CN102543074A - Agricultural product data acquisition system as well as voice recognition system and method of mobile equipment - Google Patents

Abstract

The invention discloses an agricultural product data acquisition system as well as a voice recognition system and method of mobile equipment. An adopted voice recognition engine comprises a feature extracting module, an HMM (Hidden Markov Model) voice recognition system, an SVM (Support Vector Machine) feature vector extracting module, an SVM training module and a hybrid decoding module. A training sample of a voice signal is subjected to feature extraction to obtain a feature vector sequence; a feature vector with fixed length, which is used for the training of an SVM classifier, is obtained through the SVM feature vector extracting module; in the recognition process, syllable and segment information recognized by the HMM voice recognition system is constructed into feature vector information for the classification of the SVM classifier, then, distance information obtained through the SVM classifier is fitted into a posterior probability through a sigmoid function, and a hybrid decoder is input to obtain a final recognition result. According to the invention, the problem of poor operability of the mobile equipment can be effectively solved, the requirement for the informatization level of personnel in the acquisition process can be reduced, and the environmental suitability of agricultural field information acquisition can be improved.

Description

Mobile device agricultural product data acquisition system (DAS) and speech recognition system thereof, method

Technical field

The present invention relates to the intelligent information processing technology of Agricultural Information technical field, refer in particular to a kind of mobile device agricultural product data acquisition system (DAS) and speech recognition system thereof, method.

Background technology

Agricultural data acquisition is the source and the basis of IT application to agriculture; Under agriculture in modern times quick, the accurate and effective requirement; Agricultural on-site data gathering problem highlights day by day; Though " 3S " is technological, the applied research of wireless sensor technology obtains certain achievement; But the collecting method research of using towards the agricultural producer is still comparatively deficient, needs manually-operated occasion particularly outstanding at some, like data acquisition in agriculture field condition work data acquisition, the agricultural production process and agricultural product price data acquisition etc.Tradition relies on the collecting method of keyboard input and efficient and the quality that means directly influence data acquisition.Characteristics such as mobile collection equipment is little with its volume, reliability is high, price is low, good portability and support mobile operating; Be widely used in each links such as agricultural production, circulation; Make agricultural data acquisition accuracy and ageing on be greatly improved, but exist poor operability, collecting efficiency problem not high and that information acquisition personnel's the level of IT application is had relatively high expectations.

The collecting work of agricultural products in China information mainly relies on artificial the completion at present; Generally send information collector to inquire about and write down the price and the demand of various agricultural product to the scene, market by agricultural sector or market management mechanism; The breath keeper that notifies then place; Be input in the computing machine through arrangement, handle the back through Computer Analysis again and release through certain mode, promptness is not high; Receive a series of problems such as manpower, time, space constraint, these have all restricted the effective management to market to the raising of consumer service and relevant department.

Summary of the invention

The technical matters that the present invention will solve is: propose a kind of mobile device agricultural product data acquisition system (DAS) and speech recognition system thereof, method; Through phonetic entry recognition methods collection site data based on mobile device; Break away from the restriction of mobile device keypad and the small screen, improved the efficient of data acquisition simultaneously.

Technical solution of the present invention is following:

A kind of speech recognition system of the mobile device towards the agricultural product data acquisition; This system adopts the speech recognition engine based on the HMM/SVM mixed architecture; This speech recognition engine comprises characteristic extracting module, HMM speech recognition system, SVM proper vector extraction module and SVM training module and hybrid decoding module, and the training sample of voice signal is through obtaining feature vector sequence after the feature extraction; The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module; In identifying; The syllable paragraph information that the HMM speech recognition system is identified is built into the eigenvector information that can be used in the classification of svm classifier device; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.

A kind of mobile device agricultural product data acquisition system (DAS), this agricultural product data acquisition system (DAS) comprises the data acquisition module of being located in this mobile device, this data acquisition module comprises the speech recognition system of aforesaid mobile device towards the agricultural product data acquisition.

A kind of mobile device audio recognition method towards the agricultural product data acquisition, this method are that the speech recognition system with HMM is the basis, and the training sample of voice signal is through obtaining feature vector sequence after the feature extraction; The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module; In identifying; The syllable paragraph information that the HMM speech recognition system is identified is built into the eigenvector information that can be used in the classification of svm classifier device; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.

