KR101179915B1 - Apparatus and method for cleaning up vocalization data in Voice Recognition System provided Statistical Language Model - Google Patents

Abstract

PURPOSE: A speaking data filtering device of a voice recognizing system using a statistical language model and a method thereof are provided to consistently filter a sentence made by voice recognition and update statistical language model grammar of an existing voice recognizing engine by a trained statistical language model. CONSTITUTION: A speaking voice collecting unit(1) collects speaking voices of users and corresponding keywords. A voice processing unit(100) converts a voice signal outputted from the speaking voice collecting unit into voice data. A voice recognizing unit(200) recognizes the voice data by stored statistical language model grammar. The voice recognizing unit updates the statistical language model grammar. A filtering unit(300) corrects a sentence made by voice recognition. The filtering unit generates an update version of a statistical language model about the corrected sentence.

Description

Apparatus and method for cleaning up vocalization data in Voice Recognition System provided Statistical Language Model}

The present invention relates to a speech recognition system using a statistical language model (SLM), and more particularly, to collect collected customer speech (or "talk") data in a consistent form after speech recognition. The present invention relates to an apparatus and method for refining speech data of a speech recognition system to which a statistical language model, which can be applied by updating a statistical language model grammar of a speech recognition engine by learning a statistical language model.

In general, one of the important factors that determine the satisfaction of a customer's purchase of a product or service is customer-to-customer service such as an enterprise or a public office for resolving customer's AS and customer complaints and providing fast and prompt after-sales service. Recently, the automatic response system is to perform the window area dedicated to customer service.

Automated answering systems have evolved into third generation CTI automated answering systems (or "CTI call center systems"), combining telephones and computers, through the second generation, which relied on exchange and answering machine functions.

Automated answering system can reduce the cost of customer service processing in the enterprise, act as a branch office, obtain customer information quickly and efficiently, easily, and provide rapid service. It is developing.

Currently, CTI auto answering system that integrates both telephone and computer is the mainstream, and CTI auto answering system combines computer and telephone to distribute the incoming calls efficiently and to help the counselor to consult effectively. Information is grasped and provided in advance, and various additional functions are provided through a computer.

In addition, in order to further develop the CTI auto answering system, an auto answering system including a voice recognition system for performing voice recognition in order to quickly recognize and process a customer's requirements has been developed and applied.

Statistical language model (SLM) is widely applied as a speech recognition technology applied to an automatic answering system including the speech recognition system.

Statistical language model (SLM) has the advantage of improving the accuracy of the system and reducing the execution time in various fields of natural language processing.

However, very different vocalization patterns appear in the same sentence depending on the user's age, language learning ability, dialect, and pronunciation accuracy. In order to train them with statistical language models, the user texts corresponding to age, language learning ability, dialect, and accuracy of pronunciation, and analyzes speech patterns. In this case, since the recognition target is learned in various forms, the recognition result has a high probability of appearing in a meaning other than that of the original form, thereby degrading speech recognition performance.

For example, if the recognition target is A, the recognition result A is more likely to be represented as B, C, D, or the like instead of A according to the user's age, language learning ability, dialect, and accuracy of pronunciation. Therefore, speech recognition performance is poor.

Therefore, an object of the present invention is to apply a statistical language model that can be applied to update the statistical language model of the speech recognition engine by refining the collected speech data of the customer in a consistent form after speech recognition and then learning the statistical language model An apparatus and method for refining speech data of a speech recognition system are provided.

An apparatus for reproducing speech data of a speech recognition system to which the statistical language model of the present invention is applied to achieve the above object; A speech voice that matches and collects the user's speech voice and the keyword of the speech voice, collects and stores the voice signals for the speech voice collected when the refining processing request event is generated, and outputs a keyword corresponding to the output voice signal. A speech processing unit for converting the speech signal output from the speech speech collecting unit into speech data and outputting the speech data; and a speech recognition result of the speech data output from the speech processing unit by pre-stored statistical language model grammar, Outputs a speech recognition unit for receiving a statistical language model update version and updating the statistical language model grammar, and inputting a keyword corresponding to the speech recognition result and the speech recognition result and consistently correcting a sentence of the speech recognition result. Update statistical language model for corrected sentences It characterized in that it comprises a tablet unit for generating a version to provide the speech recognition unit.

