CN106557809A - Nerve network system and the method is trained by the nerve network system - Google Patents
Nerve network system and the method is trained by the nerve network system Download PDFInfo
- Publication number
- CN106557809A CN106557809A CN201510641501.2A CN201510641501A CN106557809A CN 106557809 A CN106557809 A CN 106557809A CN 201510641501 A CN201510641501 A CN 201510641501A CN 106557809 A CN106557809 A CN 106557809A
- Authority
- CN
- China
- Prior art keywords
- neural network
- row
- modules
- network system
- module
- 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
Landscapes
- Image Analysis (AREA)
Abstract
A kind of method for disclosing nerve network system and the nerve network system being trained.The nerve network system includes：The neural network module being connected in series of two or more columns, wherein, at least at least one of string neural network module in the two or more columns is not only connected to the upper strata neural network module of the row belonging to which and is connected to the upper strata neural network module of at least one other row.In accordance with an embodiment of the present disclosure, the nerve network system is not only connected in series different neural network modules by row integrated approach, and can further connect the neural network module of different lines so that different neural network modules have complementary advantages, so as to realize higher performance.
Description
Technical field
It relates to artificial neural network field, in particular it relates to a kind of can merge different god
The nerve network system of Jing mixed-media network modules mixed-medias and the method is trained by the nerve network system.
Background technology
Artificial neural network is to compare one of popular research direction in the last few years.Recent years, nerve
Network technology such as convolutional neural networks (CNN), long memory network (LSTM) in short-term and deep layer god
Jing networks (DNN) emerge in large numbers one after another.Different neural network modules has respective advantage and each
From limitation, how to merge these different neural network modules so that different neural network modules
Mutual supplement with each other's advantages, realize higher performance, become problem demanding prompt solution.
Specifically, by taking speech recognition technology as an example, speech recognition technology is extensively applied in many fields,
Including voice user interface such as phonetic dialing, call routing, remote household electrical appliance control, search, simple
Data input, structurized document compiling, speech-to-text process, and civil aviaton application etc..From
The advanced technology Deq of deep learning and big data based on CNN, LSTM or DNN, voice
The degree of accuracy of identification has obtained significantly being lifted.How different deep learning structures are merged so that no
With the mutual supplement with each other's advantages of deep learning structure so as to realize higher recognition accuracy, become current concern
One of hot issue.
The content of the invention
The brief overview with regard to the disclosure has been given below, so as to provide with regard to the disclosure some
The basic comprehension of aspect.It is understood, however, that this general introduction is not the exhaustive with regard to the disclosure
General introduction.It is not intended to critical component or pith for determining the disclosure, nor meaning
Figure is used for limiting the scope of the present disclosure.Its purpose is only given in simplified form with regard to the disclosure
Some concepts, in this, as preamble in greater detail given later.
In view of problem above, the purpose of the disclosure is to provide and a kind of can merge different neutral net moulds
The nerve network system of block and the method is trained by the nerve network system.
According to the one side of the disclosure, there is provided a kind of nerve network system, including：Two row or more
The neural network module being connected in series of row, wherein, at least string in the two or more columns
At least one neural network module be not only connected to the row belonging to which upper strata neural network module and
And it is connected to the upper strata neural network module of at least one other row.
According to another aspect of the present disclosure, additionally provide one kind and above-mentioned nerve network system is trained
Method, including：For each column in the two or more columns, using training data in each column
All neural network modules be trained；And entered based on all neural network modules in each column
Row the obtained result of training, using the training data to the two or more columns in all row
Combine and be trained.
According to the other side of the disclosure, additionally provide above-mentioned according to disclosed method for realizing
Computer program code and computer program and record has this to be used to realize above-mentioned basis thereon
The computer-readable recording medium of the computer program code of disclosed method.
The other side of the embodiment of the present disclosure is given in following description part, wherein, specifically
The bright preferred embodiment for fully disclosing the embodiment of the present disclosure, and which is not applied to limit.
