CN109036390A - A kind of broadcast keyword recognition method based on integrated gradient elevator - Google Patents
A kind of broadcast keyword recognition method based on integrated gradient elevator Download PDFInfo
- Publication number
- CN109036390A CN109036390A CN201810929482.7A CN201810929482A CN109036390A CN 109036390 A CN109036390 A CN 109036390A CN 201810929482 A CN201810929482 A CN 201810929482A CN 109036390 A CN109036390 A CN 109036390A
- Authority
- CN
- China
- Prior art keywords
- broadcast
- keyword
- data
- training
- classifier
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 22
- 238000012549 training Methods 0.000 claims description 45
- 238000005070 sampling Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000009467 reduction Effects 0.000 claims description 7
- 238000011946 reduction process Methods 0.000 claims description 6
- 230000010354 integration Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 abstract description 6
- 238000010586 diagram Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000001228 spectrum Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 241001672694 Citrus reticulata Species 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000003631 expected effect Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000003786 synthesis reaction 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
- G10L15/08—Speech classification or search
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
- G10L25/24—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being the cepstrum
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Acoustics & Sound (AREA)
- Human Computer Interaction (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Signal Processing (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of broadcast keyword recognition methods based on integrated gradient elevator.In the presence of single keyword, the recall rate of keyword recognition (can be evaluated one of the important indicator of keyword recognition, and often need most promotion) promotion to 80% or more by the present invention, while keep 70% or so whole accuracy rate.Meanwhile the index F1 score for measuring uneven specimen discerning, the present invention can promote it to 0.31 or so from 0.04 or so of single gradient elevator in test sample, be greatly improved the reliability of identification.
Description
Technical field
The present invention relates to technology for information acquisition fields, and in particular to a kind of broadcast keyword based on integrated gradient elevator
Recognition methods.
Background technique
Broadcast keyword recognition is mainly used in broadcasted content analysis, obtains, efficiently in the relevant information of broadcasted content
Data mining and radio-frequency spectrum control etc. are widely used in.The working principle for broadcasting keyword recognition is from one
The segment containing special key words is found out automatically in section sound program recording, and broadcasted content is automatically analyzed according to keyword segment.It passes
On system, the work of broadcasted content analysis has the shortcomings that at high cost, time-consuming, is easy to appear fault mostly by being accomplished manually.And from
Dynamic broadcast keyword recognition can then be identified by reliable algorithm by computer or integrated form system, thus reduce at
Originally, improve efficiency and avoid the mistake being manually likely to occur.
The core of broadcast keyword recognition is to find out the algorithm of keyword from broadcast segment.Intuitively say, we
Rule-based algorithm can be designed, keyword is determined according to the feature of broadcast paragraph.However, since broadcast singal belongs to voice
One kind of signal, information content is larger, data structure is complex, so simple rule-based method tends not to reach
Expected effect.In addition to rule-based method, since broadcast singal belongs to one kind of voice signal, also have in previous method
The design handled using speech recognition system.But since broadcast singal is often compared with general voice signal difference
It is larger, there are the interference informations such as special noise and music background, and in radio-frequency spectrum control, often for secrecy
It needing, identifying system needs use under offline environment, so general speech recognition system is also difficult to know in broadcast keyword
Ideal effect Shang not obtained.In addition, since broadcast keyword recognition is generally required in face of a large amount of uneven sample (keyword phases
Very small part is only accounted for for whole broadcast), so general algorithm is easy to omit keyword, or non-keyword is accidentally divided into pass
Key word causes the mistake of identification.
Summary of the invention
For above-mentioned deficiency in the prior art, a kind of broadcast based on integrated gradient elevator provided by the invention is crucial
Word recognition methods solves the problems, such as that broadcast keyword recognition is easy mistake.
