KR20210020294A - Method And Apparatus for Providing Speech Recognition of Word Unit or Sentence Unit - Google Patents

Abstract

Disclosed are a method and device for recognizing speech in units of words or sentences. The method according to an embodiment of the present invention checks partial matching of a word corresponding to inputted speech, extracts a candidate word by combining a plurality of preset algorithms, and generates a final candidate word for filtering the candidate word, and checks whether a word with the smallest value is a standard word to output the word as the standard word after comparing approximations of the final candidate words if there are a plurality of final candidate words.

Description

TECHNICAL FIELD [Method And Apparatus for Providing Speech Recognition of Word Unit or Sentence Unit}

The present embodiment relates to a method and apparatus for recognizing speech in units of words or sentences.

The contents described below merely provide background information related to the present embodiment and do not constitute the prior art.

In general, speech recognition is a series of phonological and linguistic information that extracts phonological and linguistic information from the acoustic information included in the speech, and after modeling the vocabulary registered in the speech vocabulary dictionary through the speech recognizer, finds the most similar data and makes it react. It's a process.

Therefore, in the entire process of speech recognition, a large or small number of words to be recognized occupies an important part in determining recognition rate improvement and recognition speed.

The current speech recognition technology goes beyond the level of isolated word recognition for small-volume vocabulary, and studies on continuous speech recognition for large-scale vocabulary are active, and a lot of effort is being put into processing the vocabulary to be recognized.

In this regard, as large-scale speech vocabulary is being recognized as a recognition target, the speech recognition rate has emerged as a large problem. This is because the recognition rate is relatively high when the number of words to be recognized is small, while the recognition rate decreases significantly as the number of words to be recognized increases.

Currently, in the case of the recognition rate of more than 1,000 words for speech recognition, it is announced as 90-95% in a refined laboratory environment, but in real life, 80% is due to the surrounding environmental noise and channel noise in various communication networks and communication devices themselves. It is less than that. There is a problem that a successful commercial service is not possible with this level of recognition rate.

In this embodiment, a partial match for a word corresponding to an input voice, a candidate word is extracted by combining a plurality of preset algorithms, a final candidate word for filtering the candidate word is generated, and there are a plurality of final candidate words, An object of the present invention is to provide a method and apparatus for speech recognition in units of words or sentences that compares approximate values of final candidate words and then determines whether a word having the smallest value is a standard word and outputs a standard word.

According to an aspect of the present embodiment, in a word-by-word speech recognition method, the method includes: receiving a word-based speech input and converting it into word text through Speech-To-Text (STT); A pre-processing step of classifying tokens for the word text, checking different meanings of the tokens, performing alphabetic processing and numeric processing based on the result of performing exception processing, and then generating a final input word combining tokens; A candidate selection process of confirming partial match of the final input word, extracting a candidate word by combining a plurality of preset algorithms, and generating a final candidate word filtering the candidate word; When there are a plurality of final candidate words, a final candidate filtering process of comparing the approximate values of the final candidate words, checking whether the word having the smallest value is a standard word, and outputting a standard word. Provides a way.

According to another aspect of the present embodiment, in a sentence-by-sentence speech recognition method, the method includes: receiving a sentence-by-sentence voice input and converting it into sentence text through Speech-To-Text (STT); A pre-processing step of classifying tokens for the sentence text, checking different meanings of the tokens, performing alphabetic processing and numeric processing based on a result of performing exception processing, and then generating a final input word combining tokens; An input sentence classification and processing process of classifying and processing the final input word as a search reference sentence; A process of extracting a partial match for food nouns when extracting a food noun starting or ending with an input word in the final input word, extracting only when the input word is a complete match; A food noun partial match verification process of indicating whether or not there is a food noun that partially matches the food noun partial match result through true and false values; A combined word extraction process of combining words in order and checking whether the combined word is a valid word when the food noun partial match verification result is an input sentence that is not divided by a conjunctive shooting survey; Generates an initial candidate word based on the final input word, partial match word, and all match words, selects the final candidate word from the initial candidate word when the initial candidate word exists in an internal database, and selects the initial candidate word A group extraction process of selecting the final candidate word by using the final input word if it does not exist in the internal database; A word retrieval process of checking partial coincidence of the final input word and selecting a final candidate word by comparing approximate values and consonants using a plurality of preset algorithms; Providing a sentence-by-sentence speech recognition method comprising: converting a dialect stored in a final result and another meaning result among the final candidate words into an appropriate standard language to remove redundancies in the result of each input value; do.

As described above, according to the present embodiment, a partial match for a word corresponding to an input voice, a candidate word is extracted by combining a plurality of preset algorithms, a final candidate word for filtering the candidate word is generated, and the final When there are a plurality of candidate words, after comparing approximate values of the final candidate words, there is an effect of outputting the standard word by checking whether the word having the smallest value is a standard word.