The speech recognition unit may include a simulation unit configured to output a simulation result by performing a simulation on the updated statistical language model when the statistical language model grammar is updated by the statistical language model updated version.

The refiner may include a word detector that divides and outputs a word and word unit from a sentence unit recognition result output from the speech recognizer, and a correction type for performing the correction of the word, word, and sentence unit according to the correction type. A word word and word phrases and keywords for a sentence of a recognition result output from the word detection part, and a word corrected by performing a correction according to the type of correction by referring to the dictionary part; The correction unit for outputting words by correction type and the sentence output by each correction type are combined to generate a corrected sentence, and the statistical language model is trained by the statistical language model on the generated sentence. And a statistical language model generator for generating and outputting an updated version. .

The dictionary unit may include a grammar dictionary for defining grammar of words for correcting a typo in a word unit, a spacing dictionary for storing a spacing grammar specialized for each application field to correct a spacing of a word or a plurality of word units, and a word unit. Standard conversions for phonetic writing and phonetic writing for different words, and a variable conversion dictionary for each of a number of standard endings for matching words on a word-by-word basis. And a match dictionary which defines a representative ending to match the variable endings.

The correction unit receives words constituting the sentence of the speech recognition result from the word detection unit, checks whether there is a typo among the words by referring to the grammar dictionary, corrects a normal word if there is a typo, and generates a statistical language model generator. Corrects a word-based spacing for the words of the sentence by referring to the spacing dictionary by receiving a typo correcting unit and a sentence unit of words from the word detecting unit, and receiving a keyword corresponding to the sentence from the spoken speech collector. A spacing correction unit for outputting to a statistical language model generator and outputting words for a sentence of a recognition result from a word detection unit, and referring to the writing conversion dictionary, whether the words are different from standard writing and writing as sounds. Judge the standard writing and phonetic writing If it is a different word, the phonetic writing correction unit which converts the phonetic writing into a phonetic writing and outputs it to the statistical language model generating unit, and receives the words for the sentence of the recognition result from the word detecting unit, and refers to the And a mother matching correction unit for matching to a representative mother corresponding to the mother and outputting to the statistical language model generator.

The writing conversion dictionary may further define a plurality of additional pronunciations which are actually pronounced according to the pronunciation accuracy and dialect for basic vocabulary that is standard writing and basic vocabularies that vary depending on pronunciation accuracy and dialect among the basic vocabulary. The phonetic writing correction unit receives a word for the sentence of the recognition result from the word detection unit, and when there are a plurality of pronunciations according to pronunciation accuracy and dialect for the basic vocabulary of each of the words by referring to the writing conversion dictionary. Characterized by the words for the actual utterance (writing as spoken) corresponding to the region to which the speech recognition system is applied is characterized in that the output to the statistical language model generator.

The spoken voice collecting unit may collect and match the spoken voices of the users who call the automatic answering system using the automatic answering system and the keywords of the corresponding stages of the scenario inducing the spoken voice.

Speech data purification method of the speech recognition system to which the statistical language model of the present invention is applied to achieve the above object; A collection process of collecting and matching the speech voices of the users with the speech voices and the keywords corresponding to the speech voices; and when the refining processing request event occurs, the speech voice collection unit uses the collected speech voice units to match the keywords with the speech voices. A process of initiating and outputting the inspection and refining, a voice data conversion process of receiving a voice signal for the spoken voice and converting the voice signal into voice data, and a voice recognition unit receiving the voice data to perform voice recognition to perform the speech A voice recognition process of outputting a sentence of a speech recognition result corresponding to a voice, and a correction unit receives a sentence of the speech recognition result, and receives a keyword matched to the speech voice from the speech speech collection unit receives a sentence of the speech recognition result The language consistently and update the statistical language model for the corrected statement Purification to produce the former to provide parts of the speech recognition, and receives the speech recognition, an input unit for statistical language model updated version of the statistical language model that updates the statistical language model dry Yes Yes Do characterized in that it comprises the updating process.

 The speech recognition unit may further include a simulation process of performing a simulation on the updated statistical language model grammar and outputting the performed simulation result.