Description of the drawings
The disclosure can be by reference to obtaining below in association with the detailed description given by accompanying drawing more preferably
Understanding, wherein in all of the figs using same or analogous reference come represent it is identical or
Similar part.The accompanying drawing is included in this manual and is formed together with detailed description below
A part for description, for preferred embodiment of the present disclosure is further illustrated and the disclosure is explained
Principle and advantage.Wherein：
Fig. 1 is the structural representation for illustrating nerve network system in accordance with an embodiment of the present disclosure；
Fig. 2 is the figure for illustrating the integrated deep learning model of row in accordance with an embodiment of the present disclosure；
Fig. 3 is the figure for illustrating trellis depth learning model in accordance with an embodiment of the present disclosure；
Fig. 4 is the figure for illustrating the DNN-HMM structures for speech recognition；
Fig. 5 be illustrate in accordance with an embodiment of the present disclosure for being trained to nerve network system
The flow chart of method；And
Fig. 6 is the personal meter for being shown as adoptable message processing device in embodiment of the disclosure
The block diagram of the exemplary construction of calculation machine.
Specific embodiment
It is described hereinafter in connection with accompanying drawing one exemplary embodiment of this disclosure.In order to clear and
All features of actual embodiment, for the sake of simple and clear, are not described in the description.However, should
Solution, must make many specific to embodiment during any this practical embodiments are developed
Determine, to realize the objectives of developer, for example, meet that related to system and business
Restrictive conditions, and these restrictive conditions a bit may be changed with the different of embodiment.
Additionally, it also should be appreciated that, although development is likely to be extremely complex and time-consuming, but to Deq
For those skilled in the art of present disclosure, this development is only routine task.
Here, in addition it is also necessary to which explanation is a bit, in order to avoid having obscured this public affairs because of unnecessary details
Open, illustrate only in the accompanying drawings the system structure closely related with the scheme according to the disclosure and/or
Process step, and eliminate other little with disclosure relation details.
Describe in detail below in conjunction with the accompanying drawings in accordance with an embodiment of the present disclosure.
First, by the nerve network system with reference to Fig. 1 descriptions in accordance with an embodiment of the present disclosure.Fig. 1
It is the structural representation for illustrating nerve network system 100 in accordance with an embodiment of the present disclosure.
Nerve network system 100 in accordance with an embodiment of the present disclosure includes the series connection of two or more columns
The neural network module of connection, wherein, at least string at least in the two or more columns
Individual neural network module is not only connected to the upper strata neural network module of the row belonging to which and is connected to
The upper strata neural network module of at least one other row.
As shown in figure 1, nerve network system 100 includes the neural network module that three row are connected in series,
Wherein each column includes three-layer neural network module；In each column, with C modules as ground floor nerve
Mixed-media network modules mixed-media, with L modules as second layer neural network module and with D-module as the 3rd
Layer neural network module.By taking the ground floor neural network module C in first row as an example, which not only connects
The second layer neural network module L being connected in first row, and it is connected to the god of the second layer in secondary series
Jing mixed-media network modules mixed-media L.In addition, by taking the ground floor neural network module C in secondary series as an example, which is not
But it is connected to the second layer neural network module L in secondary series, and second be connected in first row
Second layer neural network module L in the row of layer neural network module L and the 3rd.
But above-mentioned to be only exemplary rather than limiting, nerve network system 100 can be connected in series including two row
Neural network module or four neural network modules that are connected in series of row or the string more than four row
The neural network module of connection connection；Each column can include two-layer neural network module, four layers of neutral net
Module or the neural network module more than four layers；It is each arrange in same layer neural network module can be with
Partly different or mutually different, for example, the ground floor neural network module in first row can be C
During ground floor neural network module in module, secondary series can be C modules and the 3rd row
Ground floor neural network module can be L modules, in addition, the ground floor neutral net in first row
Module can be C modules, the ground floor neural network module in secondary series can be L modules, with
And the 3rd row in ground floor neural network module can be D-module, this is to other layers in each row
It is same to set up.Preferably, can be with the arrangement of each layer neural network module in empirically determined each column.
In addition, although fig 1 illustrate that in nerve network system 100, existing and being not only connected to its institute
The upper strata neural network module of the row of category and it is connected to the upper strata neutral net of at least one other row
The more than one neural network module of module, but simply by the presence of such neural network module be
Can.Preferably, the annexation of the neural network module in different lines can be determined by testing.
