CN109949803A - Building service facility control method and system based on semantic instructions intelligent recognition - Google Patents

Abstract

The invention discloses the building service facility control methods based on semantic instructions intelligent recognition, method includes the following steps: acquisition instructions voice signal first, then primary character string is converted by speech recognition by instruction voice signal, low confidence character is determined from primary character string again, then the characteristic parameter of low confidence character is substituted into Hopfield neural network, obtain regular character string, signal is finally controlled accordingly according to regular text string generation, and executes corresponding function according to control signal control building service facility.This method allows user to pass through instruction voice to control the operation of various service equipments in building service facility, so that control mode is more intelligent and convenient；Correct instruction voice can also be obtained when the instruction voice pronunciation that user issues is lack of standardization simultaneously, avoided since the problem of voice causes user can not normal control building service facility.

Description

Building service facility control method and system based on semantic instructions intelligent recognition

Technical field

The present invention relates to field of intelligent control technology, in particular to based on the building service facility of semantic instructions intelligent recognition Control method, and the building service facility control system based on semantic instructions intelligent recognition.

Background technique

Building service facility is the various daily fortune for city including public service facility, service facilities Facility in the building of the offers service such as row, daily life, the operation for building service facility be unable to do without various clothes in facility The operation for equipment of being engaged in, such as crouch in the elevator in public place in mansion and the gate inhibition at residential building doorway and Private Lounge Indoor air-conditioning, illuminator etc., these service equipments operate normally the normal operation that could ensure building service facility jointly. And with the development of the city and the improvement of people's living standards, building service facility be also intended to further it is intelligent, convenient and Multifunction.

Service equipment in existing building service facility is equipped with the control panel for starting and controlling it, but services The operation of equipment be usually all realized by carrying out manual manipulation to the control panel for being provided with button, key, such as with Family hand presses floor key in elevator, the digital button of residential building access control system, air-conditioning panel and light on indoor wall Button etc. on panel can just make corresponding control system receive control instruction, realize elevator lifting, number input, temperature Degree is adjusted and light is adjusted.Such control mode can cause inconvenience to the user in some cases, such as user is no both hands Perhaps the disabled person or user's both hands to blind holds article, or can not touch because of environment obstruction, apart from the factors such as remote Situations such as to control panel.

Accordingly, it is desirable to be able to a kind of method that building service facility operation is controlled by user's sending voice is provided, so that User is not necessarily to hand manipulations, only can control building service facility operation by issuing instruction voice.For example, user can be It says " room XXX please be call " beside gate inhibition, " please open the door, door-opening password is the instruction voices such as XXXXXX ", " please lock a door "；Or The instruction language such as " air-conditioning please be open ", " air-conditioning please be close ", " please heat up X degree ", " warm X degree of begging to surrender " is said near air-conditioning panel Sound；Or the instruction voices such as " pendent lamp light please be lighten ", " bedside lamp light please be dim " are said in light nearby panels, with control Gate inhibition processed or air-conditioning or lamps and lanterns execute corresponding operation.

However, use is all in the current method for running and realizing a variety of different function by voice control device Relatively simple speech recognition device, it is lack of standardization for different accents, the pronunciation of the user that speaks and the case where pronounce indistinctly without Method is preferably coped with and is handled, so that instruction voice possibly can not be correctly validated, leads to not control facility operation.

Summary of the invention

(1) goal of the invention

To overcome above-mentioned at least one defect of the existing technology, user is built by instruction voice to control The operation of various service equipments in service facility is built, and can be obtained just when the instruction voice pronunciation of user's sending is lack of standardization Really semantic instruction, the invention discloses following technical schemes.

(2) technical solution

As the first aspect of the present invention, the invention discloses the building service facility controls based on semantic instructions intelligent recognition Method processed, comprising:

Acquisition instructions voice signal；

Primary character string is converted by speech recognition by described instruction voice signal；

Low confidence character is determined from the primary character string；

The characteristic parameter of the low confidence character is substituted into neural network, regular character string is obtained；

Signal is controlled accordingly according to the regular text string generation, and is controlled building service according to the control signal and set Apply execution corresponding function.

In a kind of possible embodiment, primary is converted by speech recognition by described instruction voice signal described Before character string, further includes:

Noise reduction and/or echo cancellation processing are carried out to collected described instruction voice signal.

In a kind of possible embodiment, primary is converted by speech recognition by described instruction voice signal described Before character string, further includes:

The vocal print of described instruction voice signal is identified by sound groove recognition technology in e；

Judge whether the vocal print that identifies is possessed of control power limit, and the case where the vocal print identified is without control authority Lower stopping executes subsequent step.

In a kind of possible embodiment, described before determining low confidence character in the primary character string, Further include:

Identify whether in described instruction voice signal or the primary character string include the voice signal or character for identifying word, And in the case where the voice signal or character of the unidentified mark word described out, stop executing subsequent step.

It is described to determine that low confidence character includes: from the primary character string in a kind of possible embodiment

Each character in the primary character string is calculated to wrap in all sample character strings that instruction voice sample database includes The quantity of the sample character string containing each character；

The quantity of the sample character string comprising each character will be calculated compared with minimum confidence threshold, and by quantity Character lower than the minimum confidence threshold is determined as low confidence character.

In a kind of possible embodiment, substituted into neural network in the characteristic parameter by the low confidence character Before obtaining regular character string, further includes:

High confidence character is determined from the primary character string；

Determine the high confidence character of forerunner and/or subsequent high confidence character of the low confidence character；Also,

Obtaining regular character string in the characteristic parameter substitution neural network by the low confidence character includes:

The subsequent character collection and/or the subsequent height of the high confidence character of the forerunner are determined from instruction voice sample database Forerunner's character set of confidence character；

Acoustic vector training according to character some or all of in the subsequent character collection and/or forerunner's character set Neural network；

The corresponding acoustic vector of the low confidence character is substituted into the neural network and carries out Regularization identification, with from corresponding The subsequent character collection and/or forerunner's character set in determine substitute character, and replace the low confidence character obtain it is regular Character string.

It is described to determine that high confidence character includes: from the primary character string in a kind of possible embodiment

All characters for being not determined to low confidence character in the primary character string are determined as high confidence character； Or,

The quantity of the sample character string comprising each character will be calculated compared with maximum confidence threshold, and by quantity Character higher than the maximum confidence threshold is determined as high confidence character.

It is described that the high confidence character of the forerunner is determined from instruction voice sample database in a kind of possible embodiment Subsequent character collection and/or forerunner's character set of the subsequent high confidence character include:

All subsequent characters of the high confidence character of the forerunner are determined in described instruction speech samples library, and/or really Make all forerunner's characters of the subsequent high confidence character；

Count the frequency that the subsequent character and/or forerunner's character occur in all sample character strings；

The highest multiple subsequent characters of the frequency and/or institute are determined according to frequency collating or according to default frequency threshold value Forerunner's character is stated, and separately constitutes subsequent character collection and/or forerunner's character set.

In a kind of possible embodiment, the frequency is determined according to frequency collating or according to default frequency threshold value described Before highest multiple subsequent characters and forerunner's character, first to identical in the subsequent character and forerunner's character The frequency of occurrence of character is summed, using the frequency data after summing as the frequency of occurrence of character.

In a kind of possible embodiment, described according in the subsequent character collection and/or forerunner's character set Some or all of the acoustic vector training neural network of character when, only choose respective symbols in described instruction speech samples libraries Partial sound vector trains neural network.

