JPS60105037A - Production system of voice input sentence - Google Patents

Abstract

PURPOSE:To reduce the load of a processor and to improve the processing speed for production of a voice input sentence, by applying a statistic method for selection of candidates to delete undesired candidates and also to avoid the wrong setting of candidates. CONSTITUTION:The input voice is applied to an analysis/collation means 11 for calculation of resemblance degree, and a selecting/holding means 12 selects and holds a single or plural candidates according to the resemblance degrees. Then a candidate re-evaluation means 13 refers to the 1st statistic table 34 to reduce the number of candidates and to arrange the orders. A connectability evaluation means 14 has collation between the preceding and following candidates for each candidate by the 2nd and 3rd statistic tables 35 and 36 and then evaluates and reduces the appearance rate to send those candidates to a KANA (Japanese syllabary)/KANJI (Chinese character) converting part 20. A separating means 21 separates the processing unit by means of inter-syllable marks, etc. and sets an array route by the recognizing information to give the priority. Then the means 21 gives access to a word dictionary 37 for each route to check the grammatical connection for each form element and combination together with other route evaluation. Thus the input characters all undergo the KANA/KANJI conversion, and access is given to a KANJI dictionary 38 to deliver a KANJI-KANA sentence.

Description

DETAILED DESCRIPTION OF THE INVENTION (a) Technical Field of the Invention The present invention relates to a text creation method using voice input in a data processing system.

(bl Technology background) In recent years, with the development and spread of data processing technology, complex logic circuits for speech recognition algorithms and As high-speed, large-capacity devices become available through smaller, lower-cost highly integrated circuit devices (LSIs), data processing using Japanese audio is suitable for interactive formats and provides operators with special familiarization training syllables. A voice typewriter, which is an ideal target for a man-machine interface, has been put into practical use by taking advantage of the features that make it easy to use. A conceptual diagram of the voice input sentence creation method is shown. In the figure, lO is a voice recognition unit, 20 is a kana-kanji conversion unit, 40 is a document editing unit, and 50 is a voice recognition unit.
is an output control unit, 608 is a printer, 60b is a display, 70a is a microphone, and 70. b is an operating tool.

Conventionally, the voice input by the operator's utterance is inputted to the voice recognition unit 1o from the microphone 60a, and the result of the analysis is compared with the voice pattern conference voice dictionary to determine the singular or compound form corresponding to the voice input according to the degree of similarity. displayed on the screen of the Kanakane display 60b consisting of candidates,
The operator confirms the correct kana one by one starting from the middle power 1 of the candidate.
The method is to create a kana series and create a sentence containing kanji and kana that is easiest for Japanese people to understand. For this reason, although the method of creating sentences using voice input does not require special skill, the operator must constantly check the kana, and if the operator's voice becomes shaky due to fatigue etc., the display 60b
In some cases, the selection of candidates displayed on the screen may not be properly selected, or the number of candidates may be too large. This method has the disadvantage that the processing speed is not increased due to the increase in processing time.

fdl Purpose of the Present Invention The purpose of the present invention is to eliminate all of the above-mentioned drawbacks, and in the next step, for each candidate extracted according to the similarity obtained by comparing the voice input with a voice dictionary, the connection in a kana string consisting of a combination of multiple candidates is determined. Before checking with a statistical table of all possible forward connections and backward connections, each candidate is compared with a statistical table that ranks the probability of mistaking it for another kana according to the degree of similarity in its recognition. Inserting a means to re-evaluate the pattern by examining it from a statistical perspective, and depending on the matching results, the number of candidates can be further reduced, or if the correct answer is obtained after all attempts to review the priority order, cut out the kana string route after the first one. The present invention aims to provide a means for completely reducing the burden of trial repetition in kana-kanji conversion, as well as eliminating all the troublesome operations associated with sequential selection of kana characters in the past, thereby improving processing speed.