The refining processing request event may be automatically generated at one or more predetermined periods of time, days, weeks, months, and years.

The refinement process may include a word detection step of the word detection unit dividing the sentence of the recognition result into words and words, and a correction unit receiving the divided words, words and the keywords according to the type of correction, and referring to a dictionary unit. A calibration step of outputting the corrected words and phrases for each type of correction by performing a correction according to the type of correction, and receiving a sentence composed of words or words output by the statistical language model generation unit for each type of correction from the correction unit Generating a sentence reflecting the correction for each type of correction, generating a statistical language model updated version of the generated sentence, and outputting it to the speech recognition unit.

In the correcting step, the words constituting the sentence of the speech recognition result are input from the word detection unit, and a grammar dictionary is checked to see if there is a typo among the words, and if there is a typo, it is corrected to a normal word to the statistical language model generator. A word-based word correction for the words of the sentence is performed by a typo correcting step of outputting the word, a word unit of the sentence from the word detection unit, and a keyword corresponding to the sentence from the spoken speech collector. A spacing correction step of outputting to a statistical language model generation unit and receiving words for sentences of the recognition result from the word detection unit, and determining whether the words are different from standard writing and writing as described by referring to a writing conversion dictionary. Different from standard writing and phonetic writing If the word is converted into phonetic writing and output to the statistical language model generation unit, the phonetic writing correction step and the word detection unit receive the words for the sentence of the recognition result, and the word matching dictionary is referred to the word ending. And matching the corresponding representative endings to output to the statistical language model generation unit.

In the correcting step, when the words for the sentence of the recognition result are input from the word detection unit and a plurality of pronunciations are present according to pronunciation accuracy and dialect for the basic vocabulary of each word by referring to the writing conversion dictionary The apparatus further includes an additional pronunciation correction step of correcting the word for the actual utterance (writing as spoken) corresponding to the region to which the system is applied and outputting it to the statistical language model generator.

The spoken voice collected in the collection process is a voice spoken by users calling an automatic answering system using an automatic answering system, and the keyword is a keyword of a corresponding step of a scenario inducing the spoken voice. do.

The present invention is to refine the speech recognition sentences of the user uttered speech data as consistently as possible and then trained according to the statistical language model, and by updating the statistical language model grammar of the existing speech recognition engine by the trained statistical language model statistical language It has the effect of improving the speech recognition performance of the speech recognition engine applying the model.

1 is a view showing the configuration of a speech data purification apparatus of a speech recognition system to which a statistical language model is applied according to the present invention.
2 is a view showing the detailed configuration of a purification unit of the ignition data purification device according to the present invention.
3 is a flowchart illustrating a method of purifying speech data of a speech recognition system to which a statistical language model is applied according to the present invention.
Figure 4 is a flow chart illustrating in detail the purification process of the speech data purification method of the speech recognition system to which the statistical language model is applied according to the present invention.

Hereinafter, a configuration and operation of a speech data purification apparatus of a speech recognition system to which a statistical language model is applied according to the present invention will be described, and a speech data purification method in the apparatus will be described.

1 is a view showing the configuration of a speech data purification apparatus of a speech recognition system to which a statistical language model is applied according to the present invention.

Referring to FIG. 1, an apparatus for refining speech data of a speech recognition system to which a statistical language model is applied according to the present invention includes a speech collecting unit 1, a speech processing unit 100, a speech recognition unit 200, and a purification unit 300. It includes.

The spoken voice collecting unit 1 collects and matches a user's spoken voice with a keyword corresponding to the spoken voice. The spoken voice may be collected by collecting the voices of the users who call the call center in which the automatic answering system using the voice recognition is constructed, or collected by the developer directly, and the keyword may be an automatic response that induces each spoken voice. It could be a keyword for that phase of the system's scenario, or a keyword entered directly by the developer. In this case, the recorded scenario is recorded in sentence units, and the recorded speech voice and a keyword matching the speech voice are matched and stored with unique identification information.

The spoken voice collecting unit 1 outputs a voice signal for the spoken voices collected when the purification processing request event occurs to the voice processing unit 100, and outputs a keyword matching the voice signal output to the voice processing unit 100. Output to the purification unit 300. At this time, the purification unit 300 knows that the voice signal output to the voice processing unit 100 by the identification information included in the input keyword. The refining processing request event may be configured to occur at least one or more of a certain period of time, such as time, day, week, month and year, or may be generated by a voice recognition developer's request.