In a kind of possible embodiment, it is described according to the regular text string generation control accordingly signal it Afterwards, further includes:

Collected described instruction voice signal is stored in described instruction speech samples library.

It is described to control signal accordingly according to the regular text string generation and include: in a kind of possible embodiment

Semantic instructions are converted by synonym mapping by the regular character string；

Control signal is generated according to the semantic instructions, and controls building service facility according to the control signal and executes phase Answer function.

As a second aspect of the invention, the invention also discloses the building service facilities based on semantic instructions intelligent recognition Control system, comprising:

Signal acquisition module is used for acquisition instructions voice signal；

Primary character generation module, for converting primary character string by speech recognition for described instruction voice signal；

Low confidence character determining module, for determining low confidence character from the primary character string；

Regular character generation module is advised for substituting into the characteristic parameter of the low confidence character in neural network Whole character string；

Signal generation module is controlled, for controlling signal accordingly according to the regular text string generation, and according to described It controls signal control building service facility and executes corresponding function.

In a kind of possible embodiment, the control system further include:

Noise processed module, for converting primary character string by speech recognition for described instruction voice signal described Before, noise reduction is carried out to collected described instruction voice signal and/or echo cancellation is handled.

In a kind of possible embodiment, the control system further include:

Vocal print judgment module, for converting primary character string by speech recognition for described instruction voice signal described Before, the vocal print that described instruction voice signal is identified by sound groove recognition technology in e, judges whether the vocal print identified has control Permission processed, and stop executing subsequent step in the case where the vocal print identified does not have control authority.

In a kind of possible embodiment, the control system further include:

Identification module is identified, for before determining low confidence character in the primary character string, identifying institute described State in instruction voice signal or the primary character string whether voice signal or character comprising mark word, and it is unidentified go out institute In the case where the voice signal or character of stating mark word, stop executing subsequent step.

In a kind of possible embodiment, the low confidence character determining module includes:

There is quantity statistics unit, includes for calculating in the primary character string each character in instruction voice sample database All sample character strings in, the quantity of the sample character string comprising each character；

Low confidence character determination unit, for will calculate include each character the sample character string quantity with most Small confidence threshold compares, and the character by quantity lower than the minimum confidence threshold is determined as low confidence character.

In a kind of possible embodiment, the control system further include:

High confidence character determining module, for being substituted into neural network in the characteristic parameter by the low confidence character Before obtaining regular character string, high confidence character is determined from the primary character string, determines the low confidence character The high confidence character of forerunner and/or subsequent high confidence character；Also,

The regular character generation module includes:

Character set determination unit, for determining the subsequent word of the high confidence character of the forerunner from instruction voice sample database Forerunner's character set of symbol collection and/or the subsequent high confidence character；

Neural metwork training unit, for according in the subsequent character collection and/or forerunner's character set part or The acoustic vector training neural network of alphabet；

Regular character generation unit, for will the low confidence character corresponding acoustic vector substitution neural network into The identification of professional etiquette integralization to determine substitute character from the corresponding subsequent character collection and/or forerunner's character set, and replaces institute It states low confidence character and obtains regular character string.

In a kind of possible embodiment, the high confidence character determining module includes:

First character determination unit, for by all characters for being not determined to low confidence character in the primary character string It is determined as high confidence character；And/or

Second character determination unit, for the quantity and maximum of the sample character string comprising each character will to be calculated Confidence threshold compares, and the character that quantity is higher than the maximum confidence threshold is determined as high confidence character.

In a kind of possible embodiment, the character set determination unit includes:

Third character determines subelement, for determining the high confidence character of the forerunner in described instruction speech samples library All subsequent characters, and/or determine all forerunner's characters of the subsequent high confidence character；

Frequency statistics subelement, for counting the subsequent character and/or forerunner's character in all sample words The frequency occurred in symbol string；

Character set synthesizing subunit, for determining that the frequency is highest more according to frequency collating or according to default frequency threshold value A subsequent character and/or forerunner's character, and separately constitute subsequent character collection and/or forerunner's character set.

In a kind of possible embodiment, the frequency statistics subelement be also used to it is described according to frequency collating or according to Before determining the highest multiple subsequent characters of the frequency and forerunner's character according to default frequency threshold value, first to described subsequent The frequency of occurrence of identical characters is summed in character and forerunner's character, going out using the frequency data after summing as character The existing frequency.

In a kind of possible embodiment, the neural metwork training unit is only chosen in described instruction speech samples library The partial sound vector training neural network of respective symbols.

In a kind of possible embodiment, the control system further include:

Sample memory module, for it is described control signal accordingly according to the regular text string generation after, will Collected described instruction voice signal is stored in described instruction speech samples library.

In a kind of possible embodiment, the control signal generation module includes:

Semantic mapping unit, for converting semantic instructions by synonym mapping for the regular character string；

Signal generation unit for generating control signal according to the semantic instructions, and is controlled according to the control signal It builds service facility and executes corresponding function.

(3) beneficial effect

Building service facility control method and system disclosed by the invention based on semantic instructions intelligent recognition has as follows The utility model has the advantages that

1, allow user to pass through instruction voice to control the operation of various service equipments in building service facility, be not necessarily to User carries out manual manipulation control panel with hand, so that control mode is more intelligent and convenient；It can also speak simultaneously The instruction voice that user issues has an accent, it is lack of standardization to pronounce and when pronouncing indistinctly, instruction voice is correctly associated and is known Not, it to obtain correct instruction voice, avoids since the problem of voice causes user can not normal control building service facility.

2, it is spoken by before converting primary character string for instruction voice signal, first passing through sound groove recognition technology in e judgement User whether be possessed of control power limit, avoid since television set makes a sound by control system acquisition and then generates maloperation It happens, while preventing from just finding that user does not have permission and terminates and carries out subsequent step in the control system operation later period, Avoid the calculation resources of waste control system.

3, by before determining low confidence character in primary character string, or instruction voice signal is being passed through into voice Identification is converted into before primary character string, and the primary character string first converted to instruction voice signal or instruction voice signal is marked Know word identification, confirms that instruction voice signal is the control command to control system sending rather than other human conversations, thus It can be avoided and non-instruction voice is identified, avoid the calculation resources of waste control system.

4, low confidence character is determined by way of obtaining character frequency of occurrence and setting threshold value in sample database, into And semantic wrong character is navigated to, positioning accuracy is higher, and locating speed is very fast.

5, the candidate characters of low confidence character are obtained using the high confidence character before and after low confidence character, recycles nerve net Network obtains semantic correct character from candidate characters, accurately and rapidly will can be identified as semanteme because of accent the problems such as Wrong character replaces with semantic correct character, facilitates the user from different regions and controls to building service facility System, and the accuracy replaced is higher.

6, by filtering out the subsequent character and forerunner's character and only by the subsequent character of high frequency time and forerunner's word of high frequency time Accord with training neural network, and only pick out part acoustic vector substitute into neural network be trained, improve operation efficiency, Save system reaction time.

7, it is stored in instruction voice sample database by the instruction voice signal that will correctly identify, after capable of being convenient for User issues identification process when same phonetic order.

8, the relatively small number of semanteme of total quantity is converted by the regular character string of wide variety in advance using control system to refer to It enables, then regeneration controls signal and issues controlled device, so that the communication between control system and controlled device is more succinct, And reduce the requirement to controlled device data-handling capacity.

