KR20090041923A - Voice recognition method - Google Patents

Abstract

A voice recognition method is provided to model various textual language phenomenons into statistical modeling among various knowledge sources. A morpheme is interpreted for a primitive text language corpus consisting of the separate words of Korean(S201). A morpheme language corpus separated is a separate word generated to morpheme. A word trigram which is the language model consisting of a morpheme unigram about a generated morpheme language corpus as described above, and bigram and trigrams is generated(S202). A first N - best recognition candidate to the maximum N is generated for a voice(S204). Recognition result candidates applying a morph-syntactic constraints are revaluated(S205). A second N-best list generated in above step is revaluated(S206). A final N-best list is generated.

Description

Speech Recognition Method

The present invention relates to a speech recognition method, and more particularly, to a speech recognition method based on a language model for statistically modeling various textual language phenomena among knowledge sources.

The present invention is derived from the research conducted as part of the IT growth engine technology development project of the Ministry of Information and Communication and the Ministry of Information and Telecommunications Research and Development. [Task Management Number: 2006-S-036-02, Task name: Large-capacity interactive dispersion for new growth engine industries] Development of processing voice interface technology].

In general, in constructing a language model used for speech recognition of a large vocabulary, a recognition unit at a morpheme or pseudo morpheme level is set, and a language model is trained based on morphological analysis from a large text corpus.

In the general speech recognition method, a speech of a speaker is converted into text using various knowledge sources such as a language model, an acoustic model, a speech dictionary, an area model, and the like.

In this case, the normal language model process is to define the probability that a specific word string can appear consecutively, and this requires a sufficient amount of corpus to cover or represent all linguistic phenomena of a specific language. . To collect such corpus, we typically consider years of newspaper articles, novels, essays, plays, editorials, internet chat records, e-mail, and various textual materials.

The basic unit of such a language model uses a lexical structure called 'normal word'. For example, in English, since only words separated by spaces are used as the minimum recognition unit, and various statistical information and language models are created based on these, English vocabulary and grammar are relatively easy. It is possible to model the structure.

However, as in the case of Korean, various types of prefixes can be added at any time, and for languages with a lot of ending changes and advanced research, when the language model of word units is created, the memory used for speech recognition explodes. Done. Therefore, in the case of Korean, it is reasonable and realistic to construct a language model with morphemes, which are the minimum units of meaning, or phonemes, which are speech information of words.

The language model used for speech recognition is to obtain a morpheme or pseudo morpheme through morphological analysis for corpus used in establishing a language model, and to express statistically the probability of morphemes appearing adjacent to each other. In general, the probability that a specific morpheme appears in the whole corpus is expressed in unigram (1-gram), the probability that two specific morphemes appear in succession in 2-gram, and the three specific morphemes in succession In this way, tetragrams (4-grams), pentagrams (5-grams), etc. are created, such as trigrams (3-grams) of probability of occurrence, which is called 'n-grams'. This is called.

The trained engram language model is applied directly in the process of speech recognition. Currently, trigram-based search is most commonly used.

Since this method utilizes only statistical information on the concatenated structure of words, it inevitably has a problem that frequently involves semantic or syntactically incorrect recognition results or sequences.

In addition, different speech units and lexical syntactic units result in poor performance for speech events with short durations, such as the hemispherical endings such as 'b' and 'ㄹ'. The reality is that the unit of speech recognition is constructed by using expansion or adjustment.

On the other hand, a given corpus used for learning cannot reflect all vocabulary and grammatical phenomena of Korean language, and it is virtually impossible to calculate probability values used in language models for all dictionary Korean morphemes.

In other words, since the probability value is expected to be high for lexical syntactic events occurring in the limited corpus used for learning, and the bias is too low for events that do not occur, smoothing is performed. Or a technique such as back-off is applied.

A general planarization technique uses 'P (E)', which is a probability value obtained by dividing the number of occurrences of heat of specific morphemes (r) by the number of occurrences (R) of a corpus used for learning, as shown in Equation 1 below. Done.

