KR100931786B1 - Speech recognition method according to Korean spelling - Google Patents

Abstract

The present invention relates to a speech recognition method according to Korean spelling utterance, to improve the accuracy of speech recognition and to provide a speech recognition method according to Korean spelling utterance for feeding back a speech recognition result to a user in a short time.

To this end, the present invention, the speech recognition method applied to the speech recognition system, comprising: a first step of building an isolated word phonetic dictionary for recognition when the Korean spelling speech; A second step of receiving a search target vocabulary from a user according to spelling pronunciation; And a third step of outputting a recognition result in text by combining the recognized spellings based on the isolated word phonetic dictionary into syllables and vocabulary, and used for a speech recognition system.

Description

Method of Korean utterance recognition using spelling pronunciation}             

1 is an exemplary configuration of a voice recognition system to which the present invention is applied.

Figure 2 is a flow diagram of an embodiment of a voice recognition method according to the Korean spelling utterance according to the present invention.

3 is a flowchart illustrating an example of implementing a voice recognition service according to the present invention;

4 is a detailed flowchart illustrating an embodiment of a specific service implementation of FIG. 3 according to the present invention;

* Explanation of symbols for the main parts of the drawings

11: end point detector 12: feature extractor

13: Viterbi Explorer 14: dictionary

According to the present invention, a Korean syllable, a word, or spoken using a total of 31 isolated words including 19 Korean consonants, 10 vowels, and one word that distinguishes between syllables and syllables, and one word that indicates the end of the input voice. It relates to a speech recognition method for recognizing a sentence.

The voice recognition telephone information system is a system that recognizes and provides related information when a certain information is requested by inputting a voice through a telephone network. For example, in the securities information guidance system, when the voice recognition system is used, the user can listen to the stock information of the company by simply speaking the company name without having to memorize a code number corresponding to the desired company name.

In general, a widely known method of speech recognition is to use a Hidden Markov Model (HMM). Here, the Viterbi search is performed by the speech recognition process, which compares the difference between the HMM constructed by pre-training the candidate words to be recognized and the characteristics of the currently input voice to determine the most similar candidate word. It's a process.

However, since the speech recognition process requires a large amount of calculation, as the number of words that can be recognized increases, the error rate of speech recognition increases and the response time slows. Therefore, in the past, the user cannot accurately and quickly provide the desired information. there was.

That is, in the current Korean speech recognition system, the maximum number of words that can be recognized is about 2,000 to 3000 words. In the case of services with more than 10,000 words to be recognized, or many similar names among the words to be recognized, such as automatic name service, it is difficult to provide practical services with the current Korean speech recognition system. . In addition, even in the case of a service having a recognition target vocabulary of 2 to 3000 words or less, there is a problem that it is difficult to obtain a correct voice recognition result when the surrounding noise is relatively large such as a subway, a theater ticket office, and a high-speed bus terminal.

The present invention has been proposed to solve the above problems, and an object thereof is to provide a speech recognition method according to Korean spelling to increase the accuracy of speech recognition and feed back a speech recognition result to a user in a short time. .

According to an aspect of the present invention, there is provided a speech recognition method applied to a speech recognition system, comprising: a first step of constructing an isolated word phonetic dictionary for recognition in Korean spelling speech; A second step of receiving a search target vocabulary from a user according to spelling pronunciation; And a third step of outputting a recognition result in text by combining the recognized spellings based on the isolated word phonetic dictionary into syllables and vocabulary.

The present invention further includes a fourth step of searching for information using the text vocabulary of the recognition result after the third step, and outputting the retrieved text result to the speech synthesizer.

In addition, the present invention further includes a fifth step of automatically dialing a telephone number as a result of voice output after the fourth step is performed according to a user's request.

delete

delete

delete

delete

The present invention utilizes a total of 31 isolated words including 19 Korean consonants, 10 vowels, and one word to distinguish between syllables and syllables, and one word to mark the end of the input voice. By recognizing, Korean spelling speech recognition method has the advantage that can recognize an unlimited number of syllable speech, word speech, or sentence speech using only 31 isolated words. In particular, in the voice recognition 114 telephone number guide service, since the vocabulary length of a human name or a business name is one word, a highly accurate telephone number guide service can be provided by using the method proposed by the present invention. In addition, since only 31 isolated speech words need to be recognized, a large vocabulary-independent isolated word speech recognizer that is robust to ambient noise and has a recognition rate of nearly 100% is possible.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.                     

