CN107154258A - Method for recognizing sound-groove based on negatively correlated incremental learning - Google Patents
Method for recognizing sound-groove based on negatively correlated incremental learning Download PDFInfo
- Publication number
- CN107154258A CN107154258A CN201710229138.2A CN201710229138A CN107154258A CN 107154258 A CN107154258 A CN 107154258A CN 201710229138 A CN201710229138 A CN 201710229138A CN 107154258 A CN107154258 A CN 107154258A
- Authority
- CN
- China
- Prior art keywords
- network
- training
- node
- trained
- incremental learning
- 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
- 238000000034 method Methods 0.000 title claims abstract description 30
- 230000002596 correlated effect Effects 0.000 title claims abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 41
- 238000000605 extraction Methods 0.000 claims abstract description 6
- 230000007935 neutral effect Effects 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 241001269238 Data Species 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000000875 corresponding effect Effects 0.000 claims description 3
- 238000009432 framing Methods 0.000 claims description 3
- 230000010076 replication Effects 0.000 claims description 3
- 238000013139 quantization Methods 0.000 claims description 2
- 210000004218 nerve net Anatomy 0.000 claims 1
- 238000004422 calculation algorithm Methods 0.000 abstract description 30
- 230000008859 change Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 230000001755 vocal effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000003340 mental effect Effects 0.000 description 1
- 210000000056 organ Anatomy 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000001228 spectrum Methods 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
- G10L15/00—Speech recognition
- G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
-
- 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
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
-
- 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
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
- G10L19/032—Quantisation or dequantisation of spectral components
-
- 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
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Signal Processing (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Evolutionary Computation (AREA)
- Image Analysis (AREA)
Abstract
The present invention is to provide a kind of method for recognizing sound-groove based on negatively correlated incremental learning.First, the voice signal of input is pre-processed and feature extraction;2nd, system integrating is initialized, if existing system integrating before, replicates current all-network;3rd, system integrating is trained;4th, structural adjustment is carried out to each network in system integrating;5th, current network is screened, selects wherein optimal a part of network;6th, currently available network is applied, if new data arrives, then circulates execution at the beginning from step.The present invention is that Application on Voiceprint Recognition is studied in the method for incremental learning, improves its efficiency and recognition accuracy under data increment arrival scene;Incremental Learning Algorithm based on negative correlation learning can efficiently solve increment problem.The present invention is improved in terms of model training and model select two respectively, it is proposed that the problem of a kind of new algorithm is to solve above-mentioned, is then applied in incremental learning.
Description
Technical field
The present invention relates to a kind of method for recognizing sound-groove.
Background technology
Vocal print refers to the information of acoustic wave figure for reflecting human speech sound spectrum of speaking, during Application on Voiceprint Recognition is exactly the oscillogram according to voice
The speech parameter comprising speaker characteristic reflected carrys out the automatic technology that speaker's identity is identified and judged.Due to every
The difference of personal vocal organs, the sound property issued and its tone are also different, therefore, regard vocal print as individual character
Identification and checking of the feature to realize different people have the characteristics, the technology of Application on Voiceprint Recognition such as quick, stable and accuracy rate height
Also information and the every field of network are widely applied to.
But under substantial amounts of application scenarios, training data can not be obtained disposably, and Application on Voiceprint Recognition also has
Well-known defect, personally for, although voice signal has relative stability, nor definitely unalterable
, it may have mutability, easily influenceed by from health, mental emotion and sound pick-up outfit.The hair of same person
Sound is under different periods, different physiological status, or is recorded under different microphone and channel parameter, the sound of its voice
Learn feature and be likely to occur change, so that the performance of identification, which is produced, to be influenceed, the machine learning side of at this moment traditional batch type
Formula, which seems, to be difficult in adapt to.
Negative correlation learning (Negative Correlation Learning, NCL) is that a kind of Artificial neural network ensemble study is calculated
Method, can solve increment problem, negative correlation learning is a kind of Ensemble Learning Algorithms based on ANN, parallel using BP algorithm
Training it is integrated in each individual neutral net.For training sample set { (x1,t1),...,(xn,tn) system integrating it is defeated
Go out for:
Wherein, M is integrated size, i.e. the number in integrated middle individual neutral net, Fi(n) it is i-th of neutral net
For the output of n-th of sample.