The present invention is a target with the informationalized demand of market for farm products; With price quotations, the product supply-demand information of saving manpower, the market for farm products main farm produce is fixed a point in the reflection in time of increasing work efficiency is main foundation; Adopt speech recognition to design practicable agricultural product data acquisition system (DAS) as man-machine interaction mode; Can overcome the problem of mobile device operation property difference effectively; To the requirement of personal information level, improve the environmental suitability that agriculture field data is gathered in the reduction gatherer process, for the construction of quickening market for farm products information systems, the transformation of promotion market for farm products management mode provide technical support.

Description of drawings

Fig. 1 is the basic framework figure of the mobile device speech recognition system towards the agricultural product data acquisition of the present invention.

Fig. 2 is the synoptic diagram of the speech recognition modeling that is adopted towards the mobile device speech recognition system and the method for agricultural product data acquisition of the present invention.

Fig. 3 is the multi-categorizer algorithm flow chart that is adopted towards the mobile device speech recognition system and the method for agricultural product data acquisition of the present invention.

Fig. 4 is the SVM/sigmoid combined training model synoptic diagram that is adopted towards the mobile device speech recognition system and the method for agricultural product data acquisition of the present invention.

Fig. 5 is the module map of mobile device agricultural product data acquisition system (DAS) function of the present invention.

Embodiment

The portability of mobile device is that the convenience with the loss man-machine interaction is a cost to a great extent; Relative PC; The screen of handheld device is little, the sunlight direct projection can't be seen clearly down and the inconvenience of keypad operation; Touch-screen full keyboard and handwriting recognition exist also that efficient is low, the problem of poor stability, have reduced the operability of equipment to a great extent.The present invention has broken away from the restriction of mobile device keypad and the small screen through the phonetic entry recognition methods collection site data based on mobile device, has improved the efficient of data acquisition simultaneously.

The present invention proposes a kind of speech recognition system and method for the mobile device towards the agricultural product data acquisition; This system adopts the speech recognition engine based on the HMM/SVM mixed architecture; This speech recognition engine comprises characteristic extracting module, HMM speech recognition system, SVM proper vector extraction module and SVM training module (being the svm classifier device among Fig. 2) and hybrid decoding module, and the training sample of voice signal is through obtaining feature vector sequence after the feature extraction; The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module; In identifying; The HMM speech recognition system is identified the pairing eigenvector information of paragraph information be built into the eigenvector information that can be used in the classification of svm classifier device equally; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.

The present invention proposes a kind of mobile device audio recognition method towards the agricultural product data acquisition; This method combines the needs of agriculture speech data acquisition speech and the preferential identification of digital speech identification; Utilize statistical language model technology in the medium and small vocabulary continuous speech recognition system; Adopt the speech recognition modeling of improved HMM/SVM mixed architecture, adopt the mode of probability to export to classification results through introducing the sigmoid function.

This method utilizes SVM can effectively solve the principle of small sample, non-linear and dimensions classification problem; In conjunction with the concrete actual conditions of using; Adopt gaussian radial basis function as kernel function; Analyze under different signal to noise ratio (S/N ratio)s and the different vocabularies gaussian kernel parameter and error punishment combinations of parameters SVM is promoted Effect on Performance, select in the anti-noise speech recognition system of its optimum combination application and unspecified person, middle and small scale vocabulary, to obtain recognition effect and noise resisting ability preferably.

The present invention protects the mobile device agricultural product data acquisition system (DAS) with aforementioned speech recognition system simultaneously, and this agricultural product data acquisition system (DAS) comprises data acquisition module, and this data acquisition module comprises aforementioned speech recognition engine based on the HMM/SVM mixed architecture.

Mobile device with this speech recognition system designs to the daily data capture management of market for farm products, and the major function of entire equipment is divided into five major parts: data acquisition module, data maintenance enquiry module, data transmitting module, data simultaneous module and system maintaining module.The staff can utilize speech recognition engine that the agricultural product price data are gathered, inquire about, revised and delete at operation field, and the data of collection can be uploaded onto the server through grouping service wireless and WLAN dual mode; System can carry out data sync to mobile device and database server, reliability and security during with the integrality and the consistance that keep data, assurance data transmission.The system function module structure is please with reference to Fig. 5.