1 is an exemplary configuration of a voice recognition system to which the present invention is applied.

As shown in FIG. 1, the voice recognition system to which the present invention is applied includes an endpoint detector 11 for detecting both endpoints of the input voice when voice is input, and extracting features of the voice from the detected voice signal. A feature extractor 12 and a Viterbi searcher 13 for selecting words most similar to the words registered in the dictionary 14 using voice feature values.

Here, the dictionary 14 is composed of 19 Korean consonants ("Gimyeo (a)", "Ssanggi", "Nieun" (b), "Di (d)", "Ssangdi (귿)", "Leeul ( ㄹ) "," Mieum (ㅁ) "," Bieup (ㅂ) "," Ssangbi-eup (,) "," Shiot (ㅅ) "," Ssangsiot (ㅆ) "," Ieung (ㅇ) "," Ji (읒) ) "," Shinji "," Chi "," Ki "," Thi "," Fi "," Hi "(19)), Vowel 10 ("Ah", "Ya", "Oh", "Female", "Oh", "Yo", "Wo") "," Yu (ㅠ) "," U (ㅡ) "," I (ㅣ) "are ten, and one word that distinguishes syllables and syllables (here" and "is used as an example) and input voice A total of 31 isolated word phonetic dictionaries, one for the end of the word (here using "more" as an example).

The Viterbi searcher 13 determines the most similar candidate word by comparing the difference between the HMM constructed by pre-training the candidate words to be recognized and the characteristics of the currently input voice.

Since the configuration and operation of the speech recognition system to which the present invention having the configuration as described above is applied are only known techniques in the art, detailed description thereof will be omitted herein.                     

Then, a method for quickly and accurately serving a service having tens of thousands of words such as 114 telephone number guide service through a wired / wireless telephone without using an operator (guide) will be described. For example, the 114 telephone number guide service recognizes the spelling of Korean, performs a text search on the recognized result, and then sends the search result phone number as a voice through a voice synthesizer or dials directly. To make it possible.

2 is a flowchart illustrating an example of a voice recognition method in accordance with Korean spelling.

First, the user (voice talker) has 19 Korean consonants ("Lee Mo", "Ssanggi", "Nee", "Di" (")," Ssang Di "," Leeul) (ㄹ) "," Mie (ㅁ) "," Bi-eup "," Ssangbi-eup "," Siot "," Doublet "," Yeung "," Ji " "", "Ssangji", "Chi", "Ki 읔", "Tit", "Phi", "Hi". 10 vowels ("ah", "ya", "er", "female", "o", "yo", "woo" ) "," Yu (ㅠ) "," U (ㅡ) "," I (ㅣ) "are ten words, and one word that distinguishes syllables and syllables (here" and "is used as an example) If one word is used to mark the end of the voice (here, "ideal" is used as an example) and the input vocabulary is pronounced (201), the pronunciation is an endpoint detector (11), a feature extractor (12), and a Viterbi searcher. (13) and the phonetic pronunciation (202 ~ 204), if the syllable ending sound (for example, "and" is used here) When the expression is stored (206), the current input syllable is stored (207), and the spelling pronunciation input continuously is recognized (201 to 204) to recognize the end-of-vocab indication sound (for example, "above" is used here). In operation 205, the recognition results are output as text by combining the previously recognized spellings into syllables and vocabulary (208).

For example, to enter "Hong Gil-dong", "Hi 읗", "O", "Yeung", "And", "Grim", "Yi", "Lee", "And", "Diet", "O" Spell pronunciation, "eung", "ideal" sequentially.