1 is a diagram showing a preprocessing algorithm for speech recognition in units of words according to the present embodiment.
2 is a diagram showing another algorithm for checking meaning in speech recognition for each word according to the present embodiment.
3 is a diagram showing an exception processing algorithm for speech recognition in units of words according to the present embodiment.
4 is a diagram showing an algorithm for converting alphabetic representation in word-based speech recognition according to the present embodiment.
5 is a diagram showing a number representation conversion algorithm for speech recognition in units of words according to the present embodiment.
6 is a diagram showing a candidate selection algorithm in speech recognition by word according to the present embodiment.
7 is a diagram illustrating a partial matching algorithm in speech recognition for each word according to the present embodiment.
8 is a diagram illustrating a Levenshtein Distance Algorithm when speech recognition is performed in units of words according to the present embodiment.
9 is a diagram illustrating a weight formation table in speech recognition by word according to the present embodiment.
10 is a diagram illustrating a Hashing Algorithm in speech recognition by word according to the present embodiment.
11 is a diagram showing a combination of two algorithms in speech recognition by word according to the present embodiment.
12 is a diagram showing an approximate value comparison algorithm for speech recognition in units of words according to the present embodiment.
13 is a diagram illustrating an algorithm for comparing consonants in speech recognition by word according to the present embodiment.
14 is a diagram illustrating an algorithm for checking standard words in speech recognition by word according to the present embodiment.
15 is a diagram illustrating a mobile-based performance evaluation layout in speech recognition by word according to the present embodiment.
16 is a diagram illustrating a standard word mismatched word during word-by-word speech recognition according to the present embodiment.
FIG. 17 is a diagram illustrating a word that is inconsistent with dialect during speech recognition by word according to the present embodiment.
18 is a diagram illustrating an input data processing algorithm when speech recognition is performed by sentence according to the present embodiment.
19 is a diagram showing a preprocessing algorithm for speech recognition by sentence according to the present embodiment.
20 is a diagram illustrating an exception processing algorithm for speech recognition by sentence according to the present embodiment.
21 is a diagram illustrating an alphabetic notation conversion algorithm when speech recognition in sentence units according to the present embodiment.
22 is a diagram illustrating an algorithm for converting number representation in speech recognition in sentences according to the present embodiment.
23 is a diagram showing an input sentence classification and processing algorithm when speech recognition by sentence according to the present embodiment.
24 is a diagram illustrating a word extraction algorithm in speech recognition for each sentence according to the present embodiment.
25 is a diagram illustrating a candidate selection algorithm in speech recognition for each sentence according to the present embodiment.
26 is a diagram illustrating a partial matching algorithm in speech recognition for each sentence according to the present embodiment.
27 is a diagram illustrating a Levenshtein Distance Algorithm when speech recognition is performed by sentence according to the present embodiment.
28 is a diagram illustrating a weight formation table in speech recognition for each sentence according to the present embodiment.
29 is a diagram illustrating a Hashing Algorithm in sentence-by-sentence speech recognition according to the present embodiment.
FIG. 30 is a diagram illustrating a combination of two algorithms in speech recognition by sentence according to the present embodiment.
31 is a diagram illustrating an approximate value comparison algorithm for speech recognition by sentence according to the present embodiment.
32 is a diagram illustrating a consonant comparison algorithm in sentence-by-sentence speech recognition according to the present embodiment.
33 is a diagram illustrating a mobile-based performance evaluation layout when speech recognition is performed by sentence according to the present embodiment.

Hereinafter, this embodiment will be described in detail with reference to the accompanying drawings.

Word-based speech recognition

A terminal equipped with a speech recognition application processes a food noun as a result of STT (Speech-To-Text) input in a manner that receives and processes a word-based voice input.

The voice recognition application is installed on the terminal. Terminals include electronic devices such as smart phones, tablets, laptops, personal computers (PCs), portable multimedia players (PMPs), and wireless communication terminals. It can be a device.

The terminal is (i) a communication device such as a communication modem for performing communication with various devices or wired/wireless networks, (ii) a memory for storing various programs and data, and (iii) a microprocessor for operation and control by executing programs. It is a variety of devices including the like. According to at least one embodiment, the memory is a computer such as a random access memory (RAM), a read only memory (ROM), a flash memory, an optical disk, a magnetic disk, or a solid state disk (SSD). It may be a readable recording/storing medium. According to at least one embodiment, the microprocessor may be programmed to selectively perform one or more of the operations and functions described in the specification. According to at least one embodiment, the microprocessor may be implemented entirely or partially as hardware such as an Application Specific Integrated Circuit (ASIC) having a specific configuration.

Pre-processing for word-based speech recognition

As a preprocessing process, the terminal equipped with the voice recognition application checks whether there is a space in the input value output through the STT as shown in FIG. 1, and if there is space, classifies the input word based on the space and defines it as a token.

If there are spaces, the terminal equipped with the voice recognition application classifies words as spaces and checks if there is any other meaning, and checks if there is an exception word, and if there is no space, checks if the input value has a different meaning. Check for exception words.

A terminal equipped with a voice recognition application stores the corresponding original word if the input value is a different meaning of a word for checking different meanings, and ignores it if not. The terminal equipped with the speech recognition application checks the standard word in the final candidate filtering part of the stored original word and informs the user whether to find the word. The terminal equipped with the voice recognition application checks whether there is an exception word, and if there is an exception, it processes the exception, and if it does not, it performs the alphabet process.

The terminal equipped with the voice recognition application outputs the word changed due to exception or alphabet processing by replacing the corresponding token part of the input with the changed word if there is a space, and outputs the result of the exception or alphabet processing as it is if there is no space. After that, the number existing in the input value is changed through number processing.

Terminals equipped with a voice recognition application finally generate the final input word by executing a token combination based on the letter order and spacing of each word when there is a space between the multiple input words generated through number processing. Conversely, if there is no space Multiple input words that have passed through number processing without token classification are generated as final input words.

Token classification in word-based speech recognition

The terminal equipped with the voice recognition application classifies each word based on the spacing when the STT recognition result includes spacing, so that exception processing and alphabet processing can be partially performed.

Checking different meanings in word-by-word speech recognition

As shown in FIG. 2, the terminal equipped with the voice recognition application checks whether there is a word that matches the other meaning column of the other meaning (standard language, dialect) table in the input value. If there is a matching word, the terminal equipped with the speech recognition application stores the corresponding original meaning word, converts it into a standard word in the standard word check step in final candidate filtering, and outputs it together with the final result word.

Exception handling in word-based speech recognition

The terminal equipped with the voice recognition application is an algorithm that handles words that do not exist in the DB of STT and compares them with words in the exception (standard language, dialect) table and changes them to the final word. As shown in FIG. 3, the terminal equipped with the voice recognition application compares the input with the exception result column in the exception (standard language, dialect) table, converts it into a corresponding original word, and outputs it.