(el) Structure of the Invention A voice input sentence creation system in which the voice recognition unit compares unknown input voice with the feature value standard bang registered in advance in a voice dictionary, and outputs the voice bang data obtained by the voice recognition unit as a sentence containing kanji and kana by the kana-kanji conversion unit , an analysis and matching means analyzes the input speech in spectrum time series and compares the feature values point by point with all standard batons to obtain a degree of similarity, and one or more candidates are selected from the top according to the similarity obtained by the analysis and comparison means Means for selecting and holding each candidate phrase: Arranges the probability of occurrence of the candidate to be mistaken for another kana according to its degree of similarity. A means for re-evaluating, a means for evaluating the connectivity of each candidate by comparing it with the second and third statistical tables for preceding connections and subsequent connections with other kana characters. Extract and select by B7' (Morphological extraction of kana string candidates based on combinations of top candidates/and means to cut out each combination, grammar ke reference, words, kanji dictionary @i collation, \cut out length and shortness evaluation, and kana The controller is equipped with a means for converting the entire string into kanji and a means for evaluating the combination route of the extraction nest position, and the control section applies the input speech sound to the analysis and matching means of the speech recognition section and calculates the similarity score for each candidate of speech data. - causing the re-evaluation means to perform reduction and/or correction of the complement number, rank, and/or correction for the candidates selected and arranged by the selection holding means; In the first step, the top candidates are selected based on their connection validity in the kana string, and are made to perform all English in the first step.Furthermore, in the kana-kanji conversion section, a cutting means, a kana-kanji conversion means, and a route evaluation are performed for the kana string route consisting of the combination of the top candidates. This can be achieved by providing a voice input sentence creation method which has all the characteristics of executing the second means of obtaining the highest priority kana string route by means or mutual repetition thereof, and outputting the sounds of sentences containing kanji and kana.

G) Embodiment of the Invention An embodiment of the invention will be described below with reference to the drawings.

FIG. 2 is a block diagram of a voice input sentence creation method in an embodiment of the present invention, FIG. 3 is an example of its processing procedure, and FIG.
The figure shows an example of data of the first statistical table in the voice input text creation method. In the figure, 0 is a control unit, 10 is a speech recognition unit, 11 is an analysis matching unit, 12 is a candidate selection holding unit, 13 is a candidate re-evaluation unit, 14 is a connectivity evaluation unit, and 20
21 is a kana-kanji conversion unit, 21 is a kana string extraction unit, 22 is a kana-kanji conversion unit, 23 is a route evaluation unit, 30 is a storage unit,
31 is a control program, and 32 is control data.

33 is a voice dictionary, 33a=n standard batan, 34 is a first statistical table, 35 is a second statistical table, 36 is a third statistical table, 37 is a word dictionary, and 38 is a kanji dictionary.

In one embodiment of the present invention, the control unit O in FIG. Each component is controlled according to the control data 32, and the input voice is recognized according to the procedure shown in FIG. 3, and the contents are converted into kanji and kakana characters and output. The storage unit 30 stores a control program 31. Control data 3
In addition to 2, there is a monosyllabic standard pattern ``l'' that corresponds to unspecified speakers.
J: About 10 for each kana character obtained in the learning mode of a specific speaker.A speech dictionary that registers and accumulates an average standard batan tl from 0 training samples, and a certain input speech is analyzed by the analysis verification means. According to the function, the spectrum is obtained as a series-based feature quantity analysis output and held in the collation function, and compared and collated with the standard button 33a=n of the speech dictionary 33, and a plurality of kana candidates are obtained in order from the one with the highest degree of similarity. Of course, if the similarity is as high as 0.99 or 0.95, there is no problem as a single candidate, but if the similarity is not much different, for example 0.
．． 61, 0.5B.