The characteristics of negative correlation learning is maximum is the design of its error function, and the error function is by mean square error and penalty term structure
Into i-th of network is as follows for the error function of n-th of sample:
Wherein tnIt is sample i target output, pi(n) it is penalty term, λ is a control parameter, for adjusting the power of punishment
Degree, value is [0,1], referred to as punishes term coefficient, when λ is 0, and interaction is not present between each network.Use the error function
It can to reach a balance between the mean square error for influenceing individual precision and the penalty term for influenceing interindividual variation.Wherein punish
Item is penalized to be defined as:
Pi (n)=(Fi(n)-F(n))∑j≠iFj(n)-F(n)
F in formulai(n) be network i output, F (n) is integrated output.
NCL illustrates good performance in incremental learning, earliest NCL by simple application in incremental learning when, research
Person has separately designed two kinds of algorithms:One kind is the fixed NCL of integrated size (Fixed Size NCL, FSNCL), another to be
Collection increases the NCL (Growing NCL, GNCL) of formula on a large scale.The advantage and defect of both algorithms are entered in 2.2.2 sections
Went analysis, for its advantage and disadvantage, Minlong Lin et al. propose selective negative correlation learning (Selective
Negative Correlation Learning, SNCL) algorithm.
The basic thought of the algorithm is, every time when new data arrives, will it is original it is integrated in all neutral nets replicate
It is a, it is assumed that integrated size is N, is learnt using the network after duplication on new data set, by all neutral nets
Integrate and obtain 2N network, then carry out a selection operation, brush selects N number of network therein, the network so obtained
In the network trained of existing use new data, contain what is above trained again, both ensure that integrated size kept constant,
Also so that the network remained can adapt to new data, and old knowledge can preferably be retained.SNCL has preferably
Generalization Capability, but but it is far above other algorithms on time complexity, in addition, also there is the hiding node layer of individual networks
Number is difficult to determine, easy to produce the phenomenons such as over-fitting.
The content of the invention
Can improve efficiency and recognition accuracy it is an object of the invention to provide one kind, reduction time complexity based on negative
The method for recognizing sound-groove of related incremental learning.
The object of the present invention is achieved like this:
Step 1: being pre-processed to the voice signal of input and feature extraction;
Step 2: initialization system integrating, if existing system integrating before, replicates current all-network;
Step 3: being trained to system integrating;
Step 4: carrying out structural adjustment to each network in system integrating;
Step 5: being screened to current network, wherein optimal a part of network is selected;
Step 6: currently available network is applied, if new data arrives, then circulate and hold at the beginning from step
OK.
The present invention can also include:
1st, the voice signal of described pair of input is pre-processed and feature extraction is specifically included:Firstly for raw tone letter
Number carry out sample quantization, then by analog signal with certain frequency sampling processing, become discrete data, then carry out at preemphasis
Reason, followed by framing and windowing process, is divided into a frame by the signal in a period of time, is analyzed in units of frame, then
Jing Yin section is filtered out using double threshold method, mel-frequency cepstrum coefficient (Mel- next is extracted to the voice signal after treated
Frequency Cepstral Coefficients, MFCC) and first-order difference MFCC is as model training and knows another characteristic ginseng
Number.
2nd, step 2 is specifically included:If not carrying out network training before, i.e., currently available data set is first
Batch, then initializing M neutral net is used for system integrating, wherein the number of each neutral net concealed nodes uses following public affairs
Formula is determined:
Otherwise by currently the network of trained mistake is replicated.In formula, n is input node number, and l saves for output
Point number, the number of input node and output node determines by the feature and problem to be solved of input sample, α be one
Random number on interval [1,10].
3rd, described be trained to system integrating specifically includes:D is obtained to training data progress repeated sampling individual different
Data set, the number of data set is identical with the number of network, i.e. D=M.Then concurrently instructed using Negative Correlation Training Method
Practice, the speed that steepness factor accelerates network training is introduced during training, if replicated to network, that is, is worked as
Preceding network number is 2M, then trains and carried out on network after the replication, otherwise primitive network is trained.
4th, each network progress structural adjustment in system integrating is specifically included:After training terminates, it is judged
Default target error whether is met, next step is carried out if meeting, each network being otherwise trained for step 3,
The importance of each of which concealed nodes is calculated, for each insignificant node, maximum important of an associated property is found
Node, the node of two correlation maximums is merged, and the mode of merging is to merge both corresponding weights, weight
The property wanted ηiCalculation formula it is as follows:
Wherein σiIt is concealed nodes i standard deviation, the output by the concealed nodes in all sample datas is calculated;μi
It is the number of times that the node is trained to so far.