In order to achieve the above object of the invention, the technical solution adopted by the present invention are as follows: a kind of based on integrated gradient elevator
Broadcast keyword recognition method, comprising the following steps:
S1, the training broadcast segment that training broadcast is divided into 3-5s, and eigentransformation is carried out to training broadcast segment, it obtains
To training data MFCC feature;
S2, training sample is extracted in training data according to training data MFCC feature, and passes through random sampling with replacement
To multiple groups non-key word sample lack sampling at random, multiple groups balance training subset is obtained;
S3, Tomek Link noise reduction process is carried out to every group of balance training subset, obtains noise reduction balance training subset;
S4, by GBM algorithm to the independent gradient elevator model of noise reduction balance training trained of single keyword,
It obtains gradient and promotes classifier;
S5, classifier is promoted by pack algorithm integration gradient, obtains integrated gradient and promote classifier, and passes through training number
According to the probability threshold value for adjusting integrated gradient promotion classifier;
S6, the broadcast segment to be tested that broadcast to be tested is divided into 3-5s, and feature is carried out to broadcast segment to be tested
Transformation, obtains data MFCC feature to be tested;
S7, data MFCC feature to be tested is put into integrated gradient promotion classifier progress keyword identification, is identified
As a result.
Further: Tomek Link noise reduction process in the step S3 specifically: in balance training subset X in addition to
Data xiWith data xjData xk, i.e. xk∈X\{xi,xj, if meeting it to xiDistance and arrive xjDistance be all larger than xiWith
xjDistance, i.e. dist (xi,xj)<dist(xi,xk) and dist (xi,xj)<dist(xj,xk), then xiAnd xjFor a pair of of Tomek
Link, if this is to the data x in Tomek LinkiWith data xjBelong to different classes, then deletes xiOr xj。
Further: the specific steps of GBM algorithm in the step S4 are as follows:
S41, model F is enabledK(x) are as follows:
In above formula, fk(x;θk) it is the submodel that kth walks, αkFor fk(x;θk) corresponding weight, K is current step number, as
Current operation total step number, x are voice sample, i.e. training broadcast segment, θkFor fk(x;θk) parameter sets;
The distance between S42, the model predication value for calculating K+1 step and true value rK+1, calculation formula are as follows:
In above formula, L (y, FKIt (x)) is loss function, FKIt (x) is model predication value, y is true value;
The model parameter θ of S43, fitting K+1 stepK+1, fitting formula are as follows:
In above formula, θ is model parameter, fK+1(x;θk) it is the submodel that K+1 is walked;
S44, the weight α for calculating K+1 stepK+1, calculation formula are as follows:
In above formula, α is weight coefficient;
S45, by model FK(x) iteration obtains updated Optimized model FK+1(x), i.e., gradient promotes classifier:
FK+1(x)=FK(x)+αK+1fK+1(x;θK+1)。
Further: the specific method of adjustment that gradient promotes classifier probability threshold value is integrated in the step S5 are as follows:
S51, probability of the prediction example in crucial part of speech, calculation formula are calculated are as follows:
In above formula, T is the number of balance training subset, Ht(xi) be single classifier prediction result, αtIt is used
Weight takes αt=1/T;
S52, determine whether sample contains keyword, determine formula are as follows:
In above formula, δ is adjustable probability threshold value;
S53, keyword/non-key word probability threshold is determined as by probability adjustment output of the example in crucial part of speech
Value.
The invention has the benefit that the present invention can be by the recall rate (weight of evaluation keyword recognition of keyword recognition
Want one of index, and often need most promotion) promotion to 80% or more, while keep 70% or so whole accuracy rate.
Meanwhile the index F1 score for measuring uneven specimen discerning, the present invention can in test sample by it from single gradient
0.04 or so of elevator is promoted to 0.31 or so, is greatly improved the reliability of identification.
Detailed description of the invention
Fig. 1 is flow chart of the present invention;
Fig. 2 is that the present invention is based on the performance schematic diagrames of the single benchmark gradient elevator of complete data set;
Fig. 3 is performance schematic diagram of the present invention using the single gradient elevator of lack sampling scheme;
Fig. 4 is the performance schematic diagram that the present invention uses 5 integrated gradient elevators;
Fig. 5 is the performance schematic diagram that the present invention uses 10 integrated gradient elevators.
Specific embodiment
A specific embodiment of the invention is described below, in order to facilitate understanding by those skilled in the art this hair
It is bright, it should be apparent that the present invention is not limited to the ranges of specific embodiment, for those skilled in the art,
As long as various change is in the spirit and scope of the present invention that the attached claims limit and determine, these variations are aobvious and easy
See, all are using the innovation and creation of present inventive concept in the column of protection.
