on ranking results.
4. speech model training method as described in claim 1, which is characterized in that it is described to training pattern be depth nerve net
Network model, it is described to treat training pattern based on the target vocal print feature vector and be trained, obtain the target speaker's
The step of speech recognition modeling includes:
By the target vocal print feature input initialization treated the deep neural network model, it is iterated training；
After detecting deep neural network model convergence, using the deep neural network model after convergence as described in
The speech recognition modeling of target speaker.
5. such as the described in any item speech model training methods of Claims 1-4, which is characterized in that described from the multiple language
The step of extracting vocal print feature respectively in sound data segment, obtaining multiple vocal print feature vectors include:
Based on preparatory trained universal background model self-adaptive processing is carried out to the multiple voice data section respectively, obtained more
A vocal print feature vector.
6. speech model training method as claimed in claim 5, which is characterized in that described based on preparatory trained general back
Before the step of scape model carries out self-adaptive processing respectively, obtains multiple vocal print feature vectors to the multiple voice data section,
Further include:
Trained voice data gathered in advance is pre-processed；
Training phonetic feature is extracted to the pretreated trained voice data；
Universal background model training is carried out by the trained phonetic feature, obtains universal background model.
7. speech model training method as claimed in claim 6, which is characterized in that described to trained voice number gathered in advance
Include: according to pretreated step is carried out
Preemphasis, framing, adding window and endpoint detection processing are successively carried out to trained voice data gathered in advance.
8. a kind of speech model training device, which is characterized in that the speech model training device includes:
Segmentation module is divided into multiple voice data for obtaining the voice data of target speaker, and by the voice data
Section；
Extraction module obtains multiple vocal print feature vectors for extracting vocal print feature respectively from the multiple voice data section；
Order models, for being ranked up according to predetermined order mode to the multiple vocal print feature vector, and based on sequence knot
Fruit chooses target vocal print feature vector；
Training module is trained for treating training pattern based on the target vocal print feature vector, obtains the target and say
Talk about the speech recognition modeling of people.
9. a kind of speech model training equipment, which is characterized in that the speech model training equipment include memory, processor and
It is stored in the speech model training program that can be run on the memory and on the processor, the speech model training journey
The step of speech model training method as described in any one of claims 1 to 7 is realized when sequence is executed by the processor.
10. a kind of computer readable storage medium, which is characterized in that be stored with voice mould on the computer readable storage medium
Type training program is realized as described in any one of claims 1 to 7 when the speech model training program is executed by processor
The step of speech model training method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910744145.5A CN110428842A (en) | 2019-08-13 | 2019-08-13 | Speech model training method, device, equipment and computer readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910744145.5A CN110428842A (en) | 2019-08-13 | 2019-08-13 | Speech model training method, device, equipment and computer readable storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110428842A true CN110428842A (en) | 2019-11-08 |
Family
ID=68415931
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910744145.5A Pending CN110428842A (en) | 2019-08-13 | 2019-08-13 | Speech model training method, device, equipment and computer readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110428842A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110942777A (en) * | 2019-12-05 | 2020-03-31 | 出门问问信息科技有限公司 | Training method and device for voiceprint neural network model and storage medium |
| CN111179942A (en) * | 2020-01-06 | 2020-05-19 | 泰康保险集团股份有限公司 | Voiceprint recognition method, voiceprint recognition device, voiceprint recognition equipment and computer readable storage medium |
| CN111833851A (en) * | 2020-06-16 | 2020-10-27 | 杭州云嘉云计算有限公司 | Method for automatically learning and optimizing acoustic model |
| CN112002343A (en) * | 2020-08-18 | 2020-11-27 | 海尔优家智能科技（北京）有限公司 | Speech purity recognition method and device, storage medium and electronic device |