Conventionally, when candidates such as 0.57, 0.53, 0.51, and 0.48 are obtained, all of the top four candidates are sent to the connectivity evaluation means 14, and all of these candidates are sent to the connectivity evaluation means 14. The combinations were evaluated one by one, but in this embodiment, the following candidate re-evaluation means is used to compare the candidates with a first statistical table 34 that statistically arranges the percentage of candidates mistaken for other kana according to the degree of similarity. and examine it carefully. Although the 1st to 3rd statistical tables are shown as a single block in Fig. 2, the contents are similar to the standard batan, and for each kana, as shown in the example data in the 1st statistical table in Fig. 4. It shall be prepared.

In the example of FIG. 4, when candidate A's similarity is 50, B is 17/, D is 15/, and C is 107. .

Indicates that there is a probability of error at a rate of 50 50. Therefore, there is a zero possibility that A50, which is not displayed on the table, will be mistaken for A50. Of course, if the significant difference is small, the common data 1 may be used for similarity correspondence within a certain range.

As a first example, in the candidate ranking obtained by the candidate selection holding means, even in the fifth ranking, the appearance rate of the possibility of error for the first candidate, for example A50, in the first statistical table 34 is the 3rd jw.
-, for example, is in rank ■, which is higher than C, and similarly exists in rank ■, which is also omitted from the illustration, but the error rate for the second candidate, for example, B48, is higher than C, which is the third rank, the third rank. The ranking will be corrected by raising IIJ, which was 5th place, by lowering C-kai. As a second example, there is a first candidate, e.g. E80, and only the second candidate F75 exists in the second rank for E80 in the first statistical table 34, and other possibilities are O'f and X is set by the operator. Threshold value, for example, 5 or less, and F75
If only the first candidate E80, for example E70, exists in the second rank and is in a complementary state, the third and subsequent candidates are deleted.

In addition, as a third example, the above-mentioned similarity with no significant difference, for example, 0.61
Even when a large number of candidates (~0.48) are obtained, the information obtained in the first statistical table shows that only the second and third ranks are complemented. In this case, the second rank is 0.58 and the third rank is 0.58.
44 by 57 and delete the first rank Senzen. In this way, in the past, candidates were selected only according to the degree of similarity obtained by the analysis and matching means 11, but the candidate re-evaluation means 13 performs the first statistical table comparison to reduce the number of candidates and perform GV and ranking adjustment meetings. Then, candidate information is sent to the connectivity evaluation means 14. The above description has been made in the direction of deleting candidates from a plurality of candidates, but in some cases, there are few candidates and the candidate re-evaluation means 13 may add candidates with low similarity (which would otherwise be excluded). Next, the connectivity evaluation means 14 compares the kana combination of the two characters before and after each candidate with the preceding connected candidate xl l 'j2 + xs using the second statistical table, and for example, if the candidate is "a", then A, x2 a+X3
Evaluate the appearance rate of A..., reduce the preceding candidates, and use the third statistical table to connect the candidates to succeed)'++Yt+
Check with y3... and find Anne 1. A Ytr ay...
Evaluate all appearance rates of and reduce subsequent candidates. Furthermore, candidates are additionally reduced by AND conditions and/or OR conditions in the connection evaluation from before and after. The kana-kanji converter 2 converts the kana example candidates created by combining the candidates obtained in this way.
Send to 0.

The kana string cutting means 21 in the kana-kanji conversion unit 20 performs large cutting between syllable spaces or numbers, symbols, etc. to form a conversion processing unit, and then performs a conversion process according to the evaluation of the similarity information between each candidate constituting the entire kana string. A plurality of routes are set by giving an array route and its priority based on each combination of candidates.

For each route, the kana string candidates are accessed one by one to the word dictionary 37, and each morpheme and/or combination thereof is checked for grammatical connections.The length of the morphemes and their combinations is confirmed by comparing and evaluating them with the connection information in the group word dictionary 37. Check out the personality.