Preferably, each neural network module is pluggable.The pluggable property of neural network module makes
The composition for obtaining nerve network system is more flexible.For example, although every layer of neural network module is illustrated in Fig. 1
The neural network module of its last layer is all connected to, however, it is possible to utilize inserting for neural network module
Characteristic is pulled out, neural network module is connected thereto into the neural network module of two-layer.For example, can go
Fall the second layer neural network module L in the secondary series in Fig. 1, and cause the ground floor in first row
Ground floor neural network module C in neural network module C and secondary series is connected directly to second
Third layer neural network module D in row.Additionally, for example, row can be found with row by test
Between optimum connection, so as to the pluggable property using neural network module is realized in nerve network system
The optimum combination of the neural network module of each row.
Recent years, various nerual network techniques such as DNN, CNN and LSTM emerge in large numbers one after another.
These different neural network modules have respective advantage.For example, CNN modules can be by drop
Low frequency spectrum variance provides preferable feature, and LSTM modules can be carried by being provided more preferable feature
Performance is risen, and DNN modules can provide deeper network.Preferably, in the reality according to the disclosure
Apply in the nerve network system 100 of example, each neural network module can be CNN modules, LSTM
One in module and DNN modules.
In each column of nerve network system 100, the combination of different neural network modules is random.
But, we can also be combined based on some practical application experiences, for example can by
LSTM modules provide more preferable feature and carry out improving performance (and CNN modules can be by reducing frequency spectrum
Variance provides preferable feature), while can be by deepening between hidden layer module and output layer module
Mapping improving the prediction (and DNN modules can provide deeper network) of output.Preferably,
In nerve network system 100 in accordance with an embodiment of the present disclosure, in the two or more columns
At least in string, from the direction for be input to output successively include CNN modules, LSTM modules with
And DNN modules.It is by taking the first row of the nerve network system 100 in Fig. 1 as an example, defeated from being input to
On the direction for going out, ground floor neural network module can be CNN modules, second layer neutral net mould
It can be DNN modules that block can be LSTM modules and third layer neural network module.
Preferably, nerve network system 100 can also include combination layer, in the combination layer, to institute
The output for stating each column in two or more columns is combined.As shown in figure 1, to the output of each column (i.e.
The output of the third layer neural network module in each column) it is combined.Illustrate and it is unrestricted, can be right
The output of each column carries out linear combination.The combined treatment can further improve systematic function.
Preferably, as shown in figure 1, combinations thereof result can be input to HMM (implicit Ma Er
Section husband) system to be being decoded.
From the above description, it can be seen that nerve network system 100 in accordance with an embodiment of the present disclosure not only leads to
Cross row integrated approach and be connected in series different neural network modules, and can further connect different lines
Neural network module, and can be by testing the optimum connection that find between the column and the column so that it is same
Neural network module between neural network module and different lines in row has complementary advantages, so as to realize
Higher performance.
Nerve network system in accordance with an embodiment of the present disclosure is introduced with reference to speech recognition technology
100 specific example.
Recent years, in technical field of voice recognition, based on each of DNN, CNN and LSTM
Plant depth architecture and learning method is used widely.These different deep learning modules have
Respective advantage and respective limitation.Experimental data shows, based on different deep learning modules
The sentence set of speech recognition system identification mistake also differ, this diversity ensure that different deep
The fusion of degree study module can improve performance.
In in accordance with an embodiment of the present disclosure, different deep learning block coupled in series is connected to form row collection
Into deep learning model, the integrated deep learning model of the row is the nerve network system 100 shown in Fig. 1
In one row example.Fig. 2 is to illustrate the integrated deep learning mould of row in accordance with an embodiment of the present disclosure
The figure of type 200.As shown in Fig. 2 different deep learning modules is connected in series together, the row
Bottom acoustic featuress, defeated of the input of integrated deep learning model 200 for a frame or multiframe voice signal
Go out for voice class posterior probability.
Preferably, in the integrated deep learning model of row in accordance with an embodiment of the present disclosure 200, each
Deep learning module can be in CNN modules, LSTM modules and DNN modules.
In integrated deep learning model 200 is arranged, the combination of different deep learning modules is random
's.But, we can also be combined based on some practical application experiences.Preferably, in row
In integrated deep learning model 200, CNN moulds can be included from the direction for be input to output successively
Block, LSTM modules and DNN modules.For example, as shown in Fig. 2 from the side for being input to output
Upwards, ground floor deep learning module for CNN modules, second layer deep learning can be able to be
LSTM modules and third layer deep learning can be DNN modules.