Calculate two correlation of nodes formula be:
Wherein hiAnd h (p)j(p) it is respectively output to i and j-th of concealed nodes on sample p in training set, becomes
AmountWithIt is h respectivelyiAnd hjAverage value.
The present invention is that Application on Voiceprint Recognition is studied in the method for incremental learning, improves it under data increment arrival scene
Efficiency and recognition accuracy.Incremental Learning Algorithm based on negative correlation learning can efficiently solve increment problem, but existing
The selective Negative Correlation Training Method (SNCL) of classification degree of accuracy highest too high, the individual networks that remain time complexity
Hidden layer node number be difficult to determine, easily produce over-fitting the problems such as.The present invention is selected from model training and model respectively
Two aspects are improved, it is proposed that the problem of a kind of new algorithm is to solve above-mentioned, are then applied in incremental learning.
The present invention uses the incremental learning based on negative correlation learning (Negative Correlation Learning, NCL)
Algorithm is as the recognizer of vocal print, and in existing related algorithm, fixed-size FSNCL (Fixed Size NCL) is easily
Forgetting is produced to the knowledge learned, the overall Generalization Capabilities of GNCL (Growing NCL) for increasing formula are not good, based on selective collection
Into SNCL (Selective NCL) precision it is higher but the training time is long.The present invention is to SNCL algorithms from model training and mould
Type selects two aspects to be improved.First, NCL algorithms are based on BP networks and are modified so that individual net
Network error during training, to negatively correlated direction change, adds the otherness between network, but the network basis ignored
Easily there is hidden layer node number and is difficult determination in the training process in body, and the training time is long, easily produces over-fitting etc. and asks
Topic, the present invention NCL algorithms are improved for these problems so that network structure adaptive change, and by its with
Bagging algorithms are combined, referred to as ANCLBag algorithms, and the algorithm avoids purely manual setting concealed nodes number, makes its root
Situation about being trained according to current network, which adaptively increases or deleted, hides node layer, artificial error is reduced, in the mistake of adjustment
Cheng Zhong, excavates the information processing capability of network in itself, keeps node number less, on the one hand subtract in the training process as much as possible
Iterations is lacked, on the other hand, smaller unnecessary concealed nodes are likely to result in the possibility of over-fitting.In addition, with
After Bagging algorithms are combined, both influence in terms of training set generation and network training on whole integrated generation respectively so that
Difference between individual networks further increases, and ensure that overall Generalization Capability.Then, base of the present invention in the algorithm
The research of incremental learning has been carried out on plinth.During incremental learning is studied, the present invention has used for reference SNCL framework, and to it
The method of model selection is modified, and using the selective ensemble method based on sub-clustering and sequence, is considered simultaneously in the method
The accuracy and otherness of model, propose a kind of new SANCLBag algorithms, and experiment shows that the Generalization Capability of the algorithm is slightly higher
In SNCL, the latter is then substantially better than in terms of time complexity.
Finally, the model is applied in Application on Voiceprint Recognition, realize one voice signal can be pre-processed, feature
Extract, model incremental training and pattern-recognition Application on Voiceprint Recognition model, and using the model carried out parameter selection and increment test
The experiment of card, test result indicates that the model has higher recognition accuracy, and can efficiently solve asking for incremental learning
Topic.
Brief description of the drawings
Fig. 1 is FB(flow block) of the invention;
Fig. 2 is comparison diagram of the inventive algorithm with the algorithm before improving in terms of classification accuracy;
Fig. 3 is for inventive algorithm with the algorithm before improving in the time-related comparison diagram of algorithm;
Fig. 4 is inventive algorithm and other traditional algorithm classification accuracy comparison diagrams.
Embodiment
Illustrate below and the present invention is described in more detail.