The other kana sequence roots are evaluated repeatedly in the same way, and the most probable priority canal-1 is determined based on, for example, the minimum number of morphemes, and sent to the kana-kanji conversion means 22. The kana-kanji conversion means 22 receives all the priority kana route candidates, accesses the kana character string to kanji dictionary 38, converts it into a sentence containing kanji and kana, and outputs it. Means 22 obtains a valid morpheme n
This function performs formal back processing using another array route when there is no such array.

In this embodiment, the candidate re-evaluation means 13 evaluates the kana candidates inputted to the male-kana connectivity evaluation means 14 using statistical data regarding the possibility of mistaking them for other kana candidates configured as described above, and reduces the number of candidates. Since the reduction and rank correction are performed, the selection of appropriate candidates can be obtained through rational review, whereas the selection of candidates is only based on the similarity comparison of the speech patterns obtained by conventional analysis and matching means. It is possible to completely reduce the processing of each subsequent function, including food processing.

<g) As described in detail, according to the present invention, unnecessary candidates can be eliminated by adding statistical methods to the selection of candidates that would have been overlooked only by comparing the similarity based on the analysis output in Juo. Since it is possible to effectively reduce or remove incorrect candidate settings, it is possible to reduce the burden on subsequent processing functions, improve processing speed, and obtain a voice input text creation method with a low rate of erroneous conversion.

[Brief explanation of drawings]

FIG. 1 is a conceptual diagram of a voice input sentence creation method in a conventional method and an embodiment of the present invention, FIG. 2 is a block diagram of a voice input sentence creation method in an embodiment of the present invention, and FIG. 3 is an example of its processing procedure. 4 and 4 show data examples of the first statistical table in the voice input text creation method. In the figure, O is the control unit. 10 is a speech recognition unit, 11 is an analysis verification unit, 12 is a candidate selection and holding unit, 13 is a candidate re-evaluation unit, 14 is a connectivity evaluation unit, 20 is a kana-kanji conversion unit, 30 is a storage unit, 31 is a control block, 32 is control data, 33 is a voice dictionary, 3
3a”-n is the standard baton. 34 is the first statistical table, 35 is the second silk table, 3
6 is a third statistical table, and 40 is a document editing department. Reference numeral 50 represents an output open side, 60a represents a printer, and 60b represents a display. Dormitory 1lfI Dormitory?呵I- Y-3 Utaho 4 Drunkenness

Claims (1)

[Claims] In a voice input text creation system, the voice recognition unit compares the unknown input voice with the standard feature value registered in the voice dictionary.In the voice input sentence creation system, the voice recognition unit outputs the input voice as a sentence containing kanji and kana by the all-kana-kanji conversion unit. An analysis matching means that analyzes the voice in spectrum time series and compares the feature values point by point with a standard bang to calculate the similarity, and selects and retains all candidates from the top one or more according to the 36a similarity obtained by the analysis matching means. means for each candidate,
Means for re-evaluating the ranking arrangement or weighting of the candidates by comparing them with a first statistical table that arranges the occurrence rate rankings of the probability of the candidate being mistaken for another kana corresponding to the low degree, and for each candidate, prior connection and Regarding connections with other subsequent kana characters, the second. Means for evaluating connectivity by comparing with the third statistical table, means for cutting out kana sequence candidates by morphemes and/or combinations thereof, referring to grammar, matching words in kanji dictionaries, and \cutting length. The control unit is equipped with means for evaluating and converting kana strings into kanji and means for determining the combination route of cutting units. The degree of similarity is calculated for each candidate, and the number of candidates and the ranking are reduced by the re-evaluation means for the candidates selected and arranged by the selection holding means.
and correction is performed, and the connection evaluation means selects the top candidate based on its connection validity in the kana string consisting of the arrangement combination of the candidates and executes the first step, and further the kana-kanji conversion unit The cutting means, the kana-kanji conversion means and the route evaluation means, or the second means which obtains the highest priority kana string route t-V by repeating these means, are executed for the kana string route consisting of a combination of candidates, and the sound of the sentence containing kanji and kana is executed. A voice input text creation method characterized by output.