As shown in Figure 1, a kind of broadcast keyword recognition method based on integrated gradient elevator, which is characterized in that including
Following steps:
S1, the training broadcast segment that training broadcast is divided into 3-5s, and eigentransformation is carried out to training broadcast segment, it obtains
To training data MFCC feature.
S2, training sample is extracted in training data according to training data MFCC feature, and passes through random sampling with replacement
To multiple groups non-key word sample lack sampling at random, multiple groups balance training subset is obtained.Because keyword data is generally in data set
Middle accounting example is less, limited amount.If carrying out lack sampling to keyword data again, keyword lazy weight will lead to.Phase
Instead, since the data volume of non-key word itself is larger, it has been carried out to put back to random sampling, so that the non-key word sample retained
Amount is suitable with keyword sample size, then can not only reduce imbalance problem, but also will not destroy the manifold of keyword data distribution.Reason
In the case of thinking, if the number of fruit keyword data is mk, then balance sample in order to obtain, we can be arranged non-key word sample and adopt
Sample quantity p=mk.But it also needs to delete by Tomek Link noise reduction since we are subsequent some Chong Die with crucial part of speech
Non-key word example, therefore, we just need to be arranged in more some non-key word examples of sampling at the beginning
S3, Tomek Link noise reduction process is carried out to every group of balance training subset, obtains noise reduction balance training subset.
Tomek Link noise reduction process specifically: in balance training subset X in addition to data xiWith data xjData xk, i.e. xk∈
X\{xi,xj, if meeting it to xiDistance and arrive xjDistance be all larger than xiAnd xjDistance, i.e. dist (xi,xj)<dist
(xi,xk) and dist (xi,xj)<dist(xj,xk), then xiAnd xjFor a pair of of Tomek Link, if this is in Tomek Link
Data xiWith data xjBelong to different classes, then deletes xiOr xj。
S4, by GBM algorithm to the independent gradient elevator model of noise reduction balance training trained of single keyword,
It obtains gradient and promotes classifier, the specific steps of GBM algorithm are as follows:
S41, model F is enabledK(x) are as follows:
In above formula, fk(x;θk) it is the submodel that kth walks, αkFor fk(x;θk) corresponding weight, K is current step number, as
Current operation total step number, x are voice sample, i.e. training broadcast segment, θkFor fk(x;θk) parameter sets;
The distance between S42, the model predication value for calculating K+1 step and true value rK+1, calculation formula are as follows:
In above formula, L (y, FKIt (x)) is loss function, FKIt (x) is model predication value, y is true value;
The model parameter θ of S43, fitting K+1 stepK+1, fitting formula are as follows:
In above formula, θ is model parameter, fK+1(x;θk) it is the submodel that K+1 is walked;
S44, the weight α for calculating K+1 stepK+1, calculation formula are as follows:
In above formula, α is weight coefficient;
S45, by model FK(x) iteration obtains updated Optimized model FK+1(x), i.e., gradient promotes classifier:
FK+1(x)=FK(x)+αK+1fK+1(x;θK+1)。
S5, gradient promotion classifier is integrated by bagging method, obtain integrated gradient and promote classifier, and pass through training number
According to the probability threshold value for adjusting integrated gradient promotion classifier, the specific method of adjustment that gradient promotes classifier probability threshold value is integrated
Are as follows:
S51, probability of the prediction example in crucial part of speech, calculation formula are calculated are as follows:
In above formula, T is the number of balance training subset, Ht(xi) be single classifier prediction result, αtIt is used
Weight takes αt=1/T;
S52, determine whether sample contains keyword, determine formula are as follows:
In above formula, δ is adjustable probability threshold value;
S53, keyword/non-key word probability threshold is determined as by probability adjustment output of the example in crucial part of speech
Value, generally 0.5.
S6, the broadcast segment to be tested that broadcast to be tested is divided into 3-5s, and feature is carried out to broadcast segment to be tested
Transformation, obtains data MFCC feature to be tested.
S7, data MFCC feature to be tested is put into integrated gradient promotion classifier progress keyword identification, is identified
As a result.