The all parameters for arranging the different depth study module in integrated deep learning model 200 be while
What training was obtained.Row deep learning model 200 can be trained with cross entropy criterion, using random
Gradient declines (SGD) method and is optimized.
Preferably, the output for arranging integrated deep learning model 200 can be sent to single HMM solutions
Code device being decoded, so as to obtain final word sequence.
Further, in accordance with an embodiment of the present disclosure, by two or more integrated depth of row
Practise at least one of the integrated deep learning model of at least one of model 200 row deep learning mould
Block is not only connected to the upper strata deep learning module of the integrated deep learning model of row belonging to which and company
The upper strata deep learning module of the integrated deep learning model of at least one other row is connected to, grid depth is formed
Degree learning model, the trellis depth learning model is of the nerve network system 100 shown in Fig. 1
Example.
Fig. 3 is the figure for illustrating trellis depth learning model 300 in accordance with an embodiment of the present disclosure.As schemed
Shown in 3, three arrange integrated deep learning model by connecting and composing grid between deep learning module
Deep learning model 300.In trellis depth learning model 300, each integrated deep learning mould of row
Type includes three layer depth study modules；In each integrated deep learning model of row, made with C modules
For ground floor deep learning module, with L modules are as second layer deep learning module and use D
Module is used as third layer deep learning module.With first ground floor arranged in integrated deep learning model
As a example by deep learning module C, which is not only connected to first the arranged in integrated deep learning model
Two layer depth study module L, and it is connected to second second layer arranged in integrated deep learning model
Deep learning module L.In addition, with second the first layer depth arranged in integrated deep learning model
As a example by practising module C, which is not only connected to second the second layer depth arranged in integrated deep learning model
Degree study module L, and it is connected to first the second layer depth arranged in integrated deep learning model
Practise module L and the 3rd second layer deep learning module L arranged in integrated deep learning model.
Trellis depth learning model 300 shown in Fig. 3 is only example, similar in description neutral net
Mentioned during system 100, the structure of trellis depth learning model 300 can have various modifications.
Preferably, each the deep learning module in trellis depth learning model 300 is pluggable.
Relevant position in nerve network system 100 be can refer to regard to the pluggable characteristic of deep learning module
Description, here are not repeated description.
Preferably, in the particular problem of speech recognition, each arranges the defeated of integrated deep learning model
It is the acoustic featuress sequence obtained after wave filter group to enter, and it is then the other prediction knot of frame one-level to export
Really.
Preferably, as shown in figure 3, can be in the frame level of voice signal to different lines integrated study model
Output be combined.I.e., it is possible to pass through a combination layer by the defeated of different row integrated study models
Go out result to be combined.
Preferably, in order to simplify the anabolic process to arranging integrated deep learning model, we can be to institute
The posterior probability for having the frame level raw tone frame of the integrated deep learning model output of row carries out linear combination.
The linear combination defines a matrix.And the matrix is by by the different lines of frame level integrated depth
Practise model output and corresponding frame level desired value posterior probability study it is associated, so as to be trained
Arrive.It is in test phase, other come the frame level to the integrated deep learning model of different lines using the matrix
Posterior probability carry out linear combination.
Preferably, as shown in figure 3, in last layer of all of row integrated study model, will combine
Output be sent to single HMM decoders.That is, combined result can be input to one by us
The HMM systems that train, using dynamic programming being decoded.
The process of speech recognition, in order to simplify description, is described below based on DNN-HMM structures.
Fig. 4 is the figure for illustrating the DNN-HMM structures for speech recognition.
Structure as shown in Figure 4 mainly includes front end features part, DNN parts and HMM portions
Point.Front end features part is the observation portion in Fig. 4.Typically in speech recognition, fallen using Mel
Spectral coefficient feature (MFCC), that is to say will be transformed to per 20 milliseconds of voice signal a feature to
Amount.For the information using context, typically we are in present frame some frame features in front and back selected around
The DNN of rear end is sent into simultaneously.DNN parts mainly use the high-precision classification capacity of DNN
The characteristic vector of front end features part input is classified, and the label classified is mainly triphones
Status indicator in HMM model.HMM parts mainly use the status indicator being previously obtained, and enter
Row Veterbi decoding, so as to obtain final word sequence.