Detailed description of the invention

It is exemplary below with reference to the embodiment of attached drawing description, it is intended to for the explanation and illustration present invention, and cannot manage Solution is the limitation to protection scope of the present invention.

Fig. 1 is the building service facility control method first embodiment disclosed by the invention based on semantic instructions intelligent recognition Flow diagram.

Fig. 2 is the building service facility control system first embodiment disclosed by the invention based on semantic instructions intelligent recognition Structural block diagram.

Specific embodiment

To keep the purposes, technical schemes and advantages of the invention implemented clearer, below in conjunction in the embodiment of the present invention Attached drawing, technical solution in the embodiment of the present invention is further described in more detail.

It should be understood that in the accompanying drawings, from beginning to end same or similar label indicate same or similar element or Element with the same or similar functions.Described embodiments are some of the embodiments of the present invention, rather than whole implementation Example, in the absence of conflict, the features in the embodiments and the embodiments of the present application can be combined with each other.Based in the present invention Embodiment, every other embodiment obtained by those of ordinary skill in the art without making creative efforts, It shall fall within the protection scope of the present invention.

Herein, " first ", " second " etc. are only used for mutual differentiation, rather than indicate their significance level and sequence Deng.

The division of module, unit or assembly herein is only a kind of division of logic function, in actual implementation may be used To there is other division modes, such as multiple modules and/or unit can be combined or are integrated in another system.As separation The module of part description, unit, component are also possible to indiscrete may be physically separated.It is shown as a unit Component can be physical unit, may not be physical unit, it can is located at a specific place, may be distributed over grid In unit.Therefore some or all of units can be selected to realize the scheme of embodiment according to actual needs.

Below with reference to the building service facility control disclosed in Fig. 1 the present invention is described in detail based on semantic instructions intelligent recognition Method first embodiment.The present embodiment is mainly used in building service facility, allows user to pass through instruction voice to control The operation for building various service equipments in service facility carries out manual manipulation control panel with hand without user, so that controlling party Formula is more intelligent and convenient；Simultaneously can also speak user sending instruction voice have an accent, pronounce it is lack of standardization and When pronouncing indistinctly, instruction voice is correctly associated and identified, to obtain correct instruction voice, is avoided due to voice Problem causes user can not normal control building service facility.

As shown in Figure 1, the disclosed building service facility control method of the present embodiment includes the following steps:

Step 100, acquisition instructions voice signal.

User, can be by issuing instruction voice come real when needing to control the start and stop of building service facility, working order Now control.And for building service facility and its interior service equipment, then need to collect the instruction of user's sending first Voice, and obtain instruction voice signal.

It is understood that the equipment of acquisition instructions voice signal is each sky being installed in building service facility in advance Between, place, channel or build at the entrance of service facility etc. in environment, such as elevator, in bedroom etc..Under normal conditions, it uses Family issues instruction voice by speaking.Below with user in bedroom in order to control air-conditioning opening and closing and saying " please shut sky For tune " instruction voice.

Step 200, primary character string is converted by speech recognition by instruction voice signal.

After the instruction voice signal for obtaining user's sending, the speech recognition compared with based on is first carried out to instruction voice signal Technology is identified, is converted textual form for instruction voice signal, is obtained text-string.The identification of basic speech identification Fast speed, but may instruction voice that user issues there are accent, pronounce indistinctly etc. due to and cause to identify Semanteme expressed by text-string and user are practical to be want through the semantic not exactly the same of instruction voice signal representation, therefore will It converts obtained text instruction and is known as primary character string.For example, " please shut air-conditioning " that the band that user issues is had an accent identifies For the text-string of " upper control please be manage ".

Step 300, low confidence character is determined from primary character string.

Low confidence character refers to the lower character of frequency of occurrence in each speech samples of instruction voice sample database.It obtains just Grade character string after, need from determined in primary character string which character be in step 200 speech recognition process due to accent, The problems such as pronouncing indistinctly and semantic wrong character, the wrong characters of these semantemes for determining usually are exactly low confidence character.

It determines that the mode of semantic wrong character can be to be screened by pre-set instruction voice sample database.Instruction The instruction voice sample of certain amount (not necessarily all) has been included in speech samples library in advance, and with each section of instruction language Sound sample is corresponding, sample character string by accurately identifying, can pass through the confidence for judging each character in primary character string It spends to identify some or certain several characters in semantic wrong character, such as primary character string in instruction voice sample database All sample character strings in the frequency that occurs it is extremely low or even all sample character strings do not have comprising (confidence level is at this time Zero), then these characters are likely to semantic wrong character, i.e., low confidence character.Wherein, the character that confidence level is zero, must For semantic wrong character.

Specifically, in " upper control please be manage " character string, " pipe " and " control " all sample words in instruction voice sample database The frequency occurred in symbol string is extremely low, therefore " manages " and " control " is low confidence character, can be considered as semantic wrong character.If through Primary character string after speech recognition conversion is " air-conditioning is shut in celebrating ", and in instruction voice sample database in any sample character string Do not occur " celebrating ", the confidence level therefore " celebrated " is zero, belongs to low confidence character, can be determined as semantic wrong character.

Step 400, the characteristic parameter of low confidence character is substituted into neural network, obtains regular character string.

If primary character string includes low confidence character, need low confidence character carrying out error correction by neural network, So that primary character string is converted into correct regular character string of expressing the meaning.It is understood that regular character string is not necessarily meant to refer to The sample character string for having been subjected to and accurately identifying in speech samples library is enabled, as long as building service facility correctly can be identified and be expressed the meaning Correct character string.

The characteristic parameter of above-mentioned low confidence character can be text character and be converted into the later feature vector of image, such as The matrix of 256*256, wherein different for the part of ideographic character and the value of blank parts；Characteristic parameter is also possible at it The corresponding feature vector of low confidence character in preceding collected instruction voice signal.

Neural network (Neural Networks, NN) is extensive by a large amount of, simple processing unit (referred to as neuron) Ground interconnects and the complex networks system of formation.Neural network has the function of that self-learning function, connection entropy and high speed are found The ability of optimization solution can directly will likely include the primary word of wrong identification semanteme after getting well neural metwork training Symbol string (such as " please manage control ") inputs neural network, and directly obtain the regular character string after semantic Regularization (such as " air-conditioning please be shut "), wrong part is identified during speech recognition conversion to correct in step 200 with this, so that obtaining Character string and user's sheet expect the instruction issued and be consistent.

Specifically, a variety of correct templates are as nerve net corresponding to the instruction voice signal that user can may be issued The input of network, and provide corresponding desired output.When in input action to neural network, by the reality output of neural network with Desired output is compared, if the reality output of neural network is not inconsistent with desired output, adjust neural network weight and Bias.Then it repeats the above process, until the reality output of neural network is close to desired output, at this moment it is believed that nerve net Network has trained completion.Low confidence character is substituted into the neural network trained and completed, correct output can be obtained.

Such as when realizing image recognition, only formerly many different image templates and the corresponding result that should be identified are inputted Neural network, network will be by self-learning functions, and slowly association identifies similar image.

Step 500, signal is controlled accordingly according to regular text string generation, and set according to control signal control building service Apply execution corresponding function.

After obtaining regular character string, it can be ensured that building service facility can be identified smoothly and correctly execute user instruction Voice is intended to the function of executing, and realizes the service that user wants.Such as after user says the voice of " please shut air-conditioning ", Control system passes through above-mentioned steps automatically and obtains the regular text-string of " please shut air-conditioning ", and according to the character string to air-conditioning Instruction is issued, air-conditioning is closed.