Alphabet processing for word-based speech recognition

The terminal equipped with the speech recognition application is an algorithm for changing the alphabet into Korean when there is a word containing the alphabet in the input value. As shown in Fig. 4, the terminal checks whether the alphabet exists in the input value, and if there is an alphabet, Finds the Korean notation of the alphabet in the alphabet table, converts it, and prints it.

The terminal equipped with the speech recognition application uses an alphabet processing algorithm to compare the alphabet in the alphabet table for each alphabet and change it to the basic Korean pronunciation. The conversion continues until it is confirmed that no alphabet exists in the input value.

Number processing in word-based speech recognition

A terminal equipped with a voice recognition application is an algorithm for converting numbers into Korean when there is a word containing numbers in the input value, taking into account the unit of the number and the Korean notation, and in case there are sometimes food names containing numbers. The result of the number that has not been changed in Korean notation is also saved, and up to 5 change result values are output.

As shown in Fig. 5, the terminal equipped with the voice recognition application checks whether a number exists when an input is received, and if it does not exist, outputs the result as it is without any special processing, and determines whether there is a continuous number if there is a number. do.

If there are consecutive numbers, the terminal equipped with the voice recognition application outputs 4 results considering the units and numbers, and if not, outputs 2 results considering only the numbers. At this time, in both cases, the terminal equipped with the voice recognition application converts using the words for each notation and the word for each unit existing in the number table, and outputs the result without changing the number.

Token combination for word-based speech recognition

Terminals equipped with speech recognition applications are tokenized by classifying them based on spacing through 1.1 when spaces are included as a result of STT recognition, and when tokenized words are changed through 1.2, 1.3, and 1.4, as shown in [Table 1]. The final input result is created by combining the changed words based on the order of the changed input results and deleting spaces and duplicate words.

At this time, the terminal equipped with the voice recognition application generates an initial approximation value according to the character length, which is generated in the same manner as in [Equation 1], and this is called SCORE.

The terminal equipped with the voice recognition application stores the initial approximate score along with the corresponding final input word. When the token classification process is not performed because there is no space, the terminal equipped with the voice recognition application sets the initial approximation to 0.

Candidate selection for word-based speech recognition

A terminal equipped with a voice recognition application may have one or more preprocessed words prior to candidate selection. A terminal equipped with a speech recognition application is a process applied to each final input word for candidate selection, as shown in FIG. 6, and searches for a word that partially or all matches the final input word through partial match check and finds the initial candidate word. Create

If there is an initial candidate word, the terminal equipped with the speech recognition application selects a final candidate word from the initial candidate word, and if there is no initial candidate word, the final candidate word is selected using the final input word.

In order to select the final candidate word, the terminal equipped with the speech recognition application first matches the words resulting from the Hashing Algorithm and the Levenshtein Distance Algorithm when the initial approximation value created in the preprocessing is 0. Add -1 to the approximation. If it is not 0, the Hashing Algorithm is not used and the approximate value is stored as it is.

After that, the terminal equipped with the speech recognition application reduces the number of candidate words through an approximate comparison process, and if the number of candidate words is more than two, it performs consonant comparison and selects the final candidate word by reducing the number of candidate words once more.

Partial match check for word-by-word speech recognition

The terminal equipped with the voice recognition application searches for a word in which the final input word partially or all matches within the words of the standard word table and dialect table as shown in FIG. 7 to confirm partial match. In this case, if there is a partial match, the terminal equipped with the voice recognition application transmits a true value, and if there is no word, a false value.

Word-based speech recognition Levenshtein Distance Algorithm

The terminal equipped with the speech recognition application uses the Levenshtein Distance Algorithm as an algorithm that checks how similar a word among the words in the standard language table and the dialect table is to the preprocessed input word, as shown in FIG. Search for words. At this time, record how many times the letter needs to be changed to become the same as the input word, and define it as an approximation. The approximate value means that the smaller the value, the closest.

For the approximation, the initial approximation value of the final input word is added to the approximate value of the candidate word obtained through the Levenshtein Distance Algorithm.

In the current candidate selection, a terminal equipped with a speech recognition application is placed in a candidate only when the approximate value is 5 or less, and adds the approximate value of each input word that has undergone preprocessing prior to the approximate value of each candidate word.

Hashing Algorithm for word-based speech recognition

은 초성,

은 중성,

은 종성을 나타낸다. A terminal equipped with a voice recognition application executes an algorithm as shown in FIG. 10 using a Hashing Algorithm. First, the terminal equipped with the speech recognition application obtains the hash code of the input word based on the bar shown in Fig. 9, which is a weight formation table based on the pronunciation similarity, and compares the hash code value of words in the standard language table and dialect table. And search for words with the same value. However, the terminal equipped with the voice recognition application is executed only when the original input word and the number of characters are the same and no numbers are included. Here, it means when the initial approximation value is 0 and there is no number. The weight formation table shown in FIG. 9 is

Silver First Star,

Is neutral,

Represents the finality.

Combination of two algorithms in word-based speech recognition

As shown in FIG. 11, the terminal equipped with the speech recognition application has the same word among the candidate words from the Hashing Algorithm and the candidate words from the Levenshtein Distance Algorithm. Add -1 to the approximate value of. However, a terminal equipped with a voice recognition application runs only when the original input word and the number of characters are the same. Here, it means when the initial approximation value is 0.

Candidate words for word-by-word speech recognition Filtering

Approximate comparison for word-based speech recognition

The terminal equipped with the speech recognition application compares the approximate values to each candidate word and stores the candidate word having the smallest approximation value as a final result for approximate comparison, as shown in FIG. 12. At this time, words with the same approximation are also stored.