P (E) = r / R

In general, when the frequency of occurrence of specific morphemes or morpheme fever in the corpus is too low to obtain a proper probability value, they are cut-off. In order to compensate for this process, or for processing unknown morphological columns, a sub-model can be used if a given n-gram (n-gram (W1, W2, ..., Wn)) column does not occur in the corpus. The value of P (Wn | W1, ..., Wn-1) is used. This is called 'back-off'.

As described above, when applying a language modeling technique such as a flattening technique or a back-off applied in the prior art, a problem in which unintended or meaningless insertion of words or morphemes occurs frequently. have.

The technical problem to be achieved by the present invention is a language model that statistically models various text language phenomena among various knowledge sources such as language models, acoustic models, speech dictionaries, domain models, and speech recognition based on the language models. To provide a way.

In addition, an object of the present invention is to effectively control frequent word insertion errors that are semantically or lexical syntactically problematic in the conventional morpheme-based trigram search.

In order to solve this problem, according to the present invention, a language model and a search method based on the same can be used. In this case, the present invention enables efficient use of various information such as words, morphemes, and syllables by using a knowledge base structured to be suitable for establishing a language model in consideration of Korean characteristics.

In other words, the present invention improves the performance of speech recognition by efficiently knowledge of various semantic or syntactic information generated in the morpheme analysis step for a large vocabulary speech recognition system employing a morpheme-based recognition unit. In addition to morphological trigrams, stem-suffix look-up tables based on stem-to-mother combination information and stem-suffix look-up tables are used to reevaluate the strides after speech recognition. Accordingly, the present invention can improve the performance of speech recognition by effectively re-evaluating a plurality of valid recognition candidate sequences and performing error correction.

According to an aspect of the present invention, in a speech recognition method, a corpus learning process for generating a knowledge base for speech recognition from a raw text corpus, and speech recognition using a morpheme as a recognition unit using the knowledge base It provides a speech recognition method comprising a recognition process for generating a final recognition result of the engine.

Here, in the corpus learning process, the morpheme analysis is performed on the raw text corpus to generate the morpheme corpus and the morpheme sequence, and the statistical language modeling is performed on the morpheme corpus to generate the word trigram and the morpheme trigram. And inputting the morpheme corpus to generate the concatenated forms of the mother or suffixes and the actual concatenated endings to form a stem mother reference table. At this time, the morpheme corpus is a morpheme of words of the raw text corpus, and the morpheme column is composed of individual sentences only from the morpheme corpus except the morpheme corpus, and the word trigram is a morpheme unigram, a bigram, and a trigram. A morpheme trigram is generated as a part-of-speech language model consisting of unigrams, bigrams, and trigrams of morpheme parts of speech.

The recognition process may include generating a plurality of primary en-best recognition candidate sequences by performing a trigram search on the spoken voice, and performing a search based on a morpheme trigram of the primary en-best recognition candidate sequences. And performing a second en-best list by performing a re-evaluation on the second en-best list by applying a stem mother reference table to generate a final en-best list.

According to another aspect of the present invention, in the method for recognizing a spoken voice, performing a word trigram search on the spoken voice to generate a plurality of primary en-best candidate sequences, the primary en-best candidate Performing a morphological trigram search on the column to generate a secondary en-best candidate sequence, and performing a stem-mother search on the secondary en-best candidate sequence to generate a final en-best candidate sequence. It provides a voice recognition method.

Here, the generating of the above-described secondary en-best list is to re-evaluate the primary en-best candidate sequence by using the part-of-speech tag sequence of the recognition candidate sequence by applying the morpheme trigram. In addition, generating the above-described final en-best list may include: constructing a stem ending table at the time of searching for the stem-mother, and re-evaluating the secondary en-best candidate sequence based on the configured stem ending table. It includes.

At this time, the step of constructing the stem mother reference table, the step of separating the word into the stem and the mother, grouping as a mother column without separating the mother anymore, defining the mother bond string as a database, and stem Defining the database fields of the parent lookup table.