With reference to Fig. 1, method of the invention is realized by following steps:
Step 1: being pre-processed to voice signal and feature extraction.Sampled firstly for original voice signal
Quantify, handled by analog signal with certain frequency sampling, so as to realize analog-to-digital conversion, become discrete data.Then carry out
Preemphasis processing, the section for making frequency high is highlighted, and filters out interference of the frequency compared with lower curtate.Next previous step is obtained
Result framing and windowing process, the voice signal that voice signal possesses in validity in short-term, i.e. 10ms to 30ms is relatively flat
Steady.Therefore, the signal in a period of time is divided into a frame, analyzed in units of frame, then filtered out using double threshold method
Jing Yin section.Next the spy that MFCC and first-order difference MFCC is used as model training and identification is extracted to the voice signal after treated
Levy parameter;
If Step 2: not carrying out network training before, i.e., currently available data set is first, then initializes
M neutral net is for integrated, wherein the number of each neutral net concealed nodes is determined using following empirical equation:
Otherwise by currently the network of trained mistake is replicated.In formula, n is input node number, and l saves for output
Point number, the number of input node and output node determines by the feature and problem to be solved of input sample, α be one
Random number on interval [1,10].
Step 3: repeated sampling is carried out to training data obtains the different data sets of D, the number of data set and network
Number is identical, i.e. D=M.Then concurrently trained using Negative Correlation Training Method, during training introduce steepness because
Son is to accelerate the speed of network training, if previous step is replicated to network, i.e., current network number is 2M, then trains
Carry out, primitive network is trained on network after the replication otherwise;
Step 4: after training terminates, judging whether it meets default target error, carried out if meeting next
Step, each network being otherwise trained for step 3 calculates the importance of each of which concealed nodes, for each insignificant
Node, find the maximum important node of an associated property (removing insignificant node, remaining is all important node), will
The node of two correlation maximums is merged, and the mode of merging is to merge both corresponding weights, importance ηi
Calculation formula it is as follows:
Here σiConcealed nodes i standard deviation, the standard deviation by the concealed nodes all sample datas output
Calculate, this value can show situation of change of the node for different sample datas well.When standard deviation is smaller, say
The bright concealed nodes export a very close value for all samples, nearly all, and in other words, the node is for difference
The separating capacity of sample is poor.Because in three layers of feed-forward network, hidden layer and output layer are joined directly together and connect, hidden layer
The difference of the result of different pieces of information will be produced directly on the output of network to be influenceed, it can be considered that the node is for net
It is unessential for network.Another parameter μ in formulaiIt is the number of times that the node is trained to so far, used here as instruction
The reason for practicing number of times is, because adjustment operation generates new node, and new node and node frequency of training before are differences
's.The formula for calculating two correlation of nodes is as follows:
Wherein hiAnd h (p)j(p) it is respectively output to i and j-th of concealed nodes on sample p in training set.Become
AmountWithIt is h respectivelyiAnd hjAverage value, average value here calculates after all samples have been trained.If through
Training before crossing, overall error reduction is not obvious, then increases a concealed nodes, and method is that one node of random selection enters
Line splitting;
If Step 5: current network number is 2M, sub-clustering is carried out to all networks, by the net of correlation maximum
Network is divided into cluster, and a classification degree of accuracy highest network is then selected from every cluster, if current network number is M,
Then directly carry out next step;
Step 6: the network now obtained is used to apply, if there is new data to arrive afterwards, return to step one.
Claims (5)
1. a kind of method for recognizing sound-groove based on negatively correlated incremental learning, it is characterized in that:
Step 1: being pre-processed to the voice signal of input and feature extraction;
Step 2: initialization system integrating, if existing system integrating before, replicates current all-network;
Step 3: being trained to system integrating;
Step 4: carrying out structural adjustment to each network in system integrating;
Step 5: being screened to current network, wherein optimal a part of network is selected;
Step 6: currently available network is applied, if new data arrives, then execution is circulated at the beginning from step.
2. the method for recognizing sound-groove according to claim 1 based on negatively correlated incremental learning, it is characterized in that described pair of input
Voice signal carry out pretreatment and feature extraction specifically include:Sample quantization is carried out firstly for primary speech signal, then will
Analog signal with certain frequency sampling processing, become discrete data, then carry out preemphasis processing, followed by framing and
Windowing process, is divided into a frame by the signal in a period of time, is analyzed in units of frame, then using double threshold method filter out it is quiet
Segment, it is MFCC and first-order difference MFCC as mould that mel-frequency cepstrum coefficient is next extracted to the voice signal after treated
Type training and the characteristic parameter of identification.
3. the method for recognizing sound-groove according to claim 2 based on negatively correlated incremental learning, it is characterized in that step 2 is specific
Including:If not carrying out network training before, i.e., currently available data set is first, then initializes M nerve net
Network is used for system integrating, wherein the number of each neutral net concealed nodes is determined using formula below:
Otherwise by currently the network of trained mistake is replicated, in formula, n is input node number, and l is output node
The number of number, input node and output node determines that α is one in interval by the feature and problem to be solved of input sample
[1,10] random number on.