In one embodiment of the invention, in order to show effectiveness of the invention, we have been extracted by 133 radio
One data of broadcast recoding composition.Wherein, sub-fraction contains keyword " Beijing time " (mandarin), and our mesh
I.e. by keyword from broadcast segment in identify.All broadcast audios segment all processed for 5 seconds, by this
The processing of sample, we obtain 6906 records in total, wherein 197 include keyword.
Since the label of broadcast keyword recognition has very big deflection (keyword/non-keyword sample size is uneven), number
According to classification can realize high-accuracy, therefore common accuracy rate by the way that all examples are simply predicted as non-key word
Index cannot sufficiently represent the quality of algorithm.In the classification of label unbalanced data, (it will usually be looked into using precision and recall rate
Full rate) index carry out measure algorithm quality.Using TP, FP, TN, FN presentation class result is determined as true positives, and false positive is Kidney-Yin
Property and false negative, may then pass through following formula calculate positive class precision precision and recall rate recall:
Same the method also can be applied to negative class.For specific positive or negative class, we can be calculated
The F1 score of this algorithm:
In an experiment, we pay close attention to four evaluation indexes: general classification accuracy, the recall rate of non-key word, keyword
Recall rate and crucial part of speech F1 score.We using the F1 score of crucial part of speech as net assessment index because we
Task be identification keyword.That our models export is the probability that example is minority class (keyword, label 1), Wo Menke
To adjust its δ value to obtain optimum prediction result.In our experiment, the δ value of test (open interval) from 0 to 1, precision length
It is 0.05.As shown in Fig. 2, illustrating classification accuracy, majority (non-key word) recall rate and a small number of (keyword) recall rates
Variation.4 kinds of different models are tested based on gradient elevator classifier in an experiment, and parameter has been carried out most by verifying
Good adjustment.Benchmark model is a single gradient elevator (xgboost realization) classifier.X-axis indicates δ value in the figure, and
And y-axis indicates rate of precision/recall rate.From figure 2 it can be seen that keyword recall rate and rate of precision are in the second width figure (training set)
Middle highest, and keyword recall rate is greatly reduced in verifying collection and test set, illustrates that model has overfitting.
Second model of test is the single gradient elevator classifier using lack sampling method.This method can solve
It is interpreted as " single model integrated ", the advantages of it can use lack sampling and Tomek Link, but do not have based on integrated classifier
Advantage.From figure 3, it can be seen that overfitting problem is alleviated, most of example is no longer predicted as non-pass by classifier
Keyword.Although the recall rate of non-key word data is declined, aggregate performance makes moderate progress.
Finally, the 5 and 10 gradient elevator classifiers based on pack algorithm integration are tested using identical data set,
These classifiers follow technology set forth above, as a result as shown in Figure 4 and Figure 5.The two figures can be seen that there are two promoted:
First, keyword/non-key word synthesis recall rate greatly improves.Secondly, when adding 1 classifier per in overall model more
When, influence of the different δ values to performance becomes significant.The selection of δ can be completed by verifying, and we can obtain most preferably
The forecast confidence of output and each example.
Table 1 shows the best F1 score for testing a small number of (keyword) data and accuracy/recall rate under the score.
There are one additional " balance F1 score " indexs, it means that assuming that keyword and the identical feelings of non-key word example quantity
Condition is got off the F1 score of calculating.The index further emphasizes the successful identification of crucial part of speech data, and increases it and retrieve recall rate.
Table 1
Claims (4)
1. a kind of broadcast keyword recognition method based on integrated gradient elevator, which comprises the following steps:
S1, the training broadcast segment that training broadcast is divided into 3-5s, and eigentransformation is carried out to training broadcast segment, it is instructed
Practice data MFCC feature;
S2, training sample is extracted in training data according to training data MFCC feature, and by random sampling with replacement to more
The random non-key word sample lack sampling of group, obtains multiple groups balance training subset;
S3, Tomek Link noise reduction process is carried out to every group of balance training subset, obtains noise reduction balance training subset;
S4, by GBM algorithm to the independent gradient elevator model of noise reduction balance training trained of single keyword, obtain
Gradient promotes classifier;
S5, classifier is promoted by pack algorithm integration gradient, obtains integrated gradient and promote classifier, and passes through training data tune
Whole integrated gradient promotes the probability threshold value of classifier;
S6, the broadcast segment to be tested that broadcast to be tested is divided into 3-5s, and eigentransformation is carried out to broadcast segment to be tested,
Obtain data MFCC feature to be tested;
S7, data MFCC feature to be tested is put into integrated gradient promotion classifier progress keyword identification, obtains recognition result.