Therefore, when the user enters "Hong Gil-dong", when the user utters "Hi", "O", "Yeung" in sequence, and "Yes", which is the syllable ending sound (206), "Hyep, Oh, Yieung" is the syllable unit. If it is recognized and stored (207), then "base", "this" and "ri" are sequentially uttered, and the end syllable "and" is uttered (206), "base, this, ri" is the syllable unit. If it is recognized and stored (207), "Dih", "O", and "Ieung" are sequentially uttered, and "Voice", which is the end of the vocabulary, is fired (205). After recognizing as "Hi, Oh, Lee Eung", "Gimyeo, Lee, Lee Eul", and "Dong, Oh, Lee Eung", "Hong Gil Dong" with each syllable unit is output (208).

3 is a flowchart illustrating an example of implementing a voice recognition service according to the present invention.

First, a user who wants to use the service spells pronunciation using 19 Korean consonants, 10 vowels, 1 word that distinguishes between syllables and syllables, and 1 word that indicates the end of the input voice. Guide system usage (301). At this time, the announcement is clearly clarified so that the user can recognize the syllable ending sound (for example, "and" here) and the lexical ending sound (for example, "more than"). do.

As a result, the vocabulary in the form of text output as a result of FIG. 2 is outputted as a voice to a user using a Korean unlimited speech synthesizer, and the recognition result vocabulary is corrected. .

If the inputter confirms that the recognized vocabulary is incorrect (303), the user may request to re-enter the spelling pronunciation (304) to match the Korean speech recognition process 302 (see FIG. 2) and the recognition result check process 303 based on the spelling pronunciation. Repeat until it is confirmed, and if it is confirmed that the recognition result is correct (303), after performing the information search using the text vocabulary of the recognition result (305), using the unlimited speech synthesizer to the inputted text results Voice guidance or automatic dialing during search information guidance (306).
The present invention as described above can be equally applicable to not only recognizing Korean spelling utterances of Korean, but also recognizing Korean spelling utterances of foreign languages. This is the case when the foreign language is spelled in Korean. For example, when searching Japanese web information on a wired or wireless telephone that cannot input Japanese even without a Japanese input program or a Japanese input program, Hiragana (Katakana) If the phonetic spelling is spoken as it is, it can be recognized and converted into Japanese keywords, and based on this, the searched Japanese web document can be converted back to Korean web and provided. In other words, if you want to search for Korean “share price guide” as a keyword, you can say “Kabuka Annai” and spell it as “Kabuka Annai”. By converting the text into "カ ブ カ ア ン ナ イ" and web-converting Japanese web documents with `` カ ブ カ ア ン ナ イ '' as a search word back to Korean, so that Japanese web information can be retrieved from a wired or wireless telephone that does not have Japanese input program or Japanese input program. can do.

4 is a detailed flowchart illustrating an exemplary embodiment of the service implementation of FIG. 3 according to the present invention, and illustrates a wire / wireless telephone number search example of a Hanmir telephone number guide service using a Korean voice recognition method using spelling voice.

First, when the service user calls the voice recognition Hanmir telephone number guide service, the service system transmits a guidement about the service providing method (401) (see step "301" in FIG. 3). If the user knows how to use the service, the voice can be input at any time during the announcement of the service announcement.

If the correct business name is known (402), the spelling input for the business name is confirmed (403), and then the correct address is entered. . Using these two search terms, a search is made in the Hanmir telephone number guide service (405).

The business name of the search results is classified into the top-level sectors of Hanmir (currently, in the Hanmir telephone number search service, "living services", "medical services", "real estate, leasing", "travel, accommodation", "information and communication", " Research, development "," electricity, gas and water business "," food service "," financial, insurance "," exercise, entertainment "," trade, commodity brokerage "," transportation "," press and advertising ", "Social Services", "Business Related Services", "Educational Culture Services", "Foreign Institutions, Social Organizations", "Manufacturing, Wholesale, Retailing", "Administration, Defense, Social Security", "Metal, Recycled Materials, Processing" (22) There are 22 types of industry categories ("forestry, enrichment, mining, fishing" and "construction") using the unlimited Korean voice synthesizer to guide and request selection (406). All phone numbers are guided through an unlimited voice synthesizer and requested to be selected (407). If there is only one phone number of the same company name in the station number, the voice synthesizer guides the phone number and automatically connects it after 2-3 seconds (408), but if the same company name in the same phone number is 2 If more than one is found, all are guided and requested to select (407), and the service terminates by automatically connecting to the selected phone number after 2-3 seconds (408).