Consonant comparison in word-based speech recognition

The terminal equipped with the speech recognition application compares the consonant order of the final input word as shown in Figure 13, and stores the most similar word as the final result when there are two or more words after approximation for consonant comparison. However, in the case of a terminal equipped with a voice recognition application, in case of a candidate word for an input word that is not the same as the number of characters in the original word, the order of consonants must be the same to store it in the final candidate result.

Final candidate for word-based speech recognition Filtering

Approximate comparison

When the number of final candidate words is two or more, the terminal equipped with the speech recognition application compares the approximate values of each final candidate word with an approximation comparison algorithm and stores the word with the smallest value as a final result. At this time, words with the same approximation are also stored.

Standard language check

As shown in FIG. 14, the terminal equipped with the speech recognition application checks whether there is a word that matches the dialect column of the SRD table among the final result word and the original word of another meaning word found in the preprocessing step. If there is a matching word, the corresponding standard word is output along with the final result word. If there is no match, only the final result word is output.

Word-based recognition rate

Recognition rate test environment

In order to test the recognition rate by word, a terminal equipped with a voice recognition application gives three utterance opportunities, saves it if it is not the same as the test word, and checks the word that did not appear within three times.

Word-based speech recognition application UI

The original word area ① on the application execution screen shown in FIG. 5 outputs words that exist in the standard table and dialect table one by one.

The speech recognition result word area ② on the application execution screen outputs the result of the user reading the original word.

Check that the processing result area ③ on the application execution screen has been processed properly after pressing the Save or Cancel button.

When the microphone button ④ on the application execution screen is pressed, the user's voice is input.

The Back⑤ button on the application execution screen moves to the previous word.

The forward ⑥ button on the application execution screen moves to the next word.

The Save ⑦ button on the application execution screen saves when the original word and the voice recognition result word are different.

The Save Cancel ⑧ button on the application execution screen cancels saving if it is saved incorrectly.

Recognition rate result

In the case of a standard word, 4 out of 1578 words are inconsistent, showing a hit rate of about 99.7%, and the standard word inconsistent words are as shown in FIG. 16.

In the case of dialect, 21 out of 661 words are inconsistent and have a hit rate of about 96%, and when the standard words are the same, they are 77, and in this case, the hit rate is about 88%. The dialect mismatch words are as shown in FIG. 17.

Speech recognition by sentence

Sentence-based speech recognition receives voice input in one sentence unit, and receives input for the amount of one meal.

The voice recognition application is installed on the terminal. Terminals include electronic devices such as smart phones, tablets, laptops, personal computers (PCs), portable multimedia players (PMPs), and wireless communication terminals. It can be a device.

The terminal is (i) a communication device such as a communication modem for performing communication with various devices or wired/wireless networks, (ii) a memory for storing various programs and data, and (iii) a microprocessor for operation and control by executing programs. It is a variety of devices including the like. According to at least one embodiment, the memory is a computer such as a random access memory (RAM), a read only memory (ROM), a flash memory, an optical disk, a magnetic disk, or a solid state disk (SSD). It may be a readable recording/storing medium. According to at least one embodiment, the microprocessor may be programmed to selectively perform one or more of the operations and functions described in the specification. According to at least one embodiment, the microprocessor may be implemented entirely or partially as hardware such as an Application Specific Integrated Circuit (ASIC) having a specific configuration.

Input data processing for speech recognition by sentence

A terminal equipped with a voice recognition application receives voice input as a sentence unit for sentence unit voice recognition, and receives input for the amount of one meal. The structure of the input data processing algorithm is as shown in FIG. 18.

Pre-processing for speech recognition by sentence

As shown in FIG. 19, for pre-processing, the terminal equipped with the voice recognition application separates the input sentence from the STT based on spaces and checks whether there is an exception word. A terminal equipped with a voice recognition application performs token classification, exception processing, number processing, and alphabet processing during pre-processing.

First, the terminal equipped with the voice recognition application divides the input sentence based on the spacing by token classification, and then checks whether or not it partially matches the exception word list existing in the exception word table based on the order of each word.

A terminal equipped with a voice recognition application checks if a word partially matches the exception word by combining it with the word after the word and partially matches the exception word list. At this time, if the terminal equipped with the voice recognition application does not partially match, the previous word that was partially matched goes to the exception processing process, and the words that did not match partially match the exception word list again. (The criterion for partial match is “Is there any exception word that begins with that word?”)

This repetition structure is to find out by combining each word in the same manner as above when examining the exception word, since most of the cases are spaced apart from each other in the result input in sentence units. A terminal equipped with a voice recognition application converts a matching word from the exception word list into the corresponding original word.

Thereafter, the terminal equipped with the voice recognition application converts alphabets and numbers that could not be converted through the alphabet and number exception processing into Korean notation, stores the converted result, and sends it as an output value.

Token classification in sentence-level speech recognition

A terminal equipped with a voice recognition application classifies each word based on the spacing as a result of STT recognition, so that exception processing and alphabet processing can be partially performed.

Exception handling for speech recognition by sentence

A terminal equipped with a voice recognition application is an algorithm that handles words that do not exist in the DB of STT and compares them with words in the exception (standard language, dialect) table and changes them to the final word. For exception processing, the terminal equipped with the voice recognition application compares the input with the exception result column in the exception (standard language, dialect) table and converts it into a corresponding original word and outputs it if it matches.

Alphabet processing for speech recognition by sentence

The terminal equipped with the voice recognition application is an algorithm for changing the alphabet into Korean when there is a word containing the alphabet in the input value. As shown in Fig. 21, the terminal checks whether the alphabet exists in the input value, and if there is an alphabet, The Korean notation of the alphabet is found in the alphabet table, converted, and printed.