In one embodiment, primary character is converted by speech recognition by instruction voice signal in step 200 Before string, further includes:

Step 010, noise reduction is carried out to collected instruction voice signal and/or echo cancellation is handled.

It, may during user issues the instruction voice user of control system acquisition simultaneously and issues instruction voice signal There are other sound sources to generate noise jamming, such as sound, the sound of washing machine of TV etc., control system can be improved by noise reduction The signal-to-noise ratio of collected instruction voice signal, and then improve the accuracy of identification of voice signal.

Indoors under environment, the instruction voice that user issues is likely encountered indoor wall reflection and generates echo, and echo is anti- It is mapped to the acquisition that can also interfere with voice signal at the collection point of control system, by echo cancellation technology, equally can be improved The signal-to-noise ratio of the collected instruction voice signal of control system, and then improve the accuracy of identification of voice signal.

In one embodiment, primary character is converted by speech recognition by instruction voice signal in step 200 Before string, further includes:

Step 020, the vocal print of instruction voice signal is identified by sound groove recognition technology in e.

Step 030, judge whether the vocal print identified is possessed of control power limit, and do not have control in the vocal print identified Stop executing subsequent step in the case where limit.

Control system is led to after the instruction voice signal for collecting user's sending by collected voice signal It crosses speech recognition to be converted into before primary character string, whether the people that control system first has to confirmation sending instruction voice signal has the right Limit removes control building service facility.For example, the operations such as the opening and closing of room conditioning, temperature adjusting are only limitted to owner's family of three and gather around Have, other people, biology or the equipment that can issue voice such as guest, parrot, television set do not have permission, otherwise assume television set In have the lines of one " air-conditioning please be close ", then control system will do it acquisition and identify and control air-conditioning closing, generate maloperation Phenomenon.In addition, the confirmation of permission needed in the step of being arranged in front as far as possible, prevent from just finding do not have after have passed through system operations It is terminated for permission and carries out subsequent step, waste system operations resource.

Vocal print is the sound wave spectrum for the carrying verbal information that electricity consumption acoustic instrument is shown, and sound groove recognition technology in e is a kind of logical Cross the technology that sound differentiates speaker's identity.By collected voice signal by speech recognition be converted into primary character string it Before, first pass through sound groove recognition technology in e identify the vocal print of instruction voice signal whether belong to building service facility be possessed of control power The vocal print of limit.If so, continuing subsequent to be converted into primary character string；If it is not, then stopping executing subsequent step simultaneously Terminate the implementation of this control method.

Only limit is usually to be pre-set in control system, such as owner's family of three allows control system to record in advance for operation Enter and store enough vocal print samples, control system is made to can recognize that the vocal print of three people.It is understood that actually making With in the process, by sound groove recognition technology in e identify speaker whether have permission usually to collected instruction voice signal into It is carried out after row noise reduction and echo cancellation processing, to improve accuracy of identification.It should be noted that not needing to know under some occasions Other vocal print, such as in the elevator of mansion, owner can choose target elevator floor by voice, just no setting is required at this time The step of Application on Voiceprint Recognition.Alternatively, only lift attendant have elevator locked permission, then identify whether it is lift attendant Issue instruction voice signal.

In one embodiment, it is also wrapped before determining low confidence character in primary character string in step 300 It includes:

Step 030, identify in instruction voice signal or primary character string whether include the voice signal or character for identifying word, And in the case where the voice signal or character of the unidentified word of mark out, stop executing subsequent step.

Mark word is that control system can be made to confirm whether collected voice signal is to control building service facility And issue.

It is issued if it is in order to control building service facility, then user needs that mark is inside added when speaking Word, so that control system learns that this voice comprising mark word is issued to control building service facility.For example, with " please shut air-conditioning " is said at family, then word " can will be asked " as mark word, control system is after having identified " asking " word, just meeting Continue to execute subsequent step；If control system finds not including mark word in collected voice, such as collected voice is " strategic partnership " for including in dialogue between two people, then control system stops executing subsequent step and terminates this The implementation of control method.

Whether control system judges in voice signal there are many modes comprising mark word.

The first, it is straight to instruction voice signal before converting primary character string for instruction voice signal in step 200 Tap into the identification of line identifier word, that is to say, that whether comprising knowing in decision instruction voice signal in a manner of identifying audio signal The voice signal of other word.

Second, after converting primary character string for instruction voice signal in step 200, and in step 300 from Before determining low confidence character in primary character string, word is directly identified to the primary character string of instruction voice signal conversion Identification, that is to say, that whether judged in primary character string in a manner of identifying text character comprising mark word, and then judge to refer to Whether enable in voice signal includes the voice signal for identifying word.

It is understood that for above-mentioned a variety of mark word judgment modes, mark word may more than just one, such as can be with It is also mark word by " trouble ".Also, the audio signal and letter signal for identifying word are preparatory input systems, are needing to use When directly extract and voice signal to be measured or text-string identified.Instruction voice signal is directly carried out The identification of mark word carries out usually after having executed noise reduction and echo cancellation, to improve the accuracy of identification of mark word.In addition, mark The identification for knowing word can carry out after Application on Voiceprint Recognition, because Application on Voiceprint Recognition maximum probability can screen out more voice signals. In the communal facilitys such as elevator, the identification for being not provided with mark word can choose, avoid the arithmetic speed for reducing communal facility.

In one embodiment, determine that low confidence character includes: in step 300 from primary character string

Step 310, all sample character strings that each character includes in instruction voice sample database in primary character string are calculated In, the quantity of the sample character string comprising each character.

Instruction voice sample database includes a certain number of instruction voice samples, such as includes 10000 sections of the accurate knowledge of process Other speech samples character string.Each speech samples character string is corresponding with each section of instruction voice sample.Primary character string is upper State " control please be manage " that embodiment refers to, primary character string includes five characters at this time: " asking ", " pipe ", "upper", " control ", " tune ".And in 10000 sections of the instruction voice sample that instruction voice sample database is included, the speech samples character including " asking " character String is 5000, and the speech samples character string including " pipe " character is 500, including "upper" " the speech samples character string of character It is 2500, the speech samples character string including " control " character is 800, and the speech samples character string including " tune " character is 2000.

Step 320, by the quantity of the sample character string of the calculating comprising each character compared with minimum confidence threshold, and will The character that quantity is lower than minimum confidence threshold is determined as low confidence character.

Pre-set minimum confidence threshold is 1000, it is known that the appearance quantity of " pipe " and " control " character is 500 and 800 A, therefore respectively less than minimum confidence threshold " manages " and " control " is confirmed as low confidence character.

Low confidence character is determined by way of obtaining character frequency of occurrence and setting threshold value in sample database, in turn Semantic wrong character is navigated to, positioning accuracy is higher, and locating speed is very fast.

In one embodiment, it is obtained in the characteristic parameter substitution neural network by low confidence character in step 400 Before regular character string, further includes:

Step 330, high confidence character is determined from primary character string.