Mobile device agricultural product data acquisition system (DAS) based on speech recognition engine of the present invention is the concrete application towards portable agricultural product data acquisition system (DAS); Possess unspecified person, medium and small vocabulary continuous speech recognition function; System is stable, reliable; It is few to take resource, and has scalability and extendability; Man-machine interaction is friendly, and is simple to operate, meets the cognitive custom of user.

Particularly; Speech recognition system adopts Structured Design; Audio frequency input front end and speech recognition part are independently got up, and speech recognition provides the communication interface between this engine and the upper level applications simultaneously as an independent engine; For application software embeds speech identifying function convenience is provided, has made the developer of upper level applications can ignore the realization details of speech recognition technology and realize speech identifying function.This speech recognition system comprises two parts in front-end and back-end, and front end is an application program, and the rear end then is a recognition engine.Application program is towards the final user; Obtain user's input (generally being speech data and environment and interchannel noise) through the voice object, realize and the communicating by letter of engine through calling SAPI (Speech Application Programming Interface) speech recognition application programming interface development interface again., be responsible for setting up the connection between application program and the recognition engine, send the order and the data of application program to recognition engine, and the result of engine is returned to application program.Recognition engine then is the backstage service routine; It is the core of whole speech recognition system; Be in the different process spaces with application program, be responsible for receiving the information of SAPI and handling accordingly: the information from SAPI mainly contains two types, and the one, system control information; Responsible request connects, request of loading model etc.; The 2nd, identifying information is passed to engine with the various types of voice data, discerns processing by engine, returns text results or set command id.System's basic structure frame diagram is please with reference to shown in Figure 1.

According to the needs of recognition system, recognition engine be except should be able to carrying out the speech recognition, can also implementation process between communication and user data managed.The engine internal structure mainly is made up of session, communication, data management and speech recognition four parts, and is specific as follows:

Session management: be the key component of engine design, both be responsible for realizing thread scheduling, realize the function of speech recognition again.For this reason, session is divided into two types, a type systematic session, one type is the identification session.System session has only one the tunnel, and the identification session then as required can one the tunnel, or multichannel.System session is in charge of the identification session task relevant with disposal system, mainly the discerning and and discern the processing of inter-related task of identification session.To each session, it is corresponding with it that engine all can start a worker thread.For the task of requiring to accomplish in real time, or some algorithm complex low, take few task of CPU time and adopt the synchronization call mode; High for algorithm complex, take many tasks of CPU time and then adopt the asynchronous call mode.Under the situation of asynchronous call, do not lost in order to guarantee data, we have created two round-robin queues of task and result to each session.The process of asynchronous call can be expressed as:

1) application call asynchronous function is filled in task toward task queue, and function returns;

2) engine takes out task in turn from task queue, the result is filled up to result queue after executing task;

3) application program is taken out execution result from result queue.

Communication mechanism: because application program and engine be not in the same process space; Support interprocess communication (Inter process Communication; IPC) main effect is: create the pipeline be used for data transfer and exchange between process; Utilize pipeline to accept the task that application program is sent, and execution result return to application program through pipeline.Consider structuring and extensibility; The design's engine is realized communication mechanism as an independent module; Such advantage is to change if desired the communication mode of engine and application program, and we only need rewrite this module, and relate to other part as few as possible.The situation application program of unit and engine-operated on an equipment adopts the method for Memory Mapping File and its to come the communication between implementation process.Its advantage is that structure is clear, is easy to realize; And since the visit be same memory field, can guarantee the speed and the accuracy of interprocess communication.

Data management: model files such as the acoustic layer that main administrative model data and user data in recognition engine, model data are meant in the speech recognition algorithm to be adopted, linguistic level, dictionary, the phonetic syntax; User data had both comprised some essential informations, like user name, password, user's description etc., also comprised the corresponding relation of each user and the model that adopts.Before connecting the identification session, the user must provide essential information and model information, otherwise recognizer correctly stress model discern.This engine according to set data structure, is kept at user data in the binary file.Because file read-write is the process of monopolizing, adopted the critical section technology.