In trellis depth learning model 300 in accordance with an embodiment of the present disclosure, using the row collection of connection
The DNN parts in Fig. 4 are substituted into deep learning model, as trellis depth learning model 300 makes
Obtain the deep learning module and the integrated deep learning of different lines in the integrated deep learning model of same row
Deep learning module between model has complementary advantages, thus compared to above-mentioned DNN-HMM structures,
Higher recognition accuracy can be realized.
It is furthermore to be noted that, trellis depth learning model 300 here is the nerve net shown in Fig. 1
One example of network system 100, therefore in not describing in detail in trellis depth learning model 300
Appearance can be found in the description of relevant position in nerve network system 100, and here is not repeated description.
From the above description, it can be seen that trellis depth learning model 300 in accordance with an embodiment of the present disclosure
Different depth learning model is connected in series by row integrated approach not only, while further can connect making
With the integrated deep learning model of the row of different computing mechanisms, and different lines collection can be found by test
Into the optimum connection between deep learning model so that the depth in the integrated deep learning model of same row
Deep learning module between study module and the integrated deep learning model of different lines has complementary advantages, from
And high-rise stratification feature can be extracted from primary speech signal, so as to realize that higher identification is accurate
Exactness.
Additionally, except the nerve in accordance with an embodiment of the present disclosure introduced above in conjunction with speech recognition technology
Outside the application example of network system 100, those skilled in the art it is also readily appreciated that nerve network system
100 other application example, here are not repeated.
In addition, the disclosure additionally provide it is a kind of for being trained to above-mentioned nerve network system 100
Method.
By with reference to Fig. 5 describe in accordance with an embodiment of the present disclosure for carrying out to nerve network system 100
The flow example of the method for training.Fig. 5 be illustrate in accordance with an embodiment of the present disclosure for nerve net
The flow chart of the method 500 that network system 100 is trained.
As shown in figure 5, in accordance with an embodiment of the present disclosure for instructing to nerve network system 100
Experienced method 500 includes that single-row training step S502 and contigency close training step S504.
First, in single-row training step S502, for nerve network system 100 two row or more
Each column in multiple row, using training data to each column in all neural network modules be trained.Tool
Body ground, the parameter of each neural network module of each column in nerve network system 100 is by front
Obtained from all neural network modules to backward algorithm, in the row are trained simultaneously.
Then, close in training step S504 in contigency, based on all neutral net moulds in each column
Block is trained obtained result, using the training data to the institute in the two or more columns
There is contigency to close to be trained.Specifically, pass through another forward-backward algorithm algorithm using the training data
All contigencys are closed and is trained, i.e., individually train to above-mentioned Jing the result of the parameter of each column for obtaining to enter
Row fine setting obtain nerve network system 100 each row in each neural network module parameter with
And its between Connecting quantity.
Being used in accordance with an embodiment of the present disclosure is described with reference to trellis depth learning model 400
The specific example of the method 500 is trained by nerve network system.
First, each integrated deep learning model of row to trellis depth learning model 400, using instruction
Practice data to be trained all deep learning modules in each integrated deep learning model of row.Specifically
Ground, parameter of each deep learning module of each integrated deep learning model of row are to rear by front
Obtain to algorithm, to the training simultaneously of all deep learning modules in the integrated deep learning model of the row
's.
Then, for these different processes individually train the integrated deep learning model of row for obtaining, profit
The integrated deep learning model of all row is combined with the training data and be trained.Specifically, utilize
The training data is by another forward-backward algorithm algorithm to the integrated deep learning model joint of all row
It is trained, i.e., the parameter of each the integrated deep learning model of row for obtaining individually is trained to above-mentioned Jing
As a result it is finely adjusted in each the integrated deep learning model of row for obtaining trellis depth learning model 400
Each deep learning module parameter and the Connecting quantity between which.
It should be understood that the machine in storage medium and program product in accordance with an embodiment of the present disclosure can perform
Instruction can be configured to perform the above-mentioned method for being trained nerve network system, because
The content that this here is not described in detail refers to the description of previous relevant position, and here is not repeated to carry out
Description.