When determining low confidence character, high confidence character is also determined.High confidence character refers in instruction voice sample The higher character of frequency of occurrence in each speech samples in library.The higher character of frequency of occurrence can be it is opposite, such as in primary In each character of character string, all characters for being not determined to low confidence character can be determined that high confidence character, That is all lower characters of frequency of occurrence can be considered as the higher character of frequency of occurrence, and " asking ", "upper", " tune " three characters are high confidence character.The higher character of frequency of occurrence is also possible to absolute, such as can preset One for judge character whether be high confidence character maximum confidence threshold, similar to the mode for determining low confidence character, By by frequency of occurrence judges which is that height is set more than maximum confidence threshold in command language sample database in primary character string Believe character.Specifically, for example maximum confidence threshold is arranged to 2000, then " asking ", "upper", " tune " three characters are high confidence Character.

It is understood that high confidence character and low confidence character can be while determining, be also possible to according to What step was successively determined.

Step 340, the high confidence character of the forerunner of low confidence character and/or subsequent high confidence character are determined.

The high confidence character of forerunner refers to one high confidence character nearest before being located at low confidence character, subsequent high confidence Character refers to nearest one high confidence character after low confidence character.For low confidence character " pipe " and " control ", The high confidence character of the forerunner of " pipe " is " asking ", and subsequent high confidence character is "upper"；The forerunner of " control " is high, and confidence character is "upper", after After high confidence character be " tune ".

It is understood that low confidence character may only have the high confidence character of corresponding forerunner, such as primary character string is most The latter character does not have subsequent character, only forerunner's character；Low confidence character may also only have corresponding subsequent high confidence character, Such as primary character string first character does not have forerunner's character, only subsequent character.It could also be possible that having before low confidence character Multiple characters are still all not determined to have multiple characters to be still all not determined to after high confidence character or low confidence character High confidence character.

Further, the characteristic parameter of low confidence character is substituted into neural network in step 400 and obtains regular character string Include:

Step 410, the subsequent character collection and/or subsequent height of the high confidence character of forerunner are determined from instruction voice sample database Forerunner's character set of confidence character.

Subsequent character collection refers in the instruction voice sample in instruction voice sample database comprising the high confidence character of forerunner, adjacent The set of character composition after the high confidence character of the forerunner.

For example, the high confidence character of the forerunner of " pipe " is " asking ", and in instruction voice sample database include the speech samples of " asking " Character string is 5000, such as " please turn on pendent lamp ", " please shut TV " etc., then in above-mentioned two speech samples character string " beating " and "Off" are exactly the character being adjacent to after " asking ", and all in 5000 speech samples character strings comprising " asking " The set { "ON", "Off", " beating ", " liter ", " drop " ... } of this kind of character is exactly the subsequent character of the high confidence character " asking " of forerunner Collection.

Similarly, the subsequent high confidence character of " pipe " is "upper", includes the speech samples word of "upper" in instruction voice sample database In symbol string, such as " TV is shut after ten minutes ", " please draw the curtain together " etc., then in above-mentioned two speech samples character string "Off" and " drawing " are exactly the character before being adjacent to "upper", and this kind of word all in the speech samples character string comprising "upper" The set { "Off", " liter ", " conjunction ", " closing " ... } of symbol be exactly subsequent high confidence character " Guan " forerunner's character set.

In the subsequent character concentration of the high confidence character " asking " of the forerunner of " pipe " and the forerunner of subsequent high confidence character "upper" In character set, the correct character of expressing the meaning of low confidence character " pipe " must be will appear.

It is understood that the subsequent character collection and subsequent high confidence word of the high confidence character of forerunner of low confidence character " control " The method of determination of forerunner's character set of symbol can be and all determine with above-mentioned " pipe ", and when there is multiple low confidence characters The corresponding subsequent character collection and forerunner's character set of all low confidence characters and then each low confidence character is carried out respectively subsequent Neural metwork training is also possible in the corresponding subsequent character collection and forerunner's character set for determining a low confidence character and instructs After having practiced neural network and having obtained the substitute character of the low confidence character, then determine the corresponding subsequent of next low confidence character Character set and forerunner's character set.

It should be noted that if low confidence character does not have the high confidence character of corresponding forerunner, just without corresponding subsequent yet Character set, forerunner's character set is similarly.

Step 420, the acoustic vector training according to character some or all of in subsequent character collection and/or forerunner's character set Neural network.

Neural network is mainly used for the recognition mode in complicated and diversified data, target identification or voice in such as image In signal identification.Using correct " training " neural network of magnanimity training data in order to provide " live " sampling that will be analyzed When can complete their identification mission.Training neural network is with regard to itself being a complicated task and relating to the use of for example Iteration or adaptive algorithm set quantity of parameters.

For character present in each instruction voice sample database, in instruction voice sample database all the sound containing character to Amount, that is, the characteristic parameter of character.And it is all that may substitute phase that forerunner's character set and subsequent character, which concentrate include, at this time The candidate characters of low confidence character are answered, therefore forerunner's character set and the set of subsequent character collection are referred to as candidate characters collection. By taking " pipe " as an example, forerunner's character set and subsequent character that " pipe " is extracted from instruction voice sample database concentrate the sound of each character Then acoustic vector is substituted into neural network by vector, to be trained to neural network.

Can specifically discrete type two-value network: Hopfield neural metwork training mode be used:

1, acoustic vector is first subjected to binary-coding, is stored respectively with two values matrix.The mode of binaryzation can borrow Two-value halftoning method in mirror bianry image carries out binary conversion treatment to digital audio and video signals.At above-mentioned binarization The audio of reason, continued time domain waveform are converted into only two-value square wave, and the audio after binaryzation only includes 0 and 1 two kind of audio Signal, or -1 and 1 two kind of audio signal.If the acoustic vector stored in instruction voice sample database is analogue audio frequency, need First it is converted into digital audio.

2, then by the bipolar value of obtained two values matrix, and it is converted into column vector, obtains sample vector group, it is low at this time to set Letter character can also be converted into column vector simultaneously, obtain vector to be measured.Sample vector group is generated into weight using apposition rule Matrix, weight indicate the bonding strength between neuron, and the mode that neural network is remembered is stored by weight matrix. Hopfield neural network model is a kind of Recognition with Recurrent Neural Network, has feedback link, all neuron elements from input is output to It is the same, is connected with each other between them.Each neuron passes through connection weight and receives every other neuron output instead The information that feedback comes, mesh are the control in order to allow the output of each neuron that can receive every other neuron output, thus Each neuron is set mutually to restrict.The weight of each neuron is 0 at this time.Then Hamming (Hamming) distance is set.

3, sample vector group is inputted into neural network, and is learnt using synchronous working mode, obtained after training Weight matrix, neural network has trained at this time.Synchronous working mode, also referred to as concurrent working mode, refer to a period of time in office It carves, the state of all or part of neuron changes simultaneously.

4, vector to be measured is inputted into trained neural network, and by the iteration of certain number, network reaches energy level Dot exports corresponding acoustic vector as a result, being then then converted to character style, obtains respective symbols.

It is understood that when extracting the acoustic vector of character, it can be by all subsequent character collection and forerunner's character set In acoustic vector all extract and substitute into neural network, can also only extract partial character acoustic vector and substitute into nerve Network, to reduce operand and operation time.It should be noted that needing to carry out nerve net respectively to different low confidence characters The training of network prevents from obscuring.

It should be noted that if low confidence character does not have corresponding subsequent character collection, corresponding forerunner's character need to be only brought into Collection, the case where without corresponding forerunner's character set similarly.

Step 430, the corresponding acoustic vector of low confidence character is substituted into neural network and carries out Regularization identification, with from corresponding Subsequent character collection and/or forerunner's character set in determine substitute character, and replace low confidence character and obtain regular character string.