The terminal equipped with the voice recognition application uses an alphabetic notation conversion algorithm to compare each alphabet with the alphabet in the alphabet table and change it to the basic Korean pronunciation. Until it is confirmed that no alphabet exists in the input value, the terminal converts it.

Number processing in sentence-level speech recognition

A terminal equipped with a voice recognition application is an algorithm for converting numbers into Korean when there is a word containing numbers in the input value, taking into account the unit of the number and the Korean notation, and in case there are sometimes food names containing numbers. The result of the number that has not been changed in Korean notation is also saved, and up to 5 change result values are displayed.

As shown in FIG. 22 for number processing, the terminal equipped with the voice recognition application checks whether a number exists when an input is received, and if it does not exist, it outputs the result as it is without any special processing, and if there is a number, consecutive numbers Determine if there is.

If a terminal equipped with a voice recognition application has four consecutive numbers, it outputs four results that consider the unit and number representation, and if no, it outputs two results that consider only the number representation. At this time, in both cases, the terminal equipped with the voice recognition application converts using a word for each notation existing in the number table and a word for each unit, and includes the result of not changing the number.

Classification and processing of input sentences for speech recognition by sentence

In the terminal equipped with the voice recognition application, the input sentences that have passed through the pre-processing process are separated based on spaces and have the pre-processing result for each word. The order is the same as the order of the input sentences.

In a terminal equipped with a voice recognition application, the process of separating the input sentences that have undergone pre-processing based on the conjunctive fire investigation is the investigation standard sentence classification process.

The terminal equipped with the voice recognition application divides the sentences based on the words with the conjunctive fire investigation, and processes and reassembles the words based on the words as shown in FIG.

Sentence classification based on survey for speech recognition by sentence

The terminal equipped with the voice recognition application compares the input sentence that has passed through the pre-processing process for the search criteria sentence classification with the conjunctive fire investigation list, and separates words based on the word when there is a word with conjunctive fire investigation at the end of the word.

Processing for speech recognition by sentence

A terminal equipped with a voice recognition application recombines words separated by survey criteria into a single word through a process of processing. The process of processing is shown in FIG. 7. First, a terminal equipped with a voice recognition application treats words separated by survey criteria as a single word, and extracts food nouns from the word, deletes ending words such as "eaten", "eaten", and "eats", and accesses them. Carry out the shooting investigation removal process. After that, the terminal equipped with the voice recognition application is processed into a single word through a word combination process. At this time, the terminal equipped with the voice recognition application moves to the next word when the food noun extraction is successful, and enters the process of deleting the closing word when it fails. If deletion fails even in this process, the connection fire investigation and removal process is performed.

Food noun extraction

A terminal equipped with a voice recognition application checks whether a word literally exists in the food noun table for the food noun extraction process. At this time, the terminal equipped with the voice recognition application checks whether a word starting with the word or a word ending with the word exists in the food noun table. If it exists, the terminal equipped with the voice recognition application is used when combining the word as it is, and if it does not exist, it performs the process of deleting the closing word and removing the connection fire investigation.

Delete closing word

A terminal equipped with a voice recognition application checks whether there are closing words such as "eaten", "eaten", and "eaten" literally in order to delete the closing word.

A terminal equipped with a voice recognition application is to remove the verb “eaten” from the guide to check if there is a closing word. If there is a closing word, the word is deleted and excluded from the word combination process.

Elimination of conjunctive fire investigation

The terminal equipped with the voice recognition application checks and removes the search existing in the connection fire investigation table to see if it exists at the end of the word for the procedure to remove the connection fire investigation, and checks the words that could not be filtered out during the process of extracting the food name and deleting the ending word. Perform the final procedure.

Word combination

A terminal equipped with a voice recognition application is a process that only goes through words with conjunctive fire investigation for word combination, and combines words extracted through all the processes of extracting food name, ending word, and removing conjunctive fire investigation based on the order of input sentences. do.

Word extraction during unit speech recognition

In order to extract words, the terminal equipped with the voice recognition application matches and extracts the words created through the input data extraction process with words existing in the food noun table as shown in FIG. 24 for word extraction.

First, the terminal equipped with the voice recognition application determines whether it is a word divided by the connection shooting survey, and in the case of the connection shooting survey, checks whether there is an exception word that could not be converted due to the connection shooting survey in the process of extracting the input data, and processing the exception. If it is not performed once more, the result word is extracted using the process of partial match extraction of food nouns and word search.

A terminal equipped with a speech recognition application is a process created in case a speaker said an incorrect word for the process of extracting partial match for food nouns, and “a word that begins with that word or a word that ends with that word” is the standard.

A terminal equipped with a voice recognition application uses a word search process if the word cannot be extracted during the food noun matching process.

A terminal equipped with a voice recognition application finds a word that is most similar to an input word in the word search process, and can more flexibly cope with the result caused by poor STT results or poor pronunciation of the speaker.

In the case of the input sentence divided by the connection shooting investigation, the terminal equipped with the voice recognition application outputs the result obtained through the above three processes as the final result.

On the other hand, if it is not divided into contact shooting investigation, the terminal with the voice recognition application has already checked the exception word in the process of extracting the input data, so the word search process or the combination word extraction and food according to the number of words are not processed without exception processing. The noun partial match extraction process is performed. This is because in the case of input sentences that are not divided by conjunctive shooting investigation, since there is no standard to classify the sentence, it was judged that the situation that would lead to incorrect search results if the word search process was used without classification according to the number of words.

If the number of words is more than one, the terminal equipped with the voice recognition application executes a combined word extraction process to find words that can be combined based on the input order of words. After that, the terminal equipped with the speech recognition application extracts the result word using the process of extracting partial match for food nouns. In this case, if there is a combined word, the terminal equipped with the voice recognition application executes a group extraction process to organize the correlation between the combined word and the parent word lists that occupy a part of the combined word. At the end of this process, the terminal equipped with the voice recognition application sends out the result.