Speech recognition: adopted speech recognition principle in the engine, can be divided into three parts: voice signal pre-service and feature extraction based on HMM; Acoustic model and pattern match and speech recognition aftertreatment.When setting up and using this system, the speech recognition core shows as foundation, training and three processes of identification of model respectively, and the identification aftertreatment partly is a sound word transfer process; According to recognition result; Utilization transfers the binary phonetic syntax of phonetic to carry out the phonetic beta pruning based on having, and obtains simplifying the multiple-length Syllable Lattice and sends into post-processing module, structure phonetic figure; Generate speech figure, the Chinese Character Recognition result who searches for to the end by dictinary information.

In conjunction with shown in Figure 2, the present invention is that the speech recognition system with HMM is the basis, sets up SVM proper vector extraction module, SVM training and hybrid decoding module three big modules, and the training sample of voice signal obtains feature vector sequence afterwards through feature extraction; The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module; In identifying; HMM is identified syllable-based hmm characteristic of correspondence Vector Message; It is built into the eigenvector information that can be used in the classification of svm classifier device equally; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.Be elaborated in the face of technical scheme involved in the present invention down:

For speech recognition engine, the present invention adopts the speech recognition modeling that improves the HMM/SVM mixed architecture.It is to the agriculture voice messaging identification problem based on mobile device; Needs in conjunction with agriculture speech data acquisition speech and the preferential identification of digital speech identification; Utilize statistical language model gordian technique in the medium and small vocabulary continuous speech recognition system; Advantage and the SVM of dynamic perfromance that HMM is good at processed voice is to the advantages of the strong classification capacity of static data, and the speech recognition modeling of this improvement HMM/SVM mixed architecture can reach the purpose of effective raising system identification accuracy rate.HMM/SVM mixture model commonly used is that the output with SVM is converted into probability, and as the output probability of each latent state among the HMM, but training speed is slow, and has aliasing between positive sample and the negative sample data, can not accurately distinguish for the large sample characteristic.The present invention adopts the mode of probability to export through introducing the sigmoid function to classification results, has solved the uncertainty that the sample aliasing brings effectively, has effectively improved the identification accuracy.

Design concept: utilize SVM can effectively solve the principle of small sample, non-linear and dimensions classification problem; In conjunction with the concrete actual conditions of using; Adopt gaussian radial basis function as kernel function; Analyze under different signal to noise ratio (S/N ratio)s and the different vocabularies gaussian kernel parameter and error punishment combinations of parameters SVM is promoted Effect on Performance, select in the anti-noise speech recognition system of its optimum combination application and unspecified person, middle and small scale vocabulary, to obtain recognition effect and noise resisting ability preferably.

Method for designing: the speech recognition system with conventional H MM is the basis; Increased SVM proper vector extraction module; SVM training and hybrid decoding module three big modules; The training process of in Fig. 1, being represented SVM by close dotted line, the training sample of voice signal is through obtaining feature vector sequence O, i.e. observation sequence after the feature extraction.The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module.Represent identifying by solid line; HMM is identified the syllable paragraph information be built into the eigenvector information that can be used for the classification of svm classifier device; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.The main framework of speech recognition modeling is as shown in Figure 2.

As shown in the figure; The present invention adopts the SVM multi-categorizer based on the HMM/SVM mixture model; Speech recognition system based on HMM is made up of acoustic model, pronunciation dictionary, search volume, search algorithm module usually, and wherein searching algorithm is the key of decision systems performance, and the structure of search volume not only itself affects taking of internal memory; And affecting the efficient of search, the mobile device speech recognition is mostly towards specific task and environment.Because priori is abundant, have characteristics simple, that precise and high efficiency is described the things Partial Feature, the present invention is employed in and adds the problem that can effectively alleviate SVM decision tree scheme error accumulation based on the sorter of priori in the decision tree.

In conjunction with shown in Figure 3, the method for designing of this multi-categorizer is the characteristics according to Chinese speech, adopts syllable to carry out modeling as the acoustics carrier; Distinguish polytypic result for effectively carrying out syllable; SVM decision tree multi-categorizer building method of the present invention is: the distance between all kinds of in the calculated characteristics space at first, and so all there is the distance value between k-1 and other types in each type, then it sorted by descending order; Again with all kinds of optimum lineoid of two-value SVM training algorithm structure; Delete its positive sample after having constructed a node, as positive sample, other samples are the two-value sorter of negative sample structure next node with next sample; Accomplish up to all node-classifications, the corresponding algorithm flow process is as shown in Figure 3.