The voice processing unit 100 converts the voice signal output from the spoken voice collecting unit 1 into voice data and outputs the voice signal to the voice recognition unit 200.

The speech recognition unit 200 performs speech recognition on the speech data input from the speech processing unit 100 by a statistical language model grammar, which is trained by a statistical language model and stored in advance, and the recognition unit 300 receives the recognition result. Will output The voice recognition unit 200 receives the statistical language model update version from the refiner 300 and updates the statistical language model grammar that has been stored in advance. The speech recognition unit 200 also performs a simulation on the updated statistical language model when the statistical language model grammar is updated, and outputs a simulation result. The output simulation result is stored, and the stored simulation result may be displayed or printed according to the developer's choice.

The purification unit 300 receives a sentence and a keyword of the voice recognition result input from the voice recognition unit 200 and undergoes a correction process such as a typo of the sentence, keyword-based spacing, writing as it sounds and matching the ending. The sentence of the recognition result is consistently corrected, and a statistical language model updated version of the corrected sentence is generated and provided to the speech recognition unit 20. In addition, the refiner 300 stores the corrected sentences, checks the words and phrases of the stored sentences, checks when there are various spacings of a specific vocabulary, generates a list of correction rule addition targets, and outputs the output list.

In FIG. 1, only the schematic operation of the purification unit has been described. Hereinafter, a detailed configuration and operation of the purification unit 300 according to the present invention will be described.

2 is a view showing the detailed configuration of a purification unit of the ignition data purification device according to the present invention.

The refiner 300 includes a word detector 310, a corrector 320, a dictionary 330, a statistical language model generator 340, and a refinement rule generator 350.

The word detection unit 310 receives a recognition result of a sentence unit output from the voice recognition unit 200, divides the word unit into words and word units, and outputs words, words, a plurality of words, and units of sentences according to the type of correction. do.

The dictionary unit 330 may include a plurality of dictionaries according to the correction type in order to perform correction on one or more of words, phrases, a plurality of words, and sentences in sentence units according to the correction type in the correction unit 320. Equipped.

In detail, the dictionary unit 330 includes a grammar dictionary 321, a spacing dictionary 332, a writing dictionary 333 and a coincidence dictionary 334 as they are spoken.

The grammar dictionary 321 defines grammars of words for correcting a typo in a word unit.

The spacing dictionary 332 stores a spacing grammar that is specialized for each application field in order to correct spacing of a word or a plurality of word units.

The write conversion dictionary 333 matches and stores words that are different from each other in terms of phonetic writing and phonetic writing. In addition, the write conversion dictionary 333 may be pronounced in addition to the basic pronunciation dictionary of each vocabulary in the spoken data used in producing a language model, since the same word may be pronounced in various forms according to pronunciation accuracy and dialect, as shown in Table 1 below. The actual pronunciation is added. In other words, the writing conversion dictionary 333 may further define a plurality of additional pronunciations which are actually pronounced according to the pronunciation accuracy and dialect for basic vocabularies and basic vocabularies that vary depending on pronunciation accuracy and dialect among the basic vocabularies. .

Vocabulary
(Standard write) Pronunciation dictionary Real speech
(Write as you hear) Recognition result
message mzedZi Basic Pronunciation Dictionary message message mezidZi Pronunciation Pronunciation message message mesedZi Pronunciation Pronunciation Message message mesidZi Pronunciation Pronunciation Message message

The ending matching dictionary 334 defines variable endings for each of a plurality of standard endings for matching endings on a word-by-word basis and defines a representative ending for matching the variable endings. The representative and standard endings may or may not be identical.

The corrector 320 includes a typo corrector 321, a spacing corrector 322, a write corrector 323, and a coincidence corrector 324 as it sounds, from the word detector 310. The words, words, plural words, and sentence units of words are corrected and output to be consistent according to the type of correction. Types of correction include typo correction, spacing correction, phonetic write correction, and parental correction.