Correspondingly, for carrying the storage medium of the above-mentioned program product including the executable instruction of machine
It is also included within disclosure of the invention.The storage medium include but is not limited to floppy disk, CD, magneto-optic disk,
Storage card, memory stick etc..
In addition, it should also be noted that above-mentioned series of processes and system can also pass through software and/or
Firmware is realized.In the case where being realized by software and/or firmware, from storage medium or network to tool
There is the computer of specialized hardware structure, such as the general purpose personal computer 600 shown in Fig. 6 is installed and constituted
The program of the software, the computer are able to carry out various functions etc. when various programs are provided with.
In figure 6, CPU (CPU) 601 is according to read only memory (ROM) 602
The program of middle storage is loaded into random access memory (RAM) 603 from storage part 608
The various process of program performing.In RAM 603, also according to needs storage when CPU 601 is performed respectively
Plant data required during process etc..
CPU 601, ROM 602 and RAM 603 are connected to each other via bus 604.Input/defeated
Outgoing interface 605 is also connected to bus 604.
Components described below is connected to input/output interface 605：Importation 606, including keyboard, mouse
Deng；Output par, c 607, including display, such as cathode ray tube (CRT), liquid crystal display
(LCD) etc., and speaker etc.；Storage part 608, including hard disk etc.；With communications portion 609,
Including NIC such as LAN card, modem etc..Communications portion 609 is via network ratio
As the Internet performs communication process.
As needed, driver 610 is also connected to input/output interface 605.Detachable media 611
Such as disk, CD, magneto-optic disk, semiconductor memory etc. are installed in driver as needed
On 610 so that the computer program for reading out is installed in storage part 608 as needed.
In the case where above-mentioned series of processes is realized by software, it is situated between from network such as the Internet or storage
Matter such as detachable media 611 installs the program for constituting software.
It will be understood by those of skill in the art that this storage medium is not limited to shown in Fig. 6 wherein
Have program stored therein, and equipment separately distribute to provide a user with the detachable media 611 of program.
The example of detachable media 611 (includes light comprising disk (including floppy disk (registered trade mark)), CD
Disk read only memory (CD-ROM) and digital universal disc (DVD)), magneto-optic disk is (comprising mini
Disk (MD) (registered trade mark)) and semiconductor memory.Or, storage medium can be ROM 602,
Hard disk for including etc., wherein computer program stored in storage part 608, and with comprising their equipment
User is distributed to together.
Preferred embodiment of the present disclosure above by reference to Description of Drawings, but the disclosure be certainly not limited to
Upper example.Those skilled in the art can obtain various changes within the scope of the appended claims and repair
Change, and it should be understood that these changes and modification nature will be fallen in scope of the presently disclosed technology.
For example, the multiple functions being included in the embodiment above in a module can be by the dress for separating
Put to realize.As an alternative, the multiple functions of being realized by multiple modules in the embodiment above can respectively by
Separate device is realizing.In addition, one of function above can be realized by multiple modules.Needless to say,
Such configuration is included in scope of the presently disclosed technology.
In this specification, described in flow chart the step of, is not only included with the order temporally sequence
The process that row are performed, and including concurrently or individually rather than the place that must perform in temporal sequence
Reason.Additionally, or even in temporal sequence process the step of in, needless to say, suitably can also change
The order.
In addition, technical scheme below is disclosed in accordance with an embodiment of the present disclosure, including but not limited to：
A kind of 1. nerve network systems are attached, including：
The neural network module being connected in series of two or more columns,
Wherein, at least at least one of string neural network module in the two or more columns is not
But the upper strata neural network module of the row being connected to belonging to which and it is connected at least one other row
Upper strata neural network module.
Nerve network system of the note 2. according to note 1, wherein, each neural network module
It is pluggable.
Nerve network system of the note 3. according to note 1, wherein, each neural network module
It is convolutional neural networks CNN modules, long short term memory LSTM module and deep-neural-network
One in DNN modules.
Nerve network system of the note 4. according to note 3, wherein, in described two row or more
In at least string in row, include the CNN modules, institute from the direction for be input to output successively
State LSTM modules and the DNN modules.
Nerve network system of the note 5. according to note 1, also including combination layer, at described group
Close in layer, the output to each column in the two or more columns is combined.