Need to prove that for the classification problem of voice signal, the svm classifier performance is fine, but be not suitable for handling continuous input sample, need convert its output valve into available posterior probability output ability and HMM and make up mixture model.The present invention has adopted the computing method of a kind of SVM distance to posterior probability, and voice signal is carried out COMPREHENSIVE CALCULATING in identifying, and the output of SVM is converted into posterior probability.The SVM distance that the present invention adopted is following to the computing method of posterior probability:

The output format of SVM is y=sign (f (x)) (1)

Under the situation of coring, can be written as for training sample set

classification function

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{a}_i{y}_{j}K(x_i,x)+b---\left(2\right)$

In the formula, x is an input vector, x _i∈ R ⁿBe support vector, α _i∈ R is Lagrangian coefficient, y _i{+1,1} is corresponding α to ∈ _iLabel, K () is a kernel function, b ∈ R is the side-play amount of optimum lineoid.Y belongs to positive sample when f (x)＞0, and y belongs to negative sample when f (x)＜0, and y belongs to the probability of positive sample suc as formula shown in (3) under f (x) value, and f (x) is the SVM distance.

$P(y=1|f)=\frac{P\left(f\right|y=1)P_1}{P\left(f\right|y=1)P_1+P\left(f\right|y=-1)P_{-1}}---\left(3\right)$

The conditional probability hypothesis of each type by the Gaussian function model description does

Positive sample: $P\left(f\right|y=1)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_1^2}}\mathrm{Exp}\frac{-{(f-u_1)}^2}{2{\mathrm{\σ}}_1^2}---\left(4\right)$

Negative sample: $P\left(f\right|y=-1)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{-1}^2}}\mathrm{Exp}\frac{-{(f-u_{-1})}^2}{2{\mathrm{\σ}}_{-1}^2}---\left(5\right)$

Then formula (3) can be expressed as: $P(y=1|f)=\frac{1}{1+\frac{p_{-1}}{p_1}\mathrm{Exp}(-\frac{1}{2{\mathrm{\σ}}^2}({(f-u_1)}^2-{(f-u_{-1})}^2))}$

$=\frac{1}{1+\mathrm{Kexp}(-\frac{1}{2{\mathrm{\σ}}^2}((u_1^2-u_{-1}^2)+2(u_{-1}-u_1)f))}$

$=\frac{1}{1+\exp (\mathrm{Af}+B)}---\left(6\right)$

A and B are estimated parameters, when the model training parameter setting, can be obtained by the LibSVM model kit.

Describe in the face of SVM/sigmoid combined training model down:

Because its classification results of symbolic representation of SVM output valve, its absolute value representation sample be to the distance of classifying face, the therefore posterior probability in order to obtain calibrating converts the output of SVM the performance of probability model with the raising model to through the sigmoid function:

Parameter A in the formula (6) and B confirm through the maximum likelihood problem:

min _A，BF(A，B)＝∑ _i＝1(t _i(log(p _i)+(1-t _i)log(1-p _i)) (7)

P wherein _i=P (f _i),

$t_i=\left\{\begin{array}{cc}(N_{+}+1)/(N_{+}+2)& y_i=1\\ 1/(N_{-}+2)& y_i=-1\end{array}\right.---\left(8\right)$

N ₊Be positive sample number, N _-Be the negative sample number.

Training method is for training the HMM model with the voice of gathering through the Baum-Welch algorithm; The HMM topological structure that is adopted is for two shift the continuous HMM model of Multidimensional and Hybrid Gaussian density function from left to right; In training process, HMM is formed training pattern jointly as a part and the SVM of training pattern, with HMM the prime of training pattern; The training data that is used for the SVM model training there is the changing features of the property distinguished with this; And with the input vector of the multidimensional probability of HMM output as back level SVM model, the output of last SVM converts posterior probability into through Sigmoid function (6) formula of pressing, and specifically training process is as shown in Figure 4.