A terminal equipped with a speech recognition application uses a food noun partial match and word search process if the number of words is one. This is because a terminal equipped with a voice recognition application has a small number of words that are less likely to result in incorrect search results when using the word search process. First, a partial match extraction process for food nouns is performed to check whether there is a partial match for food nouns, extract the result word, and execute a word search process if there is no food noun. At the end of this process, the terminal equipped with the voice recognition application sends out the result.

The terminal equipped with the speech recognition application converts the dialect existing in the meaning result different from the result obtained through the word extraction process into a standard language, and sends the result as a final result and a different meaning.

Extraction of partial match of food nouns in sentence-by-sentence recognition

A terminal equipped with a voice recognition application extracts food nouns that start or end as an input word for the process of extracting a partial match for food nouns. A terminal equipped with a voice recognition application extracts only when the input word is a single letter and matches it completely.

Verification of partial match of food nouns in sentence-level speech recognition

A terminal equipped with a voice recognition application notifies whether there is a partial match for food nouns through true and false values rather than outputting the partially matched words for the food noun partial match verification process. The criterion at this time is “Is there a food noun that starts or ends with an input word?”, which is the same as the process of extracting partial match of food nouns. Allow.

Combination word extraction for sentence-by-sentence speech recognition

A terminal equipped with a voice recognition application finds an appropriate food noun through the process of combining words in order because there is no standard for dividing the input sentence when the input sentence is not divided by the conjunctive shooting survey. At this time, the terminal equipped with the speech recognition application checks whether the combined word is a valid word by using the food noun partial match extraction process.

Group extraction for speech recognition by sentence

A terminal equipped with a voice recognition application organizes the relationship between the combination word created by the combination word extraction process and the parent word when the number of words is more than one, but the sentence is not divided by the connection shooting survey for the group extraction process.

A terminal equipped with a voice recognition application excludes a word that is partially matched with a combination word when generating a partial match result of a parent word, and generates it without any other restrictions when creating a partial match of a combination word. At this time, the terminal equipped with the voice recognition application performs a process of extracting partial match for food nouns.

Word search in sentence-level speech recognition

A terminal equipped with a speech recognition application is a process of finding a word that is most similar to an input word for word search. As shown in FIG. 25, a partial match check is performed to find a word that matches the final input word and partially or both, and the initial candidate word Create If there is an initial candidate word, the terminal equipped with the speech recognition application selects a final candidate word within the word, and if there is no initial candidate word, the final candidate word is selected using the final input word.

The terminal equipped with the speech recognition application is the method of selecting the final candidate word. First, if the initial approximation value created in the preprocessing is 0, the words resulting from the Hashing Algorithm and the Levenshtein Distance Algorithm are matched. Add -1 to the approximation. If it is not 0, the Hashing Algorithm is not used and the approximate value is stored as it is.

After that, the terminal equipped with the speech recognition application reduces the number of candidate words through an approximate comparison process, and if the number of candidate words is more than two, selects a final candidate word through consonant comparison.

Partial coincidence check for sentence-by-sentence speech recognition

The terminal equipped with the voice recognition application searches for a word in which the final input word partially or all matches within the words of the standard language table and dialect table as shown in FIG. 26 to confirm partial match. At this time, if there is a partial match, the terminal equipped with the voice recognition application transmits a true value, and if there is no word, a false value.

Speech recognition by sentence Levenshtein Distance Algorithm

In a terminal equipped with a speech recognition application, Levenshtein Distance Algorithm is an algorithm that checks how similar a word among words in the standard language table and dialect table is to the pre-processed input word, and searches for candidate words as shown in FIG. do. At this time, the terminal equipped with the voice recognition application records how many times the letter needs to be changed to become the same as the input word, and this is called an approximation. The approximate value means that the smaller the value, the closest.

In the case of an approximation, the initial approximation value of the final input word is added to the approximate value of the candidate word obtained through the Levenshtein Distance Algorithm.

A terminal equipped with a speech recognition application places an approximate value of 5 or less in a candidate selection and stores an approximate value of each candidate word.

Hashing Algorithm for sentence-level speech recognition

A terminal equipped with a voice recognition application is executed as shown in FIG. 29 using a Hashing Algorithm. First, a terminal equipped with a speech recognition application obtains the hash code of an input word based on FIG. 28, which is a weighting table based on pronunciation similarity, and compares the hash code value of words existing in the standard language table and dialect table to the same value. Find words that have However, the terminal equipped with the voice recognition application is executed only when the original input word and the number of characters are the same and no numbers are included. This is when the initial approximation is 0 and there are no numbers.

은 초성,

은 중성,

은 종성을 나타낸다.In the weight formation table shown in FIG. 28

Silver First Star,

Is neutral,

Represents the finality.

Combination of two algorithms in sentence-based speech recognition

As shown in Fig. 30, the terminal equipped with the voice recognition application has the same word among the candidate words from the Hashing Algorithm and the candidate words from the Levenshtein Distance Algorithm, and only when the preprocessed input word and the number of letters are the same. Add -1 to the approximate value of. However, a terminal equipped with a voice recognition application runs only when the original input word and the number of characters are the same. The initial approximation value is 0.

Candidate words for speech recognition by sentence Filtering

Approximate comparison

The terminal equipped with the speech recognition application compares the approximate values to each candidate word and stores the candidate word having the smallest approximation value as a final result, as shown in FIG. 31 for approximation comparison. At this time, the terminal equipped with the voice recognition application also stores words with the same approximation.