Nerve network system of the note 6. according to note 1, wherein, the two or more columns
In each column input for a frame or multiframe voice signal bottom acoustic featuress, be output as voice class after
Test probability.
Nerve network system of the note 7. according to note 6, wherein, in the frame level of voice signal
Output to each column in the two or more columns is combined.
It is attached a kind of 8. nerve network systems to according to any one of note 1 to 7 to instruct
Experienced method, including：
For each column in the two or more columns, using training data to each column in all nerves
Mixed-media network modules mixed-media is trained；And
Obtained result is trained based on all neural network modules in each column, using described
Training data to the two or more columns in all contigencys close and be trained.
Claims (8)
1. a kind of nerve network system, including：
The neural network module being connected in series of two or more columns,
Wherein, at least at least one of string neural network module in the two or more columns is not
But the upper strata neural network module of the row being connected to belonging to which and it is connected at least one other row
Upper strata neural network module.
2. nerve network system according to claim 1, wherein, each neural network module
It is pluggable.
3. nerve network system according to claim 1, wherein, each neural network module
It is convolutional neural networks CNN modules, long short term memory LSTM module and deep-neural-network
One in DNN modules.
4. nerve network system according to claim 3, wherein, in described two row or more
In at least string in row, include the CNN modules, institute from the direction for be input to output successively
State LSTM modules and the DNN modules.
5. nerve network system according to claim 1, also including combination layer, at described group
Close in layer, the output to each column in the two or more columns is combined.
6. nerve network system according to claim 1, wherein, the two or more columns
In each column input for a frame or multiframe voice signal bottom acoustic featuress, be output as voice class after
Test probability.
7. nerve network system according to claim 6, wherein, in the frame level of voice signal
Output to each column in the two or more columns is combined.
8. one kind is instructed to nerve network system according to any one of claim 1 to 7
Experienced method, including：
For each column in the two or more columns, using training data to each column in all nerves
Mixed-media network modules mixed-media is trained；And
Obtained result is trained based on all neural network modules in each column, using described
Training data to the two or more columns in all contigencys close and be trained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510641501.2A CN106557809A (en) | 2015-09-30 | 2015-09-30 | Nerve network system and the method is trained by the nerve network system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510641501.2A CN106557809A (en) | 2015-09-30 | 2015-09-30 | Nerve network system and the method is trained by the nerve network system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106557809A true CN106557809A (en) | 2017-04-05 |
Family
ID=58417811
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510641501.2A Pending CN106557809A (en) | 2015-09-30 | 2015-09-30 | Nerve network system and the method is trained by the nerve network system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106557809A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108228782A (en) * | 2017-12-29 | 2018-06-29 | 山东科技大学 | A kind of implication relation based on deep learning finds method |
| CN108279692A (en) * | 2018-01-17 | 2018-07-13 | 哈尔滨工程大学 | A kind of UUV dynamic programming methods based on LSTM-RNN |
| CN108416744A (en) * | 2018-01-30 | 2018-08-17 | 百度在线网络技术（北京）有限公司 | Image processing method, device, equipment and computer readable storage medium |
| CN108668265A (en) * | 2017-12-29 | 2018-10-16 | 西安电子科技大学 | The method for predicting collision probability between mobile subscriber based on Recognition with Recurrent Neural Network |
| WO2018227781A1 (en) * | 2017-06-12 | 2018-12-20 | 平安科技（深圳）有限公司 | Voice recognition method, apparatus, computer device, and storage medium |
| CN110930981A (en) * | 2018-09-20 | 2020-03-27 | 深圳市声希科技有限公司 | Many-to-one voice conversion system |
| CN112968740A (en) * | 2021-02-01 | 2021-06-15 | 南京邮电大学 | Satellite spectrum sensing method based on machine learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824054A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded depth neural network-based face attribute recognition method |
| CN104732274A (en) * | 2015-03-10 | 2015-06-24 | 华南理工大学 | Intelligent computer |