Specifically, the typo correcting unit 321 refers to the grammar dictionary 331 of the dictionary unit 330 to check whether there is a typo in words constituting the sentence of the speech recognition result, and if there is a typo, The proofed word is output to the statistical language model generator 340.

The spacing correction unit 332 receives words in sentence units from the word detection unit 310, receives a keyword corresponding to the sentence from the spoken speech collection unit 1, and the spacing dictionary 332 refers to a keyword-based word. Perform spacing correction and output.

Table 1 below shows keyword-based spacing correction examples. When the first sentence of Table 1 is described as an example, the spacing correction unit 332 may generate a keyword 'payment price' for the first sentence. 1) input from the sentence, sentence units ['this time', 'month', 'payment price', 'how much'] is input. As described above, when the phrases and keywords in sentence units are input, the spacing correction unit 332 generates a statistical language model by correcting as ['this', 'month', 'payment amount', 'yi', 'how much']. Output to the unit 340. By doing so, as shown in Table 2 below, when training with a vocal data source, only the effect of learning once per sentence was separated by the keyword and trained to form a keyword that was repeatedly matched in different sentences. It has the effect of performing a lot of training. As the number of training increases, voice recognition performance also improves.

Vocal data source Learning effect Keyword separation Learning effect How much is the payment this month 1 lesson per sentence How much is the payment this month 3 times of payment Please tell me your payment. Please tell me your payment Trying to pay I'm paying the bill. I will change my plan I will change my plan My plan two lessons I want to report my loss. I want to report my loss. I want to check if the loss has been reported well. To check the loss report Lost report three times learning I will cancel the report. I will cancel the report.

The phonetic writing correction unit 323 receives words for sentences of the recognition result from the word detecting unit 310 and refers to the writing conversion dictionary 333 to determine whether the words are different from standard writing. If it is determined that the standard writing and the phonetic writing as different words, the phonetic writing is converted to the phonetic statistical model generation unit 340 is output. For example, words different from standard writing and phonetic writing may be the same as in Table 3 below.

Standard writing
(Vocabulary) Phonetic writing
(Actual speech) conversion service Service Service-> Service Diners Diners Diners-> Diners bonus Ponat Bonus-> Ponat center Canter Center-> Center Voice phishing Boys Pissing Voice Phishing-> Voice Phishing

In addition, the writing correction unit 323 as spoken refers to the writing conversion dictionary 333 as shown in Table 1 above even when a plurality of pronunciations exist for the basic vocabulary, which is the standard writing, according to pronunciation accuracy and dialect. The word about the actual utterance (writing as spoken) corresponding to the region to which the recognition system is applied is output to the statistical language model generator 340.

The word matching correction unit 324 receives the words for the sentence of the recognition result from the word detection unit 310 and refers to the word matching dictionary 334 to match the representative words corresponding to the ending words of the statistical language model. Output to the generation unit 340. In the case of conversational speech, recognition performance may deteriorate since the mother has various patterns depending on pronunciation and dialect. Therefore, by matching these mothers and training, recognition performance can be improved. For example, as shown in Table 4 below, the mother matching correction unit 324 matches [move (stem) + r (mother)] to [move (stem) + ryo (mother)], and [move (stem)]. + Drugyong (mother)] matches [move (stem) + goryo (mother)].

Word Variable mother Representative mother Same To move. I'm going to I'm going to To move To move I'm going to To change Ryeo Trying To change To change R I want to change I'm going I'm going I want to change Change R Ryeo To change Change L

Statistical language model generation unit 340 by the statistical language model for the sentence corrected by combining one or more correction data of the word, word, a plurality of words and sentence units of the word output from the correction unit 320 By performing the training, a statistical language model update version is generated and output to the speech recognition unit 200.

The refinement rule generation unit 350 may generate a single sentence of a sentence that is output from the typo correcting unit 321, the spacing correcting unit 322, the writing correcting unit 323, and the mother coincidence correcting unit 324. Each word and phrase is input and stored in sentence units, and the same vocabulary of the stored sentences is checked to generate a correction rule addition target list having different spacing but not being corrected despite the same vocabulary, and outputting it to an output file.

The developer may update the rule by viewing a list of targets for adding a correction rule.