2. the broadcast keyword recognition method according to claim 1 based on integrated gradient elevator, which is characterized in that institute
State Tomek Link noise reduction process in step S3 specifically: in balance training subset X in addition to data xiWith data xjData
xk, i.e. xk∈X\{xi,xj, if meeting it to xiDistance and arrive xjDistance be all larger than xiAnd xjDistance, i.e. dist (xi,
xj)<dist(xi,xk) and dist (xi,xj)<dist(xj,xk), then xiAnd xjFor a pair of of Tomek Link, if this is to Tomek
Data x in LinkiWith data xjBelong to different classes, then deletes xiOr xj。
3. the broadcast keyword recognition method according to claim 1 based on integrated gradient elevator, which is characterized in that institute
State the specific steps of GBM algorithm in step S4 are as follows:
S41, model F is enabledK(x) are as follows:
In above formula, fk(x;θk) it is the submodel that kth walks, αkFor fk(x;θk) corresponding weight, K is current step number, as currently
Operation total step number, x is voice sample, i.e., training broadcast segment, θkFor fk(x;θk) parameter sets;
The distance between S42, the model predication value for calculating K+1 step and true value rK+1, calculation formula are as follows:
In above formula, L (y, FKIt (x)) is loss function, FKIt (x) is model predication value, y is true value;
The model parameter θ of S43, fitting K+1 stepK+1, fitting formula are as follows:
In above formula, θ is model parameter, fK+1(x;θk) it is the submodel that K+1 is walked;
S44, the weight α for calculating K+1 stepK+1, calculation formula are as follows:
In above formula, α is weight coefficient;
S45, by model FK(x) iteration obtains updated Optimized model FK+1(x), i.e., gradient promotes classifier:
FK+1(x)=FK(x)+αK+1fK+1(x;θK+1)。
4. the broadcast keyword recognition method according to claim 1 based on integrated gradient elevator, which is characterized in that institute
It states and integrates the specific method of adjustment that gradient promotes classifier probability threshold value in step S5 are as follows:
S51, probability of the prediction example in crucial part of speech, calculation formula are calculated are as follows:
In above formula, T is the number of balance training subset, Ht(xi) be single classifier prediction result, αtFor used power
Weight, takes αt=1/T;
S52, determine whether sample contains keyword, determine formula are as follows:
In above formula, δ is adjustable probability threshold value;
S53, keyword/non-key word probability threshold value is determined as by probability adjustment output of the example in crucial part of speech.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810929482.7A CN109036390B (en) | 2018-08-15 | 2018-08-15 | Broadcast keyword identification method based on integrated gradient elevator |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810929482.7A CN109036390B (en) | 2018-08-15 | 2018-08-15 | Broadcast keyword identification method based on integrated gradient elevator |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109036390A true CN109036390A (en) | 2018-12-18 |
CN109036390B CN109036390B (en) | 2022-07-08 |
Family
ID=64631548
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810929482.7A Active CN109036390B (en) | 2018-08-15 | 2018-08-15 | Broadcast keyword identification method based on integrated gradient elevator |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109036390B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110299133A (en) * | 2019-07-03 | 2019-10-01 | 四川大学 | The method for determining illegally to broadcast based on keyword |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140303961A1 (en) * | 2013-02-08 | 2014-10-09 | Machine Zone, Inc. | Systems and Methods for Multi-User Multi-Lingual Communications |
CN106682642A (en) * | 2017-01-06 | 2017-05-17 | 竹间智能科技(上海)有限公司 | Multi-language-oriented behavior identification method and multi-language-oriented behavior identification system |
CN108010527A (en) * | 2017-12-19 | 2018-05-08 | 深圳市欧瑞博科技有限公司 | Audio recognition method, device, computer equipment and storage medium |