After the corresponding neural metwork training of low confidence character " pipe " is completed, " pipe " is extracted from instruction voice sample database Acoustic vector, the acoustic vector of " pipe " is substituted into neural network, the substitute character in corresponding candidate character set can be obtained "Off".Similarly, the acoustic vector of " control " is substituted into neural network, the substitute character in corresponding candidate character set can be obtained " sky "."Off" and " sky " are replaced into " pipe " and " control " in primary character string respectively, realizing " please manage upper control for primary character string Adjust " it converts and obtains regular character string " air-conditioning please be shut ".

It is understood that the extraction of the acoustic vector of " pipe " can be after neural metwork training completes, it can also be It is extracted simultaneously when extracting the acoustic vector of character in subsequent character collection and forerunner's character set.

The candidate characters of low confidence character are obtained using the high confidence character before and after low confidence character, recycle neural network Semantic correct character is obtained from candidate characters, accurately and rapidly will can be identified as semanteme because of accent the problems such as has Character accidentally replaces with semantic correct character, facilitates the user from different regions and controls building service facility, And the accuracy replaced is higher.

In one embodiment, determined from primary character string in step 330 high confidence character using it is following wherein A kind of mode:

The characters for being not determined to low confidence character all in primary character string are determined as high confidence word by the first Symbol.Under such mode, the character in primary character string is only possible to be low confidence character or high confidence character, and all low confidences The high confidence character of the forerunner of character is the previous character of low confidence character, and subsequent high confidence character is the latter of low confidence character A character.

Second, by the quantity of the sample character string of the calculating comprising each character compared with maximum confidence threshold, and will The character that quantity is higher than maximum confidence threshold is determined as high confidence character.Under such mode, the character in primary character string may It is low confidence character, it is also possible to high confidence character, it is also possible to neither the middle word of low confidence character nor high confidence character Symbol.And high confidence character also might not be separated with intermediate character between possibility with low confidence adjacency.

In one embodiment, in step 410 after determining the high confidence character of forerunner in instruction voice sample database Forerunner's character set after character set and subsequent high confidence character includes:

Step 411, all subsequent characters of the high confidence character of forerunner are determined in instruction voice sample database, and/or really Make all forerunner's characters of subsequent high confidence character.

For example, all subsequent characters of the high confidence character " asking " of the forerunner of low confidence character " pipe " are as follows: "ON", "Off", " beating ", " liter ", " drop " ... }, and " pipe " subsequent high confidence character "upper" all forerunner's characters are as follows: "Off", " liter ", " conjunction ", " closing " ... }.Low confidence character " control " can similarly obtain.It is understood that having formd corresponding subsequent word at this time Symbol collection and forerunner's character set, but subsequent character collection at this time and forerunner's character set are all not by screening.

It should be noted that if it is adjacent between low confidence character and the high confidence character of forerunner/subsequent high confidence character, not Between when being separated with intermediate character, then the subsequent character of the high confidence character of forerunner is also selected from the character of the high confidence adjacency of forerunner It takes, forerunner's character of subsequent high confidence character is also chosen from the character of subsequent high confidence adjacency.If low confidence character When adjoining has intermediate character between the high confidence character of forerunner/subsequent high confidence character, then the high confidence character of forerunner is being determined When subsequent character/subsequent high confidence character forerunner's character, need according to low confidence character and the high confidence character of forerunner/subsequent height The number of characters that is spaced between confidence character determines.Such as it is spaced there are two between low confidence character and the high confidence character of forerunner Between character, it is determined that subsequent character be also required between the high confidence character of forerunner be spaced two characters.

Step 412, the frequency that subsequent character and/or forerunner's character occur in all sample character strings is counted.

In the corresponding subsequent character of " pipe ", frequency that "ON", "Off", " beating ", " liter ", " drop " occur in each sample character string Secondary is respectively 4100,3900,3400,2600,2000, and similarly, other subsequent characters have the frequency of occurrence of oneself.And " pipe " Corresponding forerunner's character is same, the frequency that "Off", " liter ", " conjunction ", " closing " occur in each sample character string is respectively 4500, 3700,3500,2400, similarly, other forerunner's characters have the frequency of occurrence of oneself.It is understood that the subsequent word of " control " Symbol and forerunner's character also have respective frequency of occurrence.

If low confidence character only has corresponding subsequent character, the frequency of subsequent character is only counted, forerunner's character is similarly.

Step 413, according to frequency collating or according to default frequency threshold value determine the highest multiple subsequent characters of the frequency and/ Or forerunner's character, and separately constitute subsequent character collection and/or forerunner's character set.

After obtaining the frequency of occurrence of each subsequent character and forerunner's character, it can be selected by following one way in which High frequency Chinese characters.The first, by the frequency of occurrence of the corresponding subsequent character of same low confidence character and forerunner's character according to from more to It is ranked up less, and therefrom selects the highest multiple characters of the frequency, such as select highest two characters of the frequency as high frequency word Symbol.Second, according to pre-set frequency threshold value, frequency of occurrence is filtered out respectively from subsequent character and forerunner's character and is higher than The character filtered out is determined as high frequency Chinese characters by the character of frequency threshold value.After obtaining forerunner's high frequency Chinese characters and subsequent high frequency Chinese characters, It is forerunner's character set by the character set that corresponding forerunner's high frequency Chinese characters of same low confidence character form, by same low confidence character The character set of corresponding subsequent high frequency Chinese characters composition be subsequent character collection.

If low confidence character only has corresponding subsequent character, subsequent character is only formed, forerunner's character is similarly.

By filtering out the subsequent character and forerunner's character of high frequency time, and only by the subsequent character of high frequency time and forerunner's character Training neural network, improves operation efficiency, saves system reaction time.

In one embodiment, after executing the step 412 and before executing step 413, first to subsequent character and The frequency of occurrence of identical characters is summed in forerunner's character, using the frequency data after summing as the frequency of occurrence of character.

If low confidence character has corresponding subsequent character and forerunner's character simultaneously, subsequent character and forerunner are being counted After the frequency that character occurs in all sample character strings, first to the appearance frequency of identical characters in subsequent character and forerunner's character It is secondary to sum, such as in the corresponding subsequent character and forerunner's character of " pipe ", there are an identical characters "Off" and " liter ", therefore by "Off" The frequency of occurrence of " liter " is summed.After summation, frequency of occurrence of two identical characters using total frequency data as character, That is no matter "Off" is in subsequent character or in forerunner's character, the frequency is all 3900+4500=8400, " liter " no matter For subsequent character still in forerunner's character, the frequency is all 2600+3700=6300.When determining high frequency Chinese characters, also according to summation The frequency afterwards carries out ranking and is compared with frequency threshold value.

It should be noted that after summation, if executing the high frequency Chinese characters determined in step 413 is entirely subsequent word The character concentrated, or the character of entirely forerunner's character set are accorded with, then when subsequent execution step 420, is belonged to according to subsequent character The case where acoustic vector training neural network of partial character in collection and forerunner's character set.

In one embodiment, at step 420 according to subsequent character collection and/or part in forerunner's character set or complete When the acoustic vector training neural network of portion's character, the partial sound vector instruction of respective symbols in instruction voice sample database is only chosen Practice neural network.