Consonant comparison

When there are two or more words after approximate comparisons for consonant comparison, the terminal equipped with the speech recognition application compares the order of consonants of the final input words as shown in FIG. 32 and stores the most similar words as the final result. However, in the case of a terminal equipped with a voice recognition application, in case of a candidate word for an input word that is not the same as the number of characters in the original word, the order of consonants must be the same to store it in the final candidate result.

Final candidate for speech recognition by sentence Filtering

Approximate comparison

When the number of final candidate words is 2 or more, the terminal equipped with the speech recognition application compares the approximate values of each final candidate word with the same algorithm as the approximation comparison process of the approximation comparison algorithm shown in FIG. 31 to determine the word with the smallest value. Save as final result. At this time, the terminal equipped with the voice recognition application also stores words with the same approximation.

Standard dialect conversion for sentence-level speech recognition

The terminal equipped with the speech recognition application performs redundancy in the result of each input value by converting the dialect stored in the final result and other semantic results into an appropriate standard language.

Sentence unit recognition rate for sentence unit speech recognition

Recognition rate test environment for speech recognition by sentence

In order to test the recognition rate by sentence unit, the terminal equipped with the speech recognition application, Three women in their twenties, one by one, tested only words, ate with ~, and ate with ~, gave 3 chances of uttering, and checked words that did not appear more than 3 times.

Test application UI for speech recognition in sentences

The original word area ① on the application execution screen shown in FIG. 5 outputs words that exist in the standard table and dialect table one by one. The speech recognition result word area ② on the application execution screen outputs the result of the user reading the original word. Check that the processing result area ③ on the application execution screen is processed properly after pressing the Save or Cancel button. When the microphone button ④ on the application execution screen is pressed, the user's voice is input. The back button ⑤ on the application execution screen moves to the previous word. The forward button⑥ on the application execution screen moves to the next word. The save button ⑦ on the application execution screen saves the original word and the voice recognition result word if they are different. Save Cancel ⑧ on the application execution screen cancels saving if it is saved incorrectly.

Recognition rate result for speech recognition by sentence

The test results by simply reading the food noun are shown in [Table 2].

[Table 3] shows the test results by adding "I ate with ~" after the food noun.

Table [4] shows the results of the test by adding “I ate it” after the food noun.

The above description is merely illustrative of the technical idea of the present embodiment, and those of ordinary skill in the technical field to which the present embodiment belongs will be able to make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present exemplary embodiments are not intended to limit the technical idea of the present exemplary embodiment, but are illustrative, and the scope of the technical idea of the present exemplary embodiment is not limited by these exemplary embodiments. The scope of protection of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present embodiment.

Claims (13)