At last, the structure of the mobile device agricultural product data acquisition system (DAS) based on speech recognition engine of the present invention is following:

(1) data acquisition module: speech recognition engine is that the speech recognition input provides support; The agricultural product price data acquisition module is implemented in the working site to functions such as the collection of various agricultural product prices, preservation, modifications.

(2) data maintenance enquiry module: data maintenance module can realize to the various information that collect browse, delete, operation such as modification; The data query module provides the inquiry to price data.

(3) data transmitting module: under the remote situation, can be sent to the data of gathering on the server through the packet radio mode; Closely, realize the data transmission between server and the mobile device terminal through adopting the wireless local net mode.

(4) data simultaneous module: this module mainly adopts the merge replication technology, makes through WLAN and carries out exchanges data between mobile device and the server, to keep both consistance.

(5) system maintaining module: user management realizes the checking to login user and password, and the restriction unauthorized access guarantees the security of system; The agricultural product kind safeguard to realize functions such as the interpolation, deletion, modification to the agricultural product kind of information; Sound bank is safeguarded and is realized packing into and unloading of recognizing voice storehouse.

As for the realization of the concrete ingredient of each module, those skilled in the art can adopt kinds of schemes according to prior art, no longer is described in detail among this paper.

Though the present invention discloses with specific embodiment; But it is not in order to limit the present invention; Any those skilled in the art; The displacement of the equivalent assemblies of under the prerequisite that does not break away from design of the present invention and scope, having done, or, all should still belong to the category that this patent is contained according to equivalent variations and modification that scope of patent protection of the present invention is done.

Claims (9)

1. speech recognition system towards the mobile device of agricultural product data acquisition; It is characterized in that; This system adopts the speech recognition engine based on the HMM/SVM mixed architecture; This speech recognition engine comprises characteristic extracting module, HMM speech recognition system, SVM proper vector extraction module and SVM training module and hybrid decoding module, and the training sample of voice signal is through obtaining feature vector sequence after the feature extraction; The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module; In identifying; The syllable paragraph information that the HMM speech recognition system is identified is built into the eigenvector information that can be used in the classification of svm classifier device; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.

2. mobile device agricultural product data acquisition system (DAS); It is characterized in that; This agricultural product data acquisition system (DAS) comprises the data acquisition module of being located in this mobile device, and this data acquisition module comprises the speech recognition system of the mobile device towards the agricultural product data acquisition as claimed in claim 1.

3. mobile device agricultural product data acquisition system (DAS) as claimed in claim 2; It is characterized in that; Said data acquisition system (DAS) also comprises data transmitting module and data simultaneous module; Data collecting module collected to data upload to database server through grouping service wireless or WLAN, system can carry out data sync to mobile device and database server, keeps the integrality and the consistance of data.

4. mobile device agricultural product data acquisition system (DAS) as claimed in claim 3; It is characterized in that; Said data acquisition system (DAS) also comprises data maintenance enquiry module and system maintaining module; Wherein, data maintenance module comprises agricultural product data query module and agricultural product data maintenance module, can inquire about, browse, revise and delete the agricultural product price data at operation field; System maintaining module comprises user management module, sound bank maintenance module and agricultural product kind maintenance module, and user management module realizes the checking to login user and password, and the restriction unauthorized access guarantees the security of system; Agricultural product kind maintenance module is realized interpolation, deletion, the modification to the agricultural product kind of information; The sound bank maintenance module is realized packing into and unloading of recognizing voice storehouse.

5. the mobile device audio recognition method towards the agricultural product data acquisition is characterized in that, this method is that the speech recognition system with HMM is the basis, and the training sample of voice signal is through obtaining feature vector sequence after the feature extraction; The eigenvector that is used for the fixed length of svm classifier device training through the acquisition of SVM proper vector extraction module; In identifying; The syllable paragraph information that the HMM speech recognition system is identified is built into the eigenvector information that can be used in the classification of svm classifier device; Fit to posterior probability to the range information that obtains through the svm classifier device through the sigmoid function then, the input hybrid decoder obtains last recognition result.