| CN112034473A (en) * | 2020-08-31 | 2020-12-04 | 福建省特种设备检验研究院 | Method, device and equipment for measuring distance between guide rail brackets of elevator and storage medium |
| CN112233694A (en) * | 2020-10-10 | 2021-01-15 | 中国电子科技集团公司第三研究所 | Target identification method and device, storage medium and electronic equipment |
| CN112489628A (en) * | 2020-11-23 | 2021-03-12 | 平安科技（深圳）有限公司 | Voice data selection method and device, electronic equipment and storage medium |
| CN112509587A (en) * | 2021-02-03 | 2021-03-16 | 南京大正智能科技有限公司 | Method, device and equipment for dynamically matching mobile number and voiceprint and constructing index |
| CN112712790A (en) * | 2020-12-23 | 2021-04-27 | 平安银行股份有限公司 | Voice extraction method, device, equipment and medium for target speaker |
| CN112863476A (en) * | 2019-11-27 | 2021-05-28 | 阿里巴巴集团控股有限公司 | Method and device for constructing personalized speech synthesis model, method and device for speech synthesis and testing |
| CN112992154A (en) * | 2021-05-08 | 2021-06-18 | 北京远鉴信息技术有限公司 | Voice identity determination method and system based on enhanced voiceprint library |
| CN113053365A (en) * | 2021-03-12 | 2021-06-29 | 云知声智能科技股份有限公司 | Voice separation method, device, equipment and storage medium |
| CN113327622A (en) * | 2021-06-02 | 2021-08-31 | 云知声（上海）智能科技有限公司 | Voice separation method and device, electronic equipment and storage medium |
| CN113345466A (en) * | 2021-06-01 | 2021-09-03 | 平安科技（深圳）有限公司 | Main speaker voice detection method, device and equipment based on multi-microphone scene |
| CN113409794A (en) * | 2021-06-30 | 2021-09-17 | 平安科技（深圳）有限公司 | Optimization method and device of voiceprint recognition model, computer equipment 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 |
| US11908455B2 (en) | 2020-01-07 | 2024-02-20 | Tencent Technology (Shenzhen) Company Limited | Speech separation model training method and apparatus, storage medium and computer device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102074236A (en) * | 2010-11-29 | 2011-05-25 | 清华大学 | Speaker clustering method for distributed microphone |
| CN106971724A (en) * | 2016-01-14 | 2017-07-21 | 芋头科技（杭州）有限公司 | A kind of anti-tampering method for recognizing sound-groove and system |
| CN108777146A (en) * | 2018-05-31 | 2018-11-09 | 平安科技（深圳）有限公司 | Speech model training method, method for distinguishing speek person, device, equipment and medium |
| CN109256137A (en) * | 2018-10-09 | 2019-01-22 | 深圳市声扬科技有限公司 | Voice acquisition method, device, computer equipment and storage medium |
| CN109686377A (en) * | 2018-12-24 | 2019-04-26 | 龙马智芯（珠海横琴）科技有限公司 | Audio identification methods and device, computer readable storage medium |
-
2019
- 2019-08-13 CN CN201910744145.5A patent/CN110428842A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102074236A (en) * | 2010-11-29 | 2011-05-25 | 清华大学 | Speaker clustering method for distributed microphone |
| CN106971724A (en) * | 2016-01-14 | 2017-07-21 | 芋头科技（杭州）有限公司 | A kind of anti-tampering method for recognizing sound-groove and system |
| CN108777146A (en) * | 2018-05-31 | 2018-11-09 | 平安科技（深圳）有限公司 | Speech model training method, method for distinguishing speek person, device, equipment and medium |
| CN109256137A (en) * | 2018-10-09 | 2019-01-22 | 深圳市声扬科技有限公司 | Voice acquisition method, device, computer equipment and storage medium |
| CN109686377A (en) * | 2018-12-24 | 2019-04-26 | 龙马智芯（珠海横琴）科技有限公司 | Audio identification methods and device, computer readable storage medium |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112863476A (en) * | 2019-11-27 | 2021-05-28 | 阿里巴巴集团控股有限公司 | Method and device for constructing personalized speech synthesis model, method and device for speech synthesis and testing |
| CN110942777A (en) * | 2019-12-05 | 2020-03-31 | 出门问问信息科技有限公司 | Training method and device for voiceprint neural network model and storage medium |
| CN110942777B (en) * | 2019-12-05 | 2022-03-08 | 出门问问信息科技有限公司 | Training method and device for voiceprint neural network model and storage medium |
| CN111179942A (en) * | 2020-01-06 | 2020-05-19 | 泰康保险集团股份有限公司 | Voiceprint recognition method, voiceprint recognition device, voiceprint recognition equipment and computer readable storage medium |
| US11908455B2 (en) | 2020-01-07 | 2024-02-20 | Tencent Technology (Shenzhen) Company Limited | Speech separation model training method and apparatus, storage medium and computer device |
| CN111833851A (en) * | 2020-06-16 | 2020-10-27 | 杭州云嘉云计算有限公司 | Method for automatically learning and optimizing acoustic model |
| CN112002343A (en) * | 2020-08-18 | 2020-11-27 | 海尔优家智能科技（北京）有限公司 | Speech purity recognition method and device, storage medium and electronic device |
| CN112002343B (en) * | 2020-08-18 | 2024-01-23 | 海尔优家智能科技（北京）有限公司 | Speech purity recognition method and device, storage medium and electronic device |
| CN112034473A (en) * | 2020-08-31 | 2020-12-04 | 福建省特种设备检验研究院 | Method, device and equipment for measuring distance between guide rail brackets of elevator and storage medium |