Subsequent character collection and forerunner's character set composition candidate characters concentrate, the character for no matter including how many, it is therein The acoustic vector of each character (namely candidate characters) might have very much, such as the frequency of "Off" and " liter " is all higher, because This can therefrom choose one at this time correspondingly, the acoustic vector of "Off" in instruction voice sample database and " liters " also can be very much Partial acoustic vector substitutes into neural network and is trained, to reduce operand.

In one embodiment, controlling signal accordingly according to regular text string generation in step 500 includes:

Step 510, semantic instructions are converted by synonym mapping by regular character string.

After executing step 400 and obtaining regular character string, obtained regular character string is mapped by term vector, conversion For the semantic instructions of format more standard.Semantic instructions are can to express standardized instruction letter semantic contained by instruction voice Breath, for example, regular character string " air-conditioning please be shut ", " air-conditioning please have been close " etc. can be converted into standard term " sky please be close Adjust ", wherein " shutting ", " pass " and other possible same semantic terms are all converted into during synonym maps Standard term " closing ".

Control system can come with semantic instructions collection, wherein preserve several can generate synonymous mapping with semantic instructions Word, such as above-mentioned " shutting ", " pass " will be stored in semantic instructions concentration, and can be mapped in " closing ".

Step 520, control signal is generated according to semantic instructions, and executes phase according to control signal control building service facility Answer function.

The semantic instructions of standard have corresponding machine identifier, by taking above-mentioned " air-conditioning please be close " by converting as an example, Air-conditioning can be issued using the machine identifier of " air-conditioning please be close " as control signal, air-conditioning can be held according to the machine identifier Row control.

Using the high operational capability of control system itself, high data-handling capacity come in advance by the regular character of wide variety String is converted into unified, the relatively small number of semantic instructions of total quantity, then regeneration control signal, so that control system and controlled Communication between equipment is more succinct, and reduces the requirement to controlled device data-handling capacity, and controlled device only needs energy The i.e. executable corresponding function of enough identification control signal corresponding with to count less semantic instructions, has without laying in advance The instruction of a variety of mapping relations-signal mapping library, may be by the control signal for the wide variety that control system is sent to cope with.

It is understood that can only say that the partial words of regular character string are mapped by synonym in step 510 and convert For semantic instructions.Specifically, regular character string first can be carried out word cutting processing, by a word sequence before step 510 It is cut into individual word one by one, the words such as verb is then therefrom selected and is converted, such as " please shut air-conditioning " is carried out Word cutting processing, obtains " shutting ", then executes step 510 and obtains semantic instructions " closing ", then will be corresponding with " closing " Machine identifier is sent to controlled device-air-conditioning, so that air-conditioning is closed.

In one embodiment, it after controlling signal accordingly according to regular text string generation in step 500, also wraps It includes:

Step 600, collected instruction voice signal is stored in instruction voice sample database.

Collected instruction voice signal is in addition to carrying out speech recognition conversion to realize the finally control to building service facility It, can also be after finally determining corresponding real instruction other than system, instruction voice signal update that initial acquisition is arrived is to referring to It enables as instruction voice sample in speech samples library, to be extracted when other voices as speech vector for identifying later；Or Voice signal reference when as Application on Voiceprint Recognition；Or when confirming that vocal print meets the requirements and includes mark word in voice signal, A pre-identification step can be added, first to instruction voice signal carry out audio signal on pre-identification, and pre-identification comparison pair A plurality of voice as being exactly the habit of usually pronouncing according to user stored, that is, just deposited after being correctly validated and execute before Storage is the voice signal of the object of pre-identification comparison.

It is stored in instruction voice sample database by the instruction voice signal that will correctly identify, it can be convenient for use later Family issues identification process when same phonetic order.

Below with reference to the building service facility control disclosed in Fig. 2 the present invention is described in detail based on semantic instructions intelligent recognition System first embodiment.The present embodiment is the control system for implementing above-mentioned building service facility control method first embodiment System.The present embodiment is mainly used in building service facility, allows user to pass through instruction voice to control building service facility The operation of interior various service equipments carries out manual manipulation control panel with hand without user, so that control mode is more intelligent With it is convenient；Can also have an accent simultaneously in the instruction voice for user's sending of speaking, pronounce it is lack of standardization and when pronouncing indistinctly, it is right Instruction voice is correctly associated and is identified, to obtain correct instruction voice, avoids since the problem of voice leads to user It can not normal control building service facility.

As shown in Fig. 2, the present embodiment is disclosed to build service facility control system, comprising: signal acquisition module, primary word Accord with generation module, low confidence character determining module, regular character generation module and control signal generation module.

Signal acquisition module is used for acquisition instructions voice signal.Signal acquisition module can be using microphone or by microphone Selection with the sound pick-up etc. of audio amplifier circuit composition, sound pick-up is arranged according to the size of environment space, needs under large space The biggish sound pick-up of maximum pickup range is chosen, camera can also be used.

Primary character generation module is connect with signal acquisition module, for converting instruction voice signal by speech recognition For primary character string.

Low confidence character determining module is connect with primary character generation module, for determining low set from primary character string Believe character.

Regular character generation module is connect with low confidence character determining module, for by the characteristic parameter generation of low confidence character Enter in neural network, obtains regular character string.

Control signal generation module is connect with regular character generation module, for controlling accordingly according to regular text string generation Signal processed, and corresponding function is executed according to control signal control building service facility.

In one embodiment, control system further include: noise processed module, for passing through by instruction voice signal Speech recognition is converted into before primary character string, carries out noise reduction to collected instruction voice signal and/or echo cancellation is handled.

In one embodiment, control system further include: vocal print judgment module, for passing through by instruction voice signal Speech recognition is converted into before primary character string, and the vocal print of instruction voice signal is identified by sound groove recognition technology in e, judges to know Not Chu vocal print whether be possessed of control power limit, and in the case where the vocal print identified does not have control authority stop executing subsequent Step.

In one embodiment, control system further include: mark identification module, for being determined from primary character string Out before low confidence character, identify in instruction voice signal or primary character string whether include the voice signal or word for identifying word Symbol, and in the case where the voice signal or character of the unidentified word of mark out, stop executing subsequent step.

In one embodiment, low confidence character determining module includes: quantity statistics unit and low confidence character occur Determination unit.

There is quantity statistics unit for calculating the institute that each character includes in instruction voice sample database in primary character string Have in sample character string, the quantity of the sample character string comprising each character.

Low confidence character determination unit is connect with there is quantity statistics unit, for that will calculate the sample word comprising each character It accords with the quantity of string and is determined as low confidence character compared with minimum confidence threshold, and by the character that quantity is lower than minimum confidence threshold.

In one embodiment, control system further include: high confidence character determining module, for by low confidence character Characteristic parameter substitute into neural network in obtain regular character string before, high confidence character is determined from primary character string, really The high confidence character of the forerunner for making low confidence character and/or subsequent high confidence character.

Also, regular character generation module includes: that character set determination unit, neural metwork training unit and regular character are raw At unit.

Character set determination unit is used to determine the subsequent character collection of the high confidence character of forerunner from instruction voice sample database And/or forerunner's character set of subsequent high confidence character.

Neural metwork training unit is connect with character set determination unit, for according to subsequent character collection and/or forerunner's character The acoustic vector training neural network of some or all of concentration character.

Regular character generation unit is connect with neural metwork training unit, is used for the corresponding acoustic vector of low confidence character It substitutes into neural network and carries out Regularization identification, to determine replacement word from corresponding subsequent character collection and/or forerunner's character set Symbol, and replace low confidence character and obtain regular character string.