As described above, according to the present invention, it is possible to effectively control frequent word insertion errors that are semantically and lexical syntactically problematic in the morpheme-based trigram search of the prior art, and effectively re-evaluate a plurality of valid recognition candidate sequences and make errors. By performing the correction, the performance of speech recognition can be improved.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is said to "include" a certain component, it means that it can further include other components, without excluding other components unless specifically stated otherwise. In addition, the terms “… unit”, “… unit”, “… module” described in the specification mean a unit that processes at least one function or operation, which may be implemented by hardware or software or a combination of hardware and software. Can be.

Now, a voice recognition method according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a schematic flowchart of a speech recognition method according to an exemplary embodiment of the present invention.

The speech recognition method according to an embodiment of the present invention is a multi-stage speech recognition method based on morphemes, morpheme parts of speech, and stem word reference table definitions, and as shown in FIG. 1, a word trigram search (S110) and a morpheme It consists of three stages: trigram search (S120) and stem-mother search (S130). Here, the morpheme trigram is applied to re-evaluate the N-best candidate sequence using the part-of-speech tag sequence of the recognition candidate sequence. In this case, when searching for a stem-mother, a stem-mother reference table is formed, and the en-best candidate sequence is re-evaluated based on the configured stem-mother reference table.

In other words, the multi-stage speech recognition method according to an embodiment of the present invention is an insertion error of an unintended or meaningless word or morpheme when applying a language model technique such as flattening or back-off applied in the prior art. To solve the problem of frequent error, it establishes general n-gram language model for learning corpus, calculates morpheme trigram based on morpheme parts of speech, and also calculates word of Korean. The structure is divided into stems (or stems) and affixes, and in particular, it is necessary to generate and use a stem table that explicitly defines the structural joining rules of the mothers.

In the multi-stage speech recognition method according to an embodiment of the present invention, the second and third stages of re-evaluation and ranking of the plurality of recognition candidate sequences (N-best) generated in the first-stage trigram search, which are basic search structures, are ranked. By correcting the error post-processing through speech, it contributes to the improvement of overall speech recognition performance.

Next, a process and principle of generating a stem word table from the Korean word structure according to an embodiment of the present invention will be described.

First, general Korean word structure combining rules can be categorized into stem, prefix + stem, stem + affix, prefix + stem + ending. Among these, in particular, the rule that appears in various ways in Korean is the combination of the endings.

For example, if you consider the representative word 'Go', there is an example that you can perform various applications at the end of a verb or statement as follows, and the term "+ go" is followed by various uses as follows. It performs a syntactic function.

1) + high: subtitle: suffix sequence: + high + high: window + description + lower + high + high

2) + high: suffix sequence: + li + la + high: wet + violet + sea + oh + ri + la + high

3) + high: subtitle: suffix sequence: + high: swell + swell + swell + high

4) + High: Subtitle: suffix sequence: + High + High: Window + New + Lower + High + High

In this regard, the embodiment of the present invention analyzes the lexical syntactic stem and the connection structure of the Korean language and constructs a stem mother reference table by databaseting the combining rule of the mother as follows.

First, words are divided into stems (ie, words and dictates) and endings (ie, surveys, suffixes, etc.).

Secondly, there are morphemes in Korean that derive or re-derive new stems, which, in principle, belong to the stem but can be separated into another stem and the mother is no longer separated. Group them together as a suffix sequence.

Third, this parent binding sequence is defined as a database, which appears uniquely in terms of the headings (Lemma), the morpheme parts, the combination form, and its components.

Fourth, we define the database fields of the stem ending table. That is, in the case of the headword (in the example, '+ high'), the syllable boundary in the word is represented by '+', and the last syllable is set as the heading word for syllable and vocabulary parsing. In the case of the part-of-speech tag to be combined (in the above example, 'to speak'), a combinable form such as a verb or a talk is specified. In the case of the combination information of the part of speech (in the case of the 'suffix sequence' in the above example