3 is a flowchart illustrating a method of purifying speech data of a speech recognition system to which a statistical language model is applied according to the present invention. A description with reference to FIGS. 1 to 3 is as follows.

First, the spoken voice collecting unit 1 collects and stores the keyword of each step of the scenario and the user's spoken voice derived in each step while providing a voice guidance service according to the scenario of the automatic response system (S311).

The spoken voice collecting unit 1 checks whether a refinement processing request event occurs during the collection of the user spoken voice and the keyword (S312).

When the refining processing request event occurs, the spoken voice collecting unit 1 sequentially reproduces the collected keywords and the user's spoken voices, and outputs a voice signal for the collected spoken voices to the voice processing unit 100. The keyword matching the voice signal output to the processor 100 is output to the refiner 300 (S313).

The voice processor 100 receives the voice signal output from the spoken voice collector 1, converts the voice signal into voice data, and outputs the voice signal to the voice recognizer 200 (S314).

The voice recognition unit 200 performs voice recognition by using the statistical language model grammar 210, which receives the voice data and outputs the voice recognition result to the refiner 300 (S315).

The tablet unit 300 receiving the voice recognition result performs a typo correction, a spacing correction, a writing correction and a coincidence correction as described above, and corrects a typo correction, spacing correction, or sound for a sentence having the same recognition result. After combining each sentence correcting the writing correction and the parental matching correction through the statistical language model generator, a statistical language model update version is generated for the plurality of sentences that have been finally corrected (refined) and output to the speech recognition unit 200. (S317).

When the refinement is completed, the voice recognition unit 200 receives the statistical language model update version output from the refiner 300 and applies the statistical language model grammar to update the statistical language model grammar (S319).

When the statistical language model grammar 210 is updated as described above, the voice recognition unit 200 performs a simulation on the updated statistical language model grammar through the simulation unit 220 and outputs a simulation result (S321).

Figure 4 is a flow chart illustrating in detail the purification process of the speech data purification method of the speech recognition system to which the statistical language model is applied according to the present invention. Hereinafter, a process of purifying utterance data according to the present invention will be described with reference to FIG. 4.

First, the word detection unit 310 of the purification unit 300 checks whether a voice recognition result is input (S411).

When the voice recognition result is input, the word detection unit 310 divides the word constituting the sentence of the voice recognition result into a word or word unit and outputs the divided word (S412).

The typo correcting unit 321 receives the words constituting the sentence recognized by the word detection unit 310 and performs a typo check on each word by referring to the grammar dictionary 331 (S421). In operation S423, the word with a typo is corrected as a correct word and output to the statistical language model generator 340 (S425).

The spacing correction unit 322 receives the words constituting the sentence from the word detecting unit 310, performs a spacing check with reference to the spacing dictionary 332 (S427), and determines whether there is a spacing to be converted. (S429), and if present, the corresponding spacing correction is applied to the statistical language model generation unit 340 (S431).

And the phonetic writing correction unit 323 receives the words constituting the sentence from the word detection unit 310 and by referring to the write conversion dictionary 333 determines whether there is a word to be written as a phonetic sound (S435), If present, the standard writing word of the corresponding word is corrected as a word of writing as it is sounded and output to the statistical language model generator 340 (S437).

In addition, the word matching correction unit 324 receives the words constituting the sentence of the recognition result from the word detecting unit 310, and the word matching dictionary 334 refers to a mother for checking whether there is a representative ending for the ending words. A matching test is performed (S439), and according to the presence or absence thereof (S441), the calibration is performed by matching the representative mother and then output to the statistical language model generator 340 (S443).

The statistical language model generator 340 then corrects a typo correcting unit 321, a spacing correction unit 322, a phonetic writing correction unit 323, and a mother coincidence correction unit for the sentence of one recognized result as described above. The corrected word unit or word unit sentences output from the input unit 324 are received and combined into one corrected sentence in which all corrections are reflected, and then stored (S445).

While the present invention has been described with reference to exemplary embodiments, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. It will be easily understood. If such improvement, change, substitution or addition is carried out within the scope of the appended claims, the technical spirit should also be regarded as belonging to the present invention.

1: speech collection unit 100: voice processing unit
200: speech recognition unit 210: statistical language model grammar
220: simulation unit 300: purification unit
310: word detection unit 320: correction unit
321: typo correction unit 322: spacing correction unit
323: phonetic writing correction unit 324: mother matching correction unit
330: dictionary section 331: grammar dictionary
332: Spacing dictionary 333: Write conversion dictionary
334: Matching dictionary 340: Statistical language model generator
350: refinement rule generation unit

Claims (14)

A speech voice that matches and collects the user's speech voice and the keyword of the speech voice, collects and stores the voice signals for the speech voice collected when the refining processing request event is generated, and outputs a keyword corresponding to the output voice signal. Collector,
A voice processing unit for converting and outputting a voice signal output from the spoken voice collecting unit into voice data;
A speech recognition unit for recognizing speech data output from the speech processing unit by pre-stored statistical language model grammar and outputting the recognition result, and receiving a statistical language model update version to update the statistical language model grammar;
And a refiner configured to receive the voice recognition result and the keywords corresponding to the voice recognition result, to consistently correct the sentence of the voice recognition result, and to generate a statistical language model update version of the corrected sentence to provide the voice recognition unit. An apparatus for reproducing speech data of a speech recognition system to which a statistical language model is applied.
The method of claim 1,
The voice recognition unit,
And a simulation unit configured to output a simulation result by performing a simulation on the updated statistical language model when the statistical language model grammar is updated by the statistical language model updated version of the speech recognition system to which the statistical language model is applied. Ignition data purification device.
The method of claim 1,
The purification unit,
A word detection unit for dividing and outputting word and word units from sentence recognition results output from the voice recognition unit;
A dictionary unit for storing dictionaries according to a type of correction for performing the correction of the word, word, sentence, according to the type of correction;
A correction unit for receiving words, phrases, and keywords for sentences of the recognition result output from the word detection unit, and performing correction according to the type of correction by referring to the dictionary unit and outputting the corrected words and words for each type of correction; ,
The statistical unit generates a corrected sentence by combining sentences outputted by the type of correction by the correction unit, and generates a statistical language model updated version by training a statistical language model on the generated sentence to generate and output a statistical language model. An apparatus for reproducing speech data of a speech recognition system to which a statistical language model is applied, including a unit.
The method of claim 3,
The dictionary unit,
A grammar dictionary that defines the grammar of words to correct typos in word units,
A spacing dictionary that stores spacing grammars specific to each application to correct spacing of words or multiple word units,
Write conversion dictionaries, where standard writing and phonetic writing for word-to-word phonetic correction are matched and stored for different words;
Speech Recognition with Statistical Language Models, which includes a word matching dictionary that defines variable endings for each of a plurality of standard endings for matching words on a word-by-word basis and defines a representative ending for matching the variable endings. Ignition data purification device of the system.
The method of claim 4, wherein
The correction unit,
Typing words constituting the sentence of the speech recognition result input from the word detection unit, checks if there is a typo among the words by referring to the grammar dictionary, corrects the normal word if there is a typo, and outputs it to the statistical language model generator With the government,
Generates a statistical language model by inputting words in a sentence unit from a word detection unit, and receiving a keyword corresponding to the sentence from a spoken speech collector and performing keyword-based spacing correction on the words of the sentence with reference to the spacing dictionary. A spacing correction unit to output negative
When the words for the sentence of the recognition result are input from the word detection unit, the word is determined by referring to the writing conversion dictionary to determine whether the words are different from the standard writing and the phonetic writing. A phonetic writing correction unit which converts the phonetic writing into output and outputs it to the statistical language model generator;
And a word matching correction unit for receiving words from the word detection unit for the sentences of the recognition result, matching the representative words corresponding to the word endings of the word by referring to the word matching dictionary, and outputting the word to the statistical language model generator. An apparatus for refining speech data of a speech recognition system to which a statistical language model is applied.
The method of claim 5,
The write conversion dictionary,
A plurality of additional pronunciations which are actually pronounced according to the pronunciation accuracy and dialect are further defined for basic vocabularies that are standard writing and basic vocabularies that vary depending on pronunciation accuracy and dialect among the basic vocabularies,
Write correction part as mentioned above,