In one embodiment, high confidence character determining module includes: that the first character determination unit and the second character are true Order member,

First character determination unit is used for the characters for being not determined to low confidence character all in primary character string are true It is set to high confidence character.And/or

Second character determination unit is connect with the first character determination unit, for that will calculate the sample character comprising each character The quantity of string is determined as high confidence character compared with maximum confidence threshold, and by the character that quantity is higher than maximum confidence threshold.

In one embodiment, character set determination unit includes: that third character determines subelement, frequency statistics subelement With character set synthesizing subunit.

Third character determine subelement for determined in instruction voice sample database the high confidence character of forerunner it is all after After character, and/or determine all forerunner's characters of subsequent high confidence character.

Frequency statistics subelement determines that subelement is connect with third character, for counting subsequent character and/or forerunner's character The frequency occurred in all sample character strings.

Character set synthesizing subunit is connect with frequency statistics subelement, for according to frequency collating or according to default frequency threshold Value determines the highest multiple subsequent characters of the frequency and/or forerunner's character, and separately constitutes subsequent character collection and/or forerunner's character Collection.

In one embodiment, frequency statistics subelement is also used to according to frequency collating or according to default frequency threshold value Before determining the highest multiple subsequent characters of the frequency and forerunner's character, first to identical characters in subsequent character and forerunner's character Frequency of occurrence is summed, using the frequency data after summing as the frequency of occurrence of character.

In one embodiment, neural metwork training unit only chooses the part of respective symbols in instruction voice sample database Acoustic vector trains neural network.

In one embodiment, control system further include: sample memory module, for concatenating according to regular character After corresponding control signal, collected instruction voice signal is stored in instruction voice sample database.

In one embodiment, control signal generation module includes: Semantic mapping unit and signal generation unit.

Semantic mapping unit is used to convert semantic instructions by synonym mapping for regular character string.

Signal generation unit is connect with Semantic mapping unit, for generating control signal according to semantic instructions, and according to control Signal control building service facility processed executes corresponding function.

The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by those familiar with the art, all answers It is included within the scope of the present invention.Therefore, protection scope of the present invention should be with the scope of protection of the claims It is quasi-.

Claims (10)

1. a kind of building service facility control method based on semantic instructions intelligent recognition characterized by comprising

Acquisition instructions voice signal；

Primary character string is converted by speech recognition by described instruction voice signal；

Low confidence character is determined from the primary character string；

The characteristic parameter of the low confidence character is substituted into neural network, regular character string is obtained；

Signal is controlled accordingly according to the regular text string generation, and is controlled building service facility according to the control signal and held Row corresponding function.

2. control method as described in claim 1, which is characterized in that described to determine low confidence from the primary character string Character includes:

Calculating each character in the primary character string includes each in all sample character strings that instruction voice sample database includes The quantity of the sample character string of the character；

By the quantity of the sample character string of the calculating comprising each character compared with minimum confidence threshold, and quantity is lower than The character of the minimum confidence threshold is determined as low confidence character.

3. control method as claimed in claim 1 or 2, which is characterized in that join in the feature by the low confidence character Number substitutes into neural network before obtaining regular character string, further includes:

High confidence character is determined from the primary character string；

Determine the high confidence character of forerunner and/or subsequent high confidence character of the low confidence character；Also,

Obtaining regular character string in the characteristic parameter substitution neural network by the low confidence character includes:

Determined from instruction voice sample database the high confidence character of the forerunner subsequent character collection and/or the subsequent high confidence Forerunner's character set of character；

Acoustic vector training nerve according to character some or all of in the subsequent character collection and/or forerunner's character set Network；

The corresponding acoustic vector of the low confidence character is substituted into the neural network and carries out Regularization identification, with from corresponding institute It states in subsequent character collection and/or forerunner's character set and determines substitute character, and replace the low confidence character and obtain regular character String.

4. control method as claimed in claim 3, which is characterized in that it is described from determined in instruction voice sample database it is described before Forerunner's character set of the subsequent character collection and/or the subsequent high confidence character that drive high confidence character includes:

All subsequent characters of the high confidence character of the forerunner are determined in described instruction speech samples library, and/or are determined All forerunner's characters of the subsequent high confidence character；

Count the frequency that the subsequent character and/or forerunner's character occur in all sample character strings；

According to frequency collating or according to default frequency threshold value determine the highest multiple subsequent characters of the frequency and/or it is described before Character is driven, and separately constitutes subsequent character collection and/or forerunner's character set.

5. control method as described in claim 1, which is characterized in that described to be controlled accordingly according to the regular text string generation Signal processed includes:

Semantic instructions are converted by synonym mapping by the regular character string；

Control signal is generated according to the semantic instructions, and controls building service facility according to the control signal and executes corresponding function Energy.

6. a kind of building service facility control system based on semantic instructions intelligent recognition characterized by comprising

Signal acquisition module is used for acquisition instructions voice signal；

Primary character generation module, for converting primary character string by speech recognition for described instruction voice signal；

Low confidence character determining module, for determining low confidence character from the primary character string；

Regular character generation module obtains regular word for substituting into the characteristic parameter of the low confidence character in neural network Symbol string；

Signal generation module is controlled, for controlling signal accordingly according to the regular text string generation, and according to the control Signal control building service facility executes corresponding function.

7. control system as described in claim 1, which is characterized in that the low confidence character determining module includes:

There is quantity statistics unit, for calculating the institute that each character includes in instruction voice sample database in the primary character string Have in sample character string, the quantity of the sample character string comprising each character；

Low confidence character determination unit, for setting the quantity for calculating the sample character string comprising each character with minimum Believe threshold value comparison, and the character by quantity lower than the minimum confidence threshold is determined as low confidence character.

8. control system as claimed in claims 6 or 7, which is characterized in that the control system further include:

High confidence character determining module is obtained for substituting into neural network in the characteristic parameter by the low confidence character Before regular character string, high confidence character is determined from the primary character string, determines the forerunner of the low confidence character High confidence character and/or subsequent high confidence character；Also,

The regular character generation module includes:

Character set determination unit, for determining the subsequent character collection of the high confidence character of the forerunner from instruction voice sample database And/or forerunner's character set of the subsequent high confidence character；

Neural metwork training unit, for according to some or all of in the subsequent character collection and/or forerunner's character set The acoustic vector training neural network of character；

Regular character generation unit, for advising the low confidence character corresponding acoustic vector substitution neural network Integralization identification to determine substitute character from the corresponding subsequent character collection and/or forerunner's character set, and is replaced described low Confidence character obtains regular character string.

9. control system as claimed in claim 8 is it is characterized in that, the character set determination unit includes:

Third character determines subelement, for determining the institute of the high confidence character of the forerunner in described instruction speech samples library There is subsequent character, and/or determines all forerunner's characters of the subsequent high confidence character；

Frequency statistics subelement, for counting the subsequent character and/or forerunner's character in all sample character strings The frequency of middle appearance；

Character set synthesizing subunit, for determining the highest multiple institutes of the frequency according to frequency collating or according to default frequency threshold value Subsequent character and/or forerunner's character are stated, and separately constitutes subsequent character collection and/or forerunner's character set.

10. control system as claimed in claim 6, which is characterized in that the control signal generation module includes:

Semantic mapping unit, for converting semantic instructions by synonym mapping for the regular character string；

Signal generation unit for generating control signal according to the semantic instructions, and is controlled according to the control signal and is built Service facility executes corresponding function.