4. the method for recognizing sound-groove according to claim 3 based on negatively correlated incremental learning, it is characterized in that described to network
Integrated be trained specifically includes:Repeated sampling is carried out to training data and obtains the different data sets of D, the number of data set and
The number of network is identical, i.e. D=M, is then concurrently trained, is introduced during training using Negative Correlation Training Method
Steepness factor accelerates the speed of network training, if replicated to network, i.e., current network number is 2M, then instructs
Practice and carried out on network after the replication, otherwise primitive network is trained.
5. the method for recognizing sound-groove according to claim 4 based on negatively correlated incremental learning, it is characterized in that described to network
Each network in integrated carries out structural adjustment and specifically included:After training terminates, judge whether it meets default target and miss
Difference, next step is carried out if meeting, and each network being otherwise trained for step 3 calculates each of which concealed nodes
Importance, for each insignificant node, finds a maximum important node of associated property, by two correlation maximums
Node merge, the mode of merging is to merge both corresponding weights, importance ηiCalculation formula it is as follows:
Wherein σiIt is concealed nodes i standard deviation, the output by the concealed nodes in all sample datas is calculated;μiIt is this
The number of times that node is trained to so far.
Calculate two correlation of nodes formula be:
Wherein hiAnd h (p)j(p) it is respectively output to i and j-th of concealed nodes on sample p in training set, variable
WithIt is h respectivelyiAnd hjAverage value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710229138.2A CN107154258A (en) | 2017-04-10 | 2017-04-10 | Method for recognizing sound-groove based on negatively correlated incremental learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710229138.2A CN107154258A (en) | 2017-04-10 | 2017-04-10 | Method for recognizing sound-groove based on negatively correlated incremental learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107154258A true CN107154258A (en) | 2017-09-12 |
Family
ID=59792651
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710229138.2A Pending CN107154258A (en) | 2017-04-10 | 2017-04-10 | Method for recognizing sound-groove based on negatively correlated incremental learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107154258A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110097183A (en) * | 2018-01-29 | 2019-08-06 | 松下电器(美国)知识产权公司 | Information processing method and information processing system |
CN110660399A (en) * | 2019-11-11 | 2020-01-07 | 广州国音智能科技有限公司 | Training method and device for voiceprint recognition, terminal and computer storage medium |
CN111310836A (en) * | 2020-02-20 | 2020-06-19 | 浙江工业大学 | Method and device for defending voiceprint recognition integrated model based on spectrogram |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102735442A (en) * | 2012-07-17 | 2012-10-17 | 华东理工大学 | Method for online monitoring and fault diagnosis of rotor |
CN103149878A (en) * | 2011-12-06 | 2013-06-12 | 中国科学院沈阳计算技术研究所有限公司 | Self-adaptive learning system of numerical control machine fault diagnosis system in multi-agent structure |
CN105021511A (en) * | 2015-07-28 | 2015-11-04 | 昆明理工大学 | Perforated tipping paper air permeability detection method applied to tobacco industry |
CN105160248A (en) * | 2015-07-02 | 2015-12-16 | 哈尔滨工程大学 | Correlation pruning neural network based identification system and method for malicious process of Xen virtual machine |
CN105373830A (en) * | 2015-12-11 | 2016-03-02 | 中国科学院上海高等研究院 | Prediction method and system for error back propagation neural network and server |
CN105572492A (en) * | 2015-10-22 | 2016-05-11 | 北京建筑大学 | City rail train auxiliary inverter fault diagnosis device |
CN106446942A (en) * | 2016-09-18 | 2017-02-22 | 兰州交通大学 | Crop disease identification method based on incremental learning |
-
2017
- 2017-04-10 CN CN201710229138.2A patent/CN107154258A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103149878A (en) * | 2011-12-06 | 2013-06-12 | 中国科学院沈阳计算技术研究所有限公司 | Self-adaptive learning system of numerical control machine fault diagnosis system in multi-agent structure |
CN102735442A (en) * | 2012-07-17 | 2012-10-17 | 华东理工大学 | Method for online monitoring and fault diagnosis of rotor |