In the word-based speech recognition method,
A process of receiving a word-based voice input and converting it into word text through Speech-To-Text (STT);
A pre-processing step of classifying tokens for the word text, checking different meanings of the tokens, performing alphabetic processing and numeric processing based on the result of performing exception processing, and then generating a final input word combining tokens;
A candidate selection process of confirming partial match of the final input word, extracting a candidate word by combining a plurality of preset algorithms, and generating a final candidate word filtering the candidate word;
When there are a plurality of final candidate words, a final candidate filtering process of comparing the approximate values of the final candidate words, checking whether the word having the smallest value is a standard word, and outputting the standard word
Word-based speech recognition method comprising a. The method of claim 1,
The pretreatment process is
A token classification process of checking whether there is a space in the word text and, if there is space, classifying an input word based on the space and defining a token as a token;
Another meaning checking process of classifying the word into spaces and checking whether there is a different meaning if there is a space in the word text;
An exception processing step of checking whether there is an exception word in the word text, and checking whether there is an exception word as it is in the input value when there is no space;
An alphabet processing step of storing the original word when the word text has a different meaning, performing exception processing when there is an exception word in the word text, and performing alphabet processing when there is no exception word;
If there is a space in the word text, the word changed due to exception processing or alphabet processing is output by replacing the corresponding token part of the input with the changed word, and if there is no space in the word text, the result of exception processing or alphabet processing is output. A number processing process of changing a number existing in the word text through number processing;
If there are spaces between the plurality of input words obtained through the above number processing, token combinations are executed based on the order of letters and spaces of each word to generate the final input word, and if there is no space, the number is processed without token classification. Token combination process of generating multiple input words as the final input words
Word-based speech recognition method comprising a. The method of claim 1,
The candidate selection process above is
This is a process applied to each final input word, and generates an initial candidate word by finding a word that partially or all matches with the final input word through partial match checking, and if the initial candidate word exists, a final candidate word within the initial candidate word Partial agreement checking process of selecting a final candidate word using the final input word if there is no initial candidate word
If the initial approximation value is 0, the words resulting from the Hashing Algorithm and Levenshtein Distance Algorithm are matched, and if there is the same word, add -1 to the approximation of the same word, and if it is not 0, the Hashing Algorithm is not used. The combination of two algorithms that stores the approximate value as it is
When reducing the number of candidate words through an approximate comparison process, when the number of candidate words is plural, a candidate word filtering process in which the final candidate word is selected by reducing the number of candidate words once more through consonant comparison
Word-based speech recognition method comprising a. The method of claim 1,
The final candidate filtering process is
An approximation comparison process of comparing approximate values of each final candidate word and storing the word having the smallest value as a final result when the number of final candidate words is plural; And
If there is a word that matches the dialect column of the table among the original words of the final result word and other meaning words found in the preprocessing step, the corresponding standard word is output together with the final result word. If there is no match, only the final result word Process of checking the output standard language
Word-based speech recognition method comprising a. In the sentence unit speech recognition method,
A process of receiving a speech input in a sentence and converting it into sentence text through Speech-To-Text (STT);
A pre-processing step of classifying tokens for the sentence text, checking different meanings of the tokens, performing alphabetic processing and numeric processing based on a result of performing exception processing, and then generating a final input word combining tokens;
An input sentence classification and processing process of classifying and processing the final input word as a search reference sentence;
A process of extracting a partial match for food nouns when extracting a food noun starting or ending with an input word in the final input word, extracting only when the input word is a complete match;
A food noun partial match verification process of indicating whether or not there is a food noun that partially matches the food noun partial match result through true and false values;
A combined word extraction process of combining words in order and checking whether the combined word is a valid word when the food noun partial match verification result is an input sentence that is not divided by a conjunctive shooting survey;
Generates an initial candidate word based on the final input word, partial match word, and all match words, selects the final candidate word from the initial candidate word when the initial candidate word exists in an internal database, and selects the initial candidate word A group extraction process of selecting the final candidate word by using the final input word if it does not exist in the internal database;
A word retrieval process of checking partial coincidence of the final input word and selecting a final candidate word by comparing approximate values and consonants using a plurality of preset algorithms;
A standard dialect conversion process of converting a dialect stored in a final result and other semantic results among the final candidate words into an appropriate standard word to remove duplicates in the result of each input value;
Sentence-based speech recognition method comprising a. The method of claim 5,
The pretreatment process,
A token classification process of checking whether there are spaces in the sentence text and, if there are spaces, classifying input words based on the spaces and then defining them as tokens;
Exception handling process of checking whether there is an exception word in the sentence text, and checking whether there is an exception word as it is in the input value when there is no space.'
An alphabet processing step of storing the original word when the sentence text has a different meaning, performing exception processing when there is an exception word in the sentence text, and performing alphabet processing when there is no exception word;
If there is a space in the sentence text, the word changed due to exception processing or alphabet processing is output by replacing the corresponding token part of the input with the changed word, and if there is no space in the sentence text, the result of exception processing or alphabet processing is output as it is, Number processing process that changes the number existing in the word text
Sentence-based speech recognition method comprising a. The method of claim 5,
The input sentence classification and processing process,
A search criterion sentence classification process of comparing the final input word with a conjunctive fire investigation list and, if there is a word with conjunctive fire investigation at the end of a word, separating words based on the word with conjunctive fire investigation; And
Processing process of recombining words separated in the search criteria sentence classification process into one word
Sentence-based speech recognition method comprising a. The method of claim 7,
The processing process;
In the process of classifying the search criteria sentence, it is checked whether or not the separated word exists in the food noun table, and if it exists, it is used when combining the word as it exists in the food noun table, deleting the ending word if it does not exist, and removing the conjunctive fire investigation. The food noun extraction process to perform;
A closing word deletion process of checking whether there is a closing word among words existing in the food noun table and deleting the closing word when there is a closing word to be excluded from the word combination process;
A conjunctive fire investigation removing process of checking and removing whether a conjunctive fire investigation among words existing in the food noun table exists at the end of a word;
A word combination process in which words extracted through all of the food noun extraction process, the closing word deletion process, and the connection fire investigation and removal process are combined based on the input sentence order
Sentence-based speech recognition method comprising a. The method of claim 5,
The group extraction process
When the number of words is not a sentence divided by the conjunctive fire survey and the number of words is more than one, the relationship between the combination word created by the combination word extraction process and the parent word is arranged, and when the result of partial matching of the parent word is generated, the combination word A sentence-by-sentence speech recognition method, characterized in that words that partially match with are excluded and are generated without any other restrictions when generating a partial match of a combined word. The method of claim 5,
The word search process
When the initial approximation value of the final input word is 0, words resulting from the Hashing Algorithm and Levenshtein Distance Algorithm are matched, and when the same word exists, a value of -1 is added to the approximation of the same word, and it is not 0. In case the Hashing Algorithm is not used, the approximate value is stored as it is, the number of candidate words is reduced through an approximate value comparison process, and if the number of candidate words is more than two, the final candidate word is selected through consonant comparison Unit speech recognition method. The method of claim 10,
The word search process
Partial match checking process in which when the final input word searches for a word that partially or all matches within a word of the standard language table or dialect table, a true value is transferred when a partial matched word exists, and a false value is transferred when the divided matched word does not exist. ;
Among the words in the standard word table and dialect table, a candidate word is searched for by examining how similar a specific word is to the pre-processed input word, and how many times the letter needs to be changed to become the same as the input word as an approximation. A Levenshtein Distance Algorithm that records and stores the approximate value of each candidate word by adding the initial approximate value of the final input word to the approximate value of the candidate word obtained through the Levenshtein Distance Algorithm, and raising it as a candidate only if the approximate value is less than 5 in the current candidate selection. Performing process;
Hashing Algorithm execution process of finding a hash code of an input word based on a weight formation table based on pronunciation similarity and comparing it with hash code values of words existing in the standard language table and the dialect table to search for words having the same value;
A combination process of two algorithms that add -1 to the approximate value of the corresponding word only when there is the same word among the candidate words from the Hashing Algorithm and the candidate words from the Levenshtein Distance Algorithm, and only when the number of letters and the preprocessed input word are the same;
A candidate word filtering process of comparing approximate values with each candidate word and comparing consonants to select the final candidate word;
Final candidate filtering process of filtering the final candidate word by comparing approximate values
Sentence-based speech recognition method comprising a. The method of claim 11,
The candidate word filtering process,
Approximate value comparison process in which the candidate word with the smallest approximation value is stored as the final result by comparing the approximate value to each candidate word, and the word with the same approximation value is also stored.
After performing the approximation comparison process, if there are two or more words, the consonant order of the final input word is compared and the most similar word is stored as the final result, and if it is a candidate word of the input word that is not the same as the original word, consonants Consonant comparison process that saves in the final candidate result only when the order of all of them is same
Sentence-based speech recognition method comprising a. The method of claim 11,
The final candidate filtering,
When the number of final candidate words is two or more, comparing approximate values of each final candidate word and storing a word having the smallest value as a final result.

Images