On the other hand, if you do not know the exact business name, spell out the type of business name (must be familiar with the name of the business handled by Hanmir) (409) and then confirm that the correct input, and then spelled the address of the business name (410) The same procedure is followed. Using these two search terms, a search is made in the Hanmir telephone number guide service (411). Based on the search result, the type of industry according to the classification of the lowest unit of the business of Hanmir is requested by using the Korean voice synthesizer to guide the selection (412). The resulting business name will be guided and requested for selection (413). Then, all the telephone numbers of the search results are guided by the voice synthesizer and requested to be selected (407). When the phone number of the same company name in the selected phone number is 1, the phone number is guided by using a voice synthesizer and automatically connected after 2-3 seconds (408). However, if two or more telephone numbers of the same business name in the same telephone number come out, all are guided and requested to be selected (407). The service is terminated by automatically connecting to the selected phone number after 2-3 seconds (408).

In order for the user to successfully receive the phone number of the business name to be searched without knowing the exact business name, the user must be familiar with the business classification of the Hanmir phone number search service.

The method of the present invention as described above may be implemented as a program and stored in a computer-readable recording medium (CD-ROM, RAM, ROM, floppy disk, hard disk, magneto-optical disk, etc.).

The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes are possible in the art without departing from the technical spirit of the present invention. It will be clear to those of ordinary knowledge.

 As described above, the present invention can effectively perform large-word speaker-independent speech recognition using only 31 isolated words. For example, in the 114 telephone number guide service using tens of thousands of words, the name of a person or a business name is the word. Since the length is one word, it is very useful for similar services, and especially, it can be used as an input technology of voice recognition internet information retrieval service using a wired / wireless phone.

In particular, when driving a car with a lot of noise, it is impossible to use your eyes and hands, so when you want to call a person / company you need to enter the name of the person or company you want to call only by voice. Since the present invention needs to recognize only 31 isolated word voices, it is possible to perform a voice recognition function that is strong against noise, thereby providing a very useful input method.

Claims (8)

In the speech recognition method applied to the speech recognition system, A first step of constructing an isolated word phonetic dictionary for recognizing Korean spelling; A second step of receiving a search target vocabulary from a user according to spelling pronunciation; And A third step of outputting a recognition result in text by combining the recognized spellings based on the isolated word phonetic dictionary into syllables and vocabulary Speech recognition method according to the Korean spelling utterance, including. The method of claim 1, After performing the third step, a fourth step of searching for information using the text vocabulary of the recognition result, and outputting the retrieved text result to the speech synthesizer Speech recognition method according to the Korean spelling utterance comprising a more. The method of claim 2, A fifth step of automatically dialing a telephone number as a result of voice output according to a user's request after performing the fourth step Speech recognition method according to the Korean spelling utterance comprising a more. The method according to any one of claims 1 to 3, The isolated word phonetic dictionary, Korean consonants and vowels, syllable separators to distinguish syllables and syllables, and vocabulary end separators to mark the end of the input voice. The method of claim 4, wherein The isolated word phonetic dictionary, 19 consonants in Korean ("Lim.", "Ssanggi", "Nee", "Di" (")," Ssang-di "," Lee "," Meeum " ㅁ) "," Bi-eup "," Ssangbi-eup "," Shikot "," Shuangsi-cloth "," Lee-eung "," Ji-ji "," Ssang-ji " ) "," Chi "," key "," tea "," pi ", and" hi ", 10 vowels (" ah " ) "," Ya "," U "," Wo "," O "," Yo "," U "," Yu " "," U (ㅡ) "," yi (ㅣ) "10), 1 syllable separator, and 1 vocabulary delimiter, a total of 31 isolated word phonetic dictionaries Recognition method. delete delete delete

Images