| CN112034473B (en) * | 2020-08-31 | 2024-02-27 | 福建省特种设备检验研究院 | Elevator guide rail bracket spacing measuring method, device, equipment and storage medium |
| CN112233694A (en) * | 2020-10-10 | 2021-01-15 | 中国电子科技集团公司第三研究所 | Target identification method and device, storage medium and electronic equipment |
| CN112233694B (en) * | 2020-10-10 | 2024-03-05 | 中国电子科技集团公司第三研究所 | Target identification method and device, storage medium and electronic equipment |
| CN112489628B (en) * | 2020-11-23 | 2024-02-06 | 平安科技（深圳）有限公司 | Voice data selection method and device, electronic equipment and storage medium |
| CN112489628A (en) * | 2020-11-23 | 2021-03-12 | 平安科技（深圳）有限公司 | Voice data selection method and device, electronic equipment and storage medium |
| CN112712790A (en) * | 2020-12-23 | 2021-04-27 | 平安银行股份有限公司 | Voice extraction method, device, equipment and medium for target speaker |
| CN112712790B (en) * | 2020-12-23 | 2023-08-15 | 平安银行股份有限公司 | Speech extraction method, device, equipment and medium for target speaker |
| CN112509587A (en) * | 2021-02-03 | 2021-03-16 | 南京大正智能科技有限公司 | Method, device and equipment for dynamically matching mobile number and voiceprint and constructing index |
| CN113053365A (en) * | 2021-03-12 | 2021-06-29 | 云知声智能科技股份有限公司 | Voice separation method, device, equipment and storage medium |
| CN113053365B (en) * | 2021-03-12 | 2023-03-24 | 云知声智能科技股份有限公司 | Voice separation method, device, equipment 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 |
| CN112992154A (en) * | 2021-05-08 | 2021-06-18 | 北京远鉴信息技术有限公司 | Voice identity determination method and system based on enhanced voiceprint library |
| CN113345466A (en) * | 2021-06-01 | 2021-09-03 | 平安科技（深圳）有限公司 | Main speaker voice detection method, device and equipment based on multi-microphone scene |
| CN113345466B (en) * | 2021-06-01 | 2024-03-01 | 平安科技（深圳）有限公司 | Main speaker voice detection method, device and equipment based on multi-microphone scene |
| CN113327622A (en) * | 2021-06-02 | 2021-08-31 | 云知声（上海）智能科技有限公司 | Voice separation method and device, electronic equipment and storage medium |
| CN113409794B (en) * | 2021-06-30 | 2023-05-23 | 平安科技（深圳）有限公司 | Voiceprint recognition model optimization method, voiceprint recognition model optimization device, computer equipment and storage medium |
| CN113409794A (en) * | 2021-06-30 | 2021-09-17 | 平安科技（深圳）有限公司 | Optimization method and device of voiceprint recognition model, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110428842A (en) | Speech model training method, device, equipment and computer readable storage medium | |
| CN110289003B (en) | Voiceprint recognition method, model training method and server | |
| CN104732978B (en) | The relevant method for distinguishing speek person of text based on combined depth study | |
| CN103971680B (en) | A kind of method, apparatus of speech recognition | |
| CN110457432B (en) | Interview scoring method, interview scoring device, interview scoring equipment and interview scoring storage medium | |
| CN108922513B (en) | Voice distinguishing method and device, computer equipment and storage medium | |
| CN101404160B (en) | Voice denoising method based on audio recognition | |
| CN106504768B (en) | Phone testing audio frequency classification method and device based on artificial intelligence | |
| CN106847292A (en) | Method for recognizing sound-groove and device | |
| CN107610707A (en) | A kind of method for recognizing sound-groove and device | |
| CN110838286A (en) | Model training method, language identification method, device and equipment | |
| CN102486922B (en) | Speaker recognition method, device and system | |
| CN108922515A (en) | Speech model training method, audio recognition method, device, equipment and medium | |
| CN107993663A (en) | A kind of method for recognizing sound-groove based on Android | |
| CN108648769A (en) | Voice activity detection method, apparatus and equipment | |
| CN108986798B (en) | Processing method, device and the equipment of voice data | |
| CN108597505A (en) | Audio recognition method, device and terminal device | |
| CN113223536B (en) | Voiceprint recognition method and device and terminal equipment | |
| CN109036470B (en) | Voice distinguishing method, device, computer equipment and storage medium | |
| CN109872713A (en) | A kind of voice awakening method and device | |
| CN102968990A (en) | Speaker identifying method and system | |
| CN105654955B (en) | Audio recognition method and device | |
| CN111899757A (en) | Single-channel voice separation method and system for target speaker extraction | |
| CN110136726A (en) | A kind of estimation method, device, system and the storage medium of voice gender | |
| CN110019741A (en) | Request-answer system answer matching process, device, equipment and readable storage medium storing program for executing |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191108 |