CN108257593A (en) * | 2017-12-29 | 2018-07-06 | 深圳和而泰数据资源与云技术有限公司 | A kind of audio recognition method, device, electronic equipment and storage medium |
-
2018
- 2018-08-15 CN CN201810929482.7A patent/CN109036390B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140303961A1 (en) * | 2013-02-08 | 2014-10-09 | Machine Zone, Inc. | Systems and Methods for Multi-User Multi-Lingual Communications |
CN106682642A (en) * | 2017-01-06 | 2017-05-17 | 竹间智能科技(上海)有限公司 | Multi-language-oriented behavior identification method and multi-language-oriented behavior identification system |
CN108010527A (en) * | 2017-12-19 | 2018-05-08 | 深圳市欧瑞博科技有限公司 | Audio recognition method, device, computer equipment and storage medium |
CN108257593A (en) * | 2017-12-29 | 2018-07-06 | 深圳和而泰数据资源与云技术有限公司 | A kind of audio recognition method, device, electronic equipment and storage medium |
Non-Patent Citations (3)
Title |
---|
E FONSECA等: ""Acoustic scene classification by ensembling gradient boosting machine and convolutional neural networks"", 《DETECTION AND CLASSIFICATION OF ACOUSTIC SCENES AND EVENTS 2017》 * |
QIONG GU等: ""Data Mining on Imbalanced Data Sets"", 《2008 INTERNATIONAL CONFERENCE ON ADVANCED COMPUTER THEORY AND ENGINEERING》 * |
张春霞: ""集成学习中有关算法的研究"", 《中国博士学位论文全文数据库》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110299133A (en) * | 2019-07-03 | 2019-10-01 | 四川大学 | The method for determining illegally to broadcast based on keyword |
CN110299133B (en) * | 2019-07-03 | 2021-05-28 | 四川大学 | Method for judging illegal broadcast based on keyword |
Also Published As
Publication number | Publication date |
---|---|
CN109036390B (en) | 2022-07-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108537544B (en) | Real-time monitoring method and monitoring system for transaction system | |
CN102799899B (en) | Special audio event layered and generalized identification method based on SVM (Support Vector Machine) and GMM (Gaussian Mixture Model) | |
Ullrich et al. | Boundary Detection in Music Structure Analysis using Convolutional Neural Networks. | |
CN107766929B (en) | Model analysis method and device | |
CN102982804A (en) | Method and system of voice frequency classification | |
CN109784388A (en) | Stealing user identification method and device | |
Cartwright et al. | SONYC-UST-V2: An urban sound tagging dataset with spatiotemporal context | |
CN105373894A (en) | Inspection data-based power marketing service diagnosis model establishing method and system | |
CN111539585B (en) | Random forest-based power customer appeal sensitivity supervision and early warning method | |
CN107293308B (en) | A kind of audio-frequency processing method and device | |
CN107507038A (en) | A kind of electricity charge sensitive users analysis method based on stacking and bagging algorithms | |
CN112287980B (en) | Power battery screening method based on typical feature vector | |
CN112417132B (en) | New meaning identification method for screening negative samples by using guest information | |
CN109257383A (en) | A kind of BGP method for detecting abnormality and system | |
CN106909642A (en) | Database index method and system | |
CN109255029A (en) | A method of automatic Bug report distribution is enhanced using weighted optimization training set | |
CN110515836B (en) | Weighted naive Bayes method for software defect prediction | |
Xue et al. | Multi long-short term memory models for short term traffic flow prediction | |
CN115358481A (en) | Early warning and identification method, system and device for enterprise ex-situ migration | |
CN113543117A (en) | Prediction method and device for number portability user and computing equipment | |
CN109036390A (en) | A kind of broadcast keyword recognition method based on integrated gradient elevator | |
CN105162643B (en) | The method, apparatus and computing device that flow is estimated | |
CN109526027A (en) | A kind of cell capacity optimization method, device, equipment and computer storage medium | |
CN116561569A (en) | Industrial power load identification method based on EO feature selection and AdaBoost algorithm | |
CN109583763A (en) | Branch trade custom power load growth feature mining algorithm |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |