JPS59119434A - Processing device for japanese language inputted by voice - Google Patents

Abstract

PURPOSE:To execute easily a conversion of an inputted voice to a ''KANA'' (Japanese syllabary) train, and to obtain automatically a ''KANA'' and Chinese character mixing sentence by adding automatically a small character, and using it as the following candidate. CONSTITUTION:When a Japanese sentence to be printed out by a printer is uttered toward a microphone 10, a voice recognizing part 14 recognizes it in each separate syllable. Subsequently, as for those syllable groups, the feature is extracted by the recognizing part 14, compared and collated with a feature of each syllable registered in advance, and the first candidate having the highest degree of similarity, the second candidate having a high degree of similarity in the next place, etc. are selected, and they are outputted as the first recognition result S1 from the voice recognizing part. This first recognition result is received, a small character and a particle are added to the first recognition result S1, and they are outputted as the second recognition result S2.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field of the Invention The present invention relates to a speech recognition method in an apparatus for creating a Japanese document through speech input.

Prior Art and Problems Currently, keyboard input is the mainstream for Japanese word processors, but voice input types have also been developed. Currently, voice input type word processors recognize the input voice, create a kana string, convert the kana string to kana-kanji, and create a kana sentence mixed with kanji.A voice recognition unit is added to the key human-powered word processor. It can be said that it was done.

Figure 1 shows an overview of a voice input word processor, with 10
1 is a microphone, 12 is an operation console, 14 is a voice recognition unit, 1
6 is a kana-kanji converter, 18 is a document editor, and 20 is a CRT.
22 is a speaker, 24 is a voice synthesis/output control unit, and 26 is a printer. When a Japanese sentence to be printed out by the printer 26 is uttered into the microphone 10, the speech recognition section 14 recognizes the half-words for each syllable (that is, kana) and passes it to the kana-kanji conversion section 16 as a kana string. At this point, the validity of the word is determined according to the word dictionary index and the grammatical connection relation check, and the word is converted into a sentence containing kana and kanji. The document editing section 18 performs line breaks, page breaks, sentence replacement, and other editing, and the edited kana-kanji text is displayed on the display 20 via the speech synthesis/output control section 24, and sent to the printer according to instructions from the operation console 12, etc. It is printed out from 26.

There are two types of speech recognition: monosyllable recognition type and word recognition type, and the latter is of course preferred, but currently only specific words such as commands can be recognized as words, and non-specific words cannot be recognized. It is usually limited to monosyllables, or individual kana. There are also various problems with monosyllable recognition. For example, in Japanese, the particles ``ha'', ``he'', and ``wo'' are pronounced and written differently; they are uttered as ``wa'', ``e'', and ``o'', but are written as ``ha''.
``to'' and ``wo''. Such devices cannot be adapted simply by increasing the accuracy of speech recognition. Therefore, in the conventional system, prior to utterance, a specific operation, such as a keystroke, is performed to specify that the next syllable is a particle whose notation is different from the utterance. For example, if you want to manually pronounce ``recognize speech'', manually input the individual syllables ``o'', ``n'', ``se'', and ``i'', then press the particle 1 designation key, and then ``o''. ” “ni” “n” “shi” “ki” “su” “
Enter the syllables "ru". The speech recognition unit treats a syllable inputted by voice immediately after the particle specification key is pressed as a particle, and if it is recognized as "o", it is determined to be the particle "wo".

Also, some kana characters are treated as lowercase letters, such as ``ya'' and ``yo''.
"Yu""Tsu""A""I""U""E""
"O" is that. In this type of case, it is not possible to determine whether or not a letter is treated as a lowercase letter simply by increasing the speech recognition accuracy, so in the conventional system, a lowercase letter designation key is pressed. For example, if you want to say "more effort" by voice, say "i", then press the lowercase letter designation key, then "tsu", "so".
Say “u”, “no”, “do”, and “su” individually,
After pressing the lowercase letter designation key again, continue with ``yo'' and ``ri''. This specifies that ``tsu'' and ``yo'' are lowercase letters, and the speech recognition a1; can perform the recognition ``Issho Nodoriyo <'''.

In such a method, it is necessary to specify the particle or lowercase letter before uttering it, and it is necessary to judge whether or not to specify at the front in the flow of a series of utterances. The drawback is that input operations are obstructed.

OBJECTS OF THE INVENTION The present invention aims to improve the above points and to make it possible to perform speech input processing of particles and lowercase letters extremely simply and easily.

Structure of the Invention The present invention provides a speech input Japanese processing device that recognizes input speech and converts it into a Japanese sentence, including a speech recognition unit that recognizes human input speech in units of syllables and converts it into kana strings, and the speech recognition unit. The present invention is characterized in that it receives a kana string outputted by the kana string, and if there is a kana that can be a particle or a lowercase letter in the kana string, the candidate addition section adds that particle and lowercase letter as the next candidate for the kana. This will be explained in detail with reference to examples.

Embodiment of the Invention The main structure of the present invention is shown in FIG. As shown in this figure, in the present invention, a candidate addition section 28 is provided after the speech recognition section 14. The group of syllables input through the microphone 10 is extracted with features in the speech recognition unit, and compared with the features of each syllable registered in advance.The first candidate with the highest degree of similarity (closest distance) is selected, followed by The second candidate with a high degree of
etc. are selected, and these are outputted from the speech recognition unit as the first recognition result S1. The candidate addition unit 28 receives this first recognition result and performs the processing shown in FIG.
The next recognition result is output as S2.

Assuming that the primary recognition result S1 has up to the n-th candidate for each syllable, the candidate addition unit 28 that receives this checks the n candidates (hereinafter referred to as recognition targets), and checks the characters that can be lowercase letters (lowercase targets). Ya, yo, yu, tsu, a, i, tsu,
Check to see if ``Eng'' and ``O'' are among them. If this judgment result in block 30 exists (Y), check whether the character is "e" or "o" in block 32.34, and if it is neither "e" nor "o", the character ( First, the lowercase letter of the character is added to the next candidate (K+1 candidate) where 1<K<n. And this (K+1
) The distance information of the candidate is the same as that of the candidate. However, K+1
)>n, (K11) The candidate is not added. If the judgment result in block 34 is "o" (Y), add "1" to the next candidate for "o", and then add a lowercase "o" to the next candidate for "wo". do. These additional “wo” and “o” should be at the same distance as the first “o”.

Next, in block 32, the lowercase letter target is "e" (
If (Y), move to block 36 and check whether there is "he" in the recognition candidates; if (N), add "he" to the next candidate for "e", and then add "he" to the next candidate for "e". Add the lowercase letter "e" to . These additional ``he'' and ``e'' should be at the same distance as the first ``e.'' On the other hand, if the judgment result in block 36 is that there is "he" in the recognition target, the process moves to block 38, and it is determined whether "he" is an earlier candidate than "e" (the distance number is small and the candidate rank number is small). If it's true (Y), we don't add ``e'' again, but simply add a lowercase ``e'' to the next option after ``e''. On the other hand, if "he" is not a faster candidate than "e", "he" is added to the next candidate after "e", and if it is a later candidate? Ni (the distance number is large and the candidate ranking number is large)" is canceled, in other words, the ranking is swapped for "to", and the next candidate for the added "to" is a lower case "e". to add. These additions “he” and “e”
is also the same distance as the first "e".

In the above process, if there is a lowercase letter in the recognition target, that lowercase letter is added as the next candidate for the lowercase target character, or among the lowercase target characters, only "e" and "o" are particles "he" and "wo". There is also a gender, so add it before the lowercase letter,
That is what it is.

If the judgment result in block 30 is that there are no lowercase letters,
Check whether there is a particle. However, since the particles "e" and "o" are checked using the route of blocks 32 and 34 described above, only the remaining particles "ha" are checked. That is, in block 40, it is checked whether "wa" is within the recognition target, and if not, no candidate is added. A direct route through these blocks 30 and 40 is one in which there is no lowercase letter target or particle target in the recognition target, and the candidate addition unit 28 does nothing with the recognition target, and the first recognition result S1
is output as is as the secondary recognition result.

If the judgment result in block 40 is that "wa" is present in the recognition target, the process moves to block 42, and the process checks eight times to see if there is "wa" in the recognition target, and if not, it is added as the next candidate for "wa". do. If there is "ha", check whether "ha" is a faster candidate than "wa", if so, then ma-1, otherwise "ha"
is set as the next candidate after ``wa'', and the later candidate ``ba'' is deleted, that is, the candidate ranking of ``ba'' is changed.

Distance information of the candidate added here (referred to as the (m+1)th candidate) (well, any information located between the distance information of the mth candidate and the (m+2)th candidate is fine, but as described above It is easy to set the same candidate as the m-th candidate, and there is an advantage that it is easy to determine whether the candidate is a candidate added later.

If particles and lowercase letters are automatically added to the primary recognition results in this way, you can simply look at the primary recognition results displayed on the display and give an instruction to select the "next candidate" and you will be able to see the lowercase letters of the corresponding characters. It can be easily transformed into particles. That is, the output of the candidate addition section 28 is sent to the speech synthesis/output control section 2 in FIG.
In addition to displaying the display 20 via step 4, if there is something that should be converted into a particle or a lowercase character in the displayed secondary recognition result S2, simply press the "Next candidate" key to replace it with the next candidate. . For example, in the above-mentioned example of "Recognize voice", when you utter "o", "n", "se", "i", and "o" and the words are displayed on the display 2o, the next candidate key on the console 12 will be pressed. If you say ``ni'', ``nJ rLJ rki'', ``su'', or ``ru'', the display on the Disfly 2o will change to ``onsei wo.'' Also, in the example of "more effort" mentioned above, if you say "i" and "tsu" and press the next candidate key when this is displayed, the display will change to "-i", followed by "so" and "u".
If you say "no", they will be displayed, and then "do"
When you say ``su'' or ``yo'' and press the next candidate key when it is displayed, the display will change to ``Hassou no Doryo'', and if you say ``<'' at the end, the display will change to ``Hassou no Doryoku''. "become.

With this operation, there is no need to check in advance and select and press the required lowercase letter key or particle key.You can simply look at what is displayed, know that it is different from what you want, and then press the next common candidate key. It is extremely easy to process, and the burden placed on the speaker is light. Furthermore, lowercase letters,
If a particle has been added as a next candidate, it can be automatically selected by a grammar check, etc., instead of being selected manually. As one method for this automatic selection, a separately filed multi-route method can be used. Next, the outline will be explained with reference to FIG.

In Figure 4, it is assumed that ``Calculate the total amount'' was inputted by voice, and the first candidate of the voice recognition result was ``Sougaouo, Kenisansuru,'' and the second candidate was ``Ofukakuwo, Heihakuus.'' . Note that each character of the second candidate corresponds to each character of the first candidate, and the degree of approximation is 2.
It is the second one. The particle "wo" was added to the second candidate in the process described above. Since the voice input is syllable by syllable, it is input manually by saying ``so'', ``tsu'', ``ga'', etc., and the comma ``,'' is input by saying the word ``comma''. Since the word input is limited to punctuation marks and other specific words (such as control words), the recognition rate is high and a second candidate is not necessary. Next, extract the conversion unit and import the first kana sequence separated by commas.

Next, the similarity information between the candidates is evaluated. In this evaluation, in this example, the degree of approximation of the second candidate to the first candidate So, Tsu, Ga is low (distance Ll of the first candidate So, Tsu, Ga, second candidate O for L2, L3), power The distance β1. R2, R3 (R)-Ll, R2-L2 are above the threshold values), so they were dropped, and only the second candidate with a high degree of approximation was left.

Of course, the second choice? lIi Leave ``wo'' as is. Note that if there is a second candidate for each syllable (kana), there are many different kana strings to consider when finding the correct answer. For example, even if the kana string is limited to two characters, ``so'', if there is a second candidate for ``off'', there will be four combinations: ``sou'', ``sofu'', ``oh'', and ``off''. 4th...If there were candidates, the combination would be an astronomical number,
In the process shown in FIG. 4, target candidates are screened out as much as possible based on the degree of approximation. In the example of FIG. 4, there are two second candidates, so there are four combinations, but the figure shows only two for simplicity.

Next, we start setting the route and assigning priority, in this example,
``Sogakuo'' is ranked 1st, and ``Sogakuwo'' is ranked 2nd. The kana string with rank 1 leads to the route to, and the kana string with rank 2 leads to the old route. These A.

In B-route 1-, kana-kanji conversion and its processing are performed. The processing content for both routes A and B is the same; first, a word dictionary is looked up, and corresponding words are selected in descending order of accuracy using the longest match method. Since this is the longest match, the one in which the dictionary headword (morpheme) and the kana string match from the beginning to the end as far as possible is considered to be the most accurate. In this example, the morphemes selected in this way were, as shown in the figure, Kusa Ryo Uo, Nail Closure, So, So, for the A route, and the total amount for the B route 1-. Perform a grammar check on these. The word dictionary has a part of speech for each morpheme, and there is also connection information about which part of speech this morpheme can be connected to, so we use these to check whether the morphemes can be connected to each other.

Next, a maximum likelihood evaluation is performed, and a high evaluation is given to items with a large length (length, number of characters), and items with a high frequency of use.

Next, we will conduct a censored evaluation. Kana strings are converted into morphemes based on the longest match, but the longest match is not necessarily the correct answer.
If morphemes are erroneously converted, subsequent morphemes will not be connected. When we enter a dead-end branch, we discontinue the morphemization in that branch, change the previous morphemization to another one, two, etc., and redo the analysis.
Efforts will continue to arrive at the correct answer, but this evaluation provides a criterion for deciding where to make the cut. The branches are evaluated based on depth, reading length, etc., and those whose evaluation values exceed the set values are discontinued. The next linguistic evaluation is a process to extract and filter out words that may be OK in a grammar check, but do not exist in Japanese, are meaningless, etc. To give a specific example, when converting kana to kanji such as ``subekudoryokusuru'', it may be converted to ``subekudoryokusuru'', but if there are three kanji in a row, each still has an l morpheme. This will be disqualified based on the criteria that it cannot be ``Ma\''.

Next, determine the highest priority route, in this example, is route A better?
Decide whether route B is better. Check whether this contains unused words, whether there are many morphemes, etc. 6A
If the root "Sugata" is a word that has never been used in this system, there will be unused words, and if the "Total amount" of route B is a word that has been used, there will be no unused words, and the latter will be rated higher. Also, the number of morphemes is 3 for the A route, and 2 for the total number of B routes, so it is better to have fewer morphemes. From these results, the kana strings of route A had a higher degree of nearness in the input speech, but the analysis results within the chain line frame showed that the kana strings of route B had a higher evaluation, and the priority was higher than that of route B. "Total amount" is changed to ■, and "Sugata Uo" of route A is changed to ■. Next, "total amount" with priority ■ is converted into kana to "sou gakuo", and since this is equal to the input data, it is determined that no error was introduced during analysis, and "total amount" becomes the output data. . In this way, if you create multiple kana sequences that have a possibility of being correct and select the one with the highest accuracy through grammatical analysis, the probability of reaching the correct answer is high, and the voice input sentence can be converted into a kana-kanji mixed sentence without human intervention. This eliminates the need to display the information on the display and ask the speaker for confirmation and correction at the stage of the kana sequence. Of course, if the speaker is asked to confirm and correct the kana string at the stage, the input data shown in FIG. 4 will be limited to the first candidate, and the kana-kanji conversion can be expected to be easier, faster, and more accurate.

As described in detail, according to the present invention, particles and lowercase letters are automatically added and used as the next candidate, making it easy to convert input speech into a kana string. The advantage is that it is possible to automatically convert sentences into kana-kanji mixed sentences without requiring the speaker to check or correct them.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an overview of a voice input Japanese word processor, and FIG. 2 is a block diagram showing the main parts of the present invention.
FIGS. 3 and 4 are flowcharts showing the processing procedure. In the drawing, 14 is a speech recognition section, and 28 is a candidate addition section.

Claims (1)

[Claims] A speech input Japanese processing device that recognizes input speech and converts it into Japanese sentences includes a speech recognition section that recognizes input speech in units of syllables and converts it into a kana string, and a speech recognition section that converts the input speech into a kana string, and a kana string that the speech recognition section outputs. 1. A voice input Japanese language processing device, comprising: a candidate addition unit that adds a kana that can be a particle and a lowercase letter to the input kana string as a next candidate for the kana.