If a word is input from the word detection unit for the sentence of the recognition result and a plurality of pronunciations are present according to pronunciation accuracy and dialect for each basic vocabulary of each word by referring to the writing conversion dictionary, An apparatus for reproducing speech data of a speech recognition system to which a statistical language model is applied, which is corrected by a word for a corresponding actual utterance (writing as a phonetic sound) and output to a statistical language model generator.
The method of claim 1,
The spoken voice collection unit,
Speech data of the speech recognition system to which the statistical language model is applied, characterized in that the speech of the users calling the automatic answering system using the automatic answering system and the keywords of the corresponding stages of the scenario inducing the speeching are matched and collected. refinery.
A collection process of collecting, by the spoken voice collector, a user's spoken voice and a keyword corresponding to the spoken voice;
A test purification initiation process of outputting a keyword matching the speech voice to the collected speech voice unit when a purification processing request event occurs;
A voice data conversion process of receiving a voice signal for the spoken voice by converting the voice signal into voice data;
A voice recognition process of receiving a voice recognition unit to receive the voice data and performing voice recognition to output a sentence of a voice recognition result corresponding to the spoken voice;
A correction unit receives a sentence of the speech recognition result, receives a keyword matched to the speech voice from the spoken speech collector, consistently corrects a sentence of the speech recognition result, and updates a statistical language model updated version of the corrected sentence. A purification process generated and provided to the voice recognition unit;
And a statistical language model grammar updating process of receiving a statistical language model update version by the speech recognition unit to update the statistical language model grammar.
9. The method of claim 8,
The speech recognition unit further comprises a simulation process for performing a simulation on the updated statistical language model grammar, and outputs the result of the simulation results, the speech data purification method of the speech recognition system to which the statistical language model is applied.
9. The method of claim 8,
The refinement processing request event, the speech data refining method of the speech recognition system to which the statistical language model is applied, characterized in that the automatic occurrence of one or more intervals of one or more of the period of time, day, week, month and year.
9. The method of claim 8,
The purification process,
A word detection step of the word detection unit dividing a sentence of the recognition result into words and words and outputting the divided word;
A correcting step of receiving a corrected word and phrases according to the type of correction by inputting the divided words and phrases and the keyword according to the type of correction, and by performing a correction according to the type of correction by referring to a dictionary;
The statistical language model generator receives a sentence composed of words or words output for each correction type from the correction unit, generates one sentence reflecting the correction for each correction type, and updates the statistical language model updated version of the generated sentence. And a statistical language model generation step of generating and outputting the speech signal to the speech recognition unit.
The method of claim 11,
The calibration step,
Step of correcting the error from the word detection unit receives the words constituting the sentence of the speech recognition result by referring to the grammar dictionary to check whether there is a typo, and if there is a typo, correct the normal word and output it to the statistical language model generator Wow,
Receives words in units of sentences from the word detection unit, receives a keyword corresponding to the sentence from the spoken speech collector, and performs keyword-based spacing correction on the words of the sentence by referring to the spacing dictionary to the statistical language model generator. A spacing correction step to output,
The word detection unit receives the words for the sentence in the recognition result, and determines whether the words are different from the standard writing and the phonetic writing by referring to the writing conversion dictionary. I will convert the writing as it is written to the statistical language model generator, as written sound correction step,
And a word matching correction step of receiving words from the word detection unit for the sentences of the recognition result, matching the representative words corresponding to the word endings of the word by referring to a word matching dictionary, and outputting the word to the statistical language model generator. A method of purifying speech data of a speech recognition system to which a statistical language model is applied.
The method of claim 12,
The calibration step,
If a word is input from the word detection unit for the sentence of the recognition result and a plurality of pronunciations are present according to pronunciation accuracy and dialect for each basic vocabulary of each word by referring to the writing conversion dictionary, A method of refining speech data of a speech recognition system to which a statistical language model is applied, further comprising an additional pronunciation correction step of correcting a word for a corresponding actual utterance (writing as a phonetic sound) and outputting it to a statistical language model generator.
9. The method of claim 8,
The speech voice collected in the collection process,
Voices spoken by users who dialed the auto answering system using the auto answering system.
The keyword is a speech data refining method of a speech recognition system to which a statistical language model is applied, wherein the keyword is a keyword of a corresponding step of a scenario inducing speech speech.

Images