6. the mobile device audio recognition method towards the agricultural product data acquisition as claimed in claim 5; It is characterized in that; This method is to adopt SVM decision tree multi-categorizer building method; This method is the characteristics according to Chinese speech, adopts syllable to carry out modeling as the acoustics carrier, and said SVM decision tree multi-categorizer building method comprises:

The distance between all kinds of in the calculated characteristics space at first;

Then separation property is estimated by descending and sorted;

Press the training set of two classification problem constructor sorters again;

Readjust training set, constructed its positive sample of deletion behind the node, and be the two-value sorter of negative sample structure next node as positive sample, other samples, up to all node-classifications completion with next sample.

7. the mobile device audio recognition method towards the agricultural product data acquisition as claimed in claim 5; It is characterized in that; This method is carried out COMPREHENSIVE CALCULATING to voice signal in identifying; The output of SVM is converted into posterior probability, and the SVM distance that it adopted comprises to the computing method of posterior probability:

The output format of SVM is y=sign (f (x)) (1)

Wherein $f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{a}_i{y}_{j}k(x_i,x)+b---\left(2\right)$

X is an input vector, and y belongs to positive sample when f (x)＞0, and y belongs to negative sample when f (x)＜0, and y belongs to the probability of positive sample suc as formula shown in (3) under f (x) value, and f (x) is the SVM distance;

$P(y=1|f)=\frac{P\left(f\right|y=1)P_1}{P\left(f\right|y=1)P_1+P\left(f\right|y=-1)P_{-1}}---\left(3\right)$

The conditional probability hypothesis of each type by the Gaussian function model description does

Positive sample: $P\left(f\right|y=1)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_1^2}}\mathrm{Exp}\frac{-{(f-u_1)}^2}{2{\mathrm{\σ}}_1^2}---\left(4\right)$

Negative sample: $P\left(f\right|y=-1)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{-1}^2}}\mathrm{Exp}\frac{-{(f-u_{-1})}^2}{2{\mathrm{\σ}}_{-1}^2}---\left(5\right)$

Then formula (3) can be expressed as: $P(y=1|f)=\frac{1}{1+\frac{p_{-1}}{p_1}\mathrm{Exp}(-\frac{1}{2{\mathrm{\σ}}^2}({(f-u_1)}^2-{(f-u_{-1})}^2))}$

$=\frac{1}{1+\mathrm{Kexp}(-\frac{1}{2{\mathrm{\σ}}^2}((u_1^2-u_{-1}^2)+2(u_{-1}-u_1)f))}$

$=\frac{1}{1+\exp (\mathrm{Af}+B)}---\left(6\right)$

A and B are estimated parameters, when the model training parameter setting, are obtained by the LibSVM model kit.

8. the mobile device audio recognition method towards the agricultural product data acquisition as claimed in claim 7 is characterized in that, said method has adopted SVM/sigmoid combined training model:

Because its classification results of symbolic representation of SVM output valve, its absolute value representation sample be to the distance of classifying face, the therefore posterior probability in order to obtain calibrating converts the output of SVM the performance of probability model with the raising model to through the sigmoid function:

Parameter A in the formula (6) and B confirm through the maximum likelihood problem:

min _A，BF(A，B)＝-∑ _i＝1(t _i(log(p _i)+(1-t _i)log(1-p _i)) (7)

P wherein _i=P (f _i),

$t_i=\left\{\begin{array}{cc}(N_{+}+1)/(N_{+}+2)& y_i=1\\ 1/(N_{-}+2)& y_i=-1\end{array}\right.---\left(8\right)$

N ₊Be positive sample number, N _-Be the negative sample number.

9. the mobile device audio recognition method towards the agricultural product data acquisition as claimed in claim 8; It is characterized in that; The training method of said SVM/sigmoid combined training model comprises: the voice of gathering are trained the HMM model through the Baum-Welch algorithm; The HMM topological structure that is adopted is for two shifting the continuous HMM model of Multidimensional and Hybrid Gaussian density functions from left to right, in training process, HMM formed training pattern jointly as a part and the SVM of training pattern; With HMM is the prime of training pattern; With this training data that is used for the SVM model training is had the changing features of the property distinguished, and with the multidimensional probability output of the HMM input vector as back level SVM model, the output of last SVM converts posterior probability into through Sigmoid function (6) formula of pressing.

Images