CN105160248A (en) * | 2015-07-02 | 2015-12-16 | 哈尔滨工程大学 | Correlation pruning neural network based identification system and method for malicious process of Xen virtual machine |
CN105021511A (en) * | 2015-07-28 | 2015-11-04 | 昆明理工大学 | Perforated tipping paper air permeability detection method applied to tobacco industry |
CN105572492A (en) * | 2015-10-22 | 2016-05-11 | 北京建筑大学 | City rail train auxiliary inverter fault diagnosis device |
CN105373830A (en) * | 2015-12-11 | 2016-03-02 | 中国科学院上海高等研究院 | Prediction method and system for error back propagation neural network and server |
CN106446942A (en) * | 2016-09-18 | 2017-02-22 | 兰州交通大学 | Crop disease identification method based on incremental learning |
Non-Patent Citations (3)
Title |
---|
HAIBO HE, ETC: "Incremental Learning from Stream Data", <IEEE TRANSACTIONS ON NEURAL NETWORKS> * |
MD. MONIRUL ISLAM, MD. ABDUS SATTAR, MD. FAIJUL AMIN, ETC: "A New Adaptive Merging and Growing Algorithm for Designing Artificial Neural Networks", <IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS CYBERNETICS—PART B: CYBERNETICS> * |
林民龙: "《博士学位论文》", 30 January 2013, 中国科学技术大学 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110097183A (en) * | 2018-01-29 | 2019-08-06 | 松下电器(美国)知识产权公司 | Information processing method and information processing system |
CN110097183B (en) * | 2018-01-29 | 2024-03-01 | 松下电器(美国)知识产权公司 | Information processing method and information processing system |
CN110660399A (en) * | 2019-11-11 | 2020-01-07 | 广州国音智能科技有限公司 | Training method and device for voiceprint recognition, terminal and computer storage medium |
CN111310836A (en) * | 2020-02-20 | 2020-06-19 | 浙江工业大学 | Method and device for defending voiceprint recognition integrated model based on spectrogram |
CN111310836B (en) * | 2020-02-20 | 2023-08-18 | 浙江工业大学 | Voiceprint recognition integrated model defending method and defending device based on spectrogram |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Cai et al. | Sensor network for the monitoring of ecosystem: Bird species recognition | |
CN108172218B (en) | Voice modeling method and device | |
CN102982809B (en) | Conversion method for sound of speaker | |
CN102509547B (en) | Method and system for voiceprint recognition based on vector quantization based | |
CN112006697B (en) | Voice signal-based gradient lifting decision tree depression degree recognition system | |
CN110164476A (en) | A kind of speech-emotion recognition method of the BLSTM based on multi output Fusion Features | |
CN108766419A (en) | A kind of abnormal speech detection method based on deep learning | |
CN112581979B (en) | Speech emotion recognition method based on spectrogram | |
CN107492382A (en) | Voiceprint extracting method and device based on neutral net | |
CN108806667A (en) | The method for synchronously recognizing of voice and mood based on neural network | |
CN110299142B (en) | Voiceprint recognition method and device based on network convergence | |
CN108281146A (en) | A kind of phrase sound method for distinguishing speek person and device | |
CN112331216A (en) | Speaker recognition system and method based on composite acoustic features and low-rank decomposition TDNN | |
CN110827857B (en) | Speech emotion recognition method based on spectral features and ELM | |
AU2020102038A4 (en) | A speaker identification method based on deep learning | |
CN110379441B (en) | Voice service method and system based on countermeasure type artificial intelligence network | |
CN108922541A (en) | Multidimensional characteristic parameter method for recognizing sound-groove based on DTW and GMM model | |
CN107154258A (en) | Method for recognizing sound-groove based on negatively correlated incremental learning | |
CN108877812B (en) | Voiceprint recognition method and device and storage medium | |
CN109272986A (en) | A kind of dog sound sensibility classification method based on artificial neural network | |
Silva et al. | Evolving spiking neural networks for recognition of aged voices | |
Cao et al. | Speaker-independent speech emotion recognition based on random forest feature selection algorithm | |
CN108492821B (en) | Method for weakening influence of speaker in voice recognition | |
Ye et al. | Attention bidirectional LSTM networks based mime speech recognition using sEMG data | |
CN114299925A (en) | Method and system for obtaining importance measurement index of dysphagia symptom of Parkinson disease patient based on voice |
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 |
Application publication date: 20170912 |
|
RJ01 | Rejection of invention patent application after publication |