| CN103529439B (en) * | 2013-10-23 | 2015-09-30 | 环境保护部卫星环境应用中心 | A kind of vegetation parameter remote sensing inversion method of nerve network system and device |
-
2015
- 2015-09-30 CN CN201510641501.2A patent/CN106557809A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103529439B (en) * | 2013-10-23 | 2015-09-30 | 环境保护部卫星环境应用中心 | A kind of vegetation parameter remote sensing inversion method of nerve network system and device |
| CN103824054A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded depth neural network-based face attribute recognition method |
| CN104732274A (en) * | 2015-03-10 | 2015-06-24 | 华南理工大学 | Intelligent computer |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018227781A1 (en) * | 2017-06-12 | 2018-12-20 | 平安科技（深圳）有限公司 | Voice recognition method, apparatus, computer device, and storage medium |
| CN108228782A (en) * | 2017-12-29 | 2018-06-29 | 山东科技大学 | A kind of implication relation based on deep learning finds method |
| CN108668265A (en) * | 2017-12-29 | 2018-10-16 | 西安电子科技大学 | The method for predicting collision probability between mobile subscriber based on Recognition with Recurrent Neural Network |
| CN108228782B (en) * | 2017-12-29 | 2020-04-21 | 山东科技大学 | Implicit relation discovery method based on deep learning |
| CN108279692A (en) * | 2018-01-17 | 2018-07-13 | 哈尔滨工程大学 | A kind of UUV dynamic programming methods based on LSTM-RNN |
| CN108279692B (en) * | 2018-01-17 | 2020-12-22 | 哈尔滨工程大学 | UUV dynamic planning method based on LSTM-RNN |
| CN108416744A (en) * | 2018-01-30 | 2018-08-17 | 百度在线网络技术（北京）有限公司 | Image processing method, device, equipment and computer readable storage medium |
| CN108416744B (en) * | 2018-01-30 | 2019-11-26 | 百度在线网络技术（北京）有限公司 | Image processing method, device, equipment and computer readable storage medium |
| CN110930981A (en) * | 2018-09-20 | 2020-03-27 | 深圳市声希科技有限公司 | Many-to-one voice conversion system |
| CN112968740A (en) * | 2021-02-01 | 2021-06-15 | 南京邮电大学 | Satellite spectrum sensing method based on machine learning |
| CN112968740B (en) * | 2021-02-01 | 2022-07-29 | 南京邮电大学 | Satellite spectrum sensing method based on machine learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106557809A (en) | Nerve network system and the method is trained by the nerve network system | |
| CN110491382B (en) | Speech recognition method and device based on artificial intelligence and speech interaction equipment | |
| CN107633842B (en) | Audio recognition method, device, computer equipment and storage medium | |
| CN104538024B (en) | Phoneme synthesizing method, device and equipment | |
| CN104575490B (en) | Spoken language pronunciation evaluating method based on deep neural network posterior probability algorithm | |
| CN103400577B (en) | The acoustic model method for building up of multilingual speech recognition and device | |
| CN108984683A (en) | Extracting method, system, equipment and the storage medium of structural data | |
| CN103049792B (en) | Deep-neural-network distinguish pre-training | |
| CN106997767A (en) | Method of speech processing and device based on artificial intelligence | |
| CN110264991A (en) | Training method, phoneme synthesizing method, device, equipment and the storage medium of speech synthesis model | |
| CN110491393A (en) | The training method and relevant apparatus of vocal print characterization model | |
| CN107220235A (en) | Speech recognition error correction method, device and storage medium based on artificial intelligence | |
| CN103366733A (en) | Text to speech system | |
| CN107301860A (en) | Audio recognition method and device based on Chinese and English mixing dictionary | |
| CN106297773A (en) | A kind of neutral net acoustic training model method | |
| CN104538028A (en) | Continuous voice recognition method based on deep long and short term memory recurrent neural network | |
| CN107103903A (en) | Acoustic training model method, device and storage medium based on artificial intelligence | |
| CN104143327A (en) | Acoustic model training method and device | |
| CN109147774B (en) | Improved time-delay neural network acoustic model | |
| CN104835493A (en) | Speech synthesis dictionary generation apparatus and speech synthesis dictionary generation method | |
| CN105845130A (en) | Acoustic model training method and device for speech recognition | |
| CN108804526A (en) | Interest determines that system, interest determine method and storage medium | |
| CN112634866B (en) | Speech synthesis model training and speech synthesis method, device, equipment and medium | |
| CN106683666A (en) | Field adaptive method based on deep neural network (DNN) | |
| CN110349569A (en) | The training and recognition methods of customized product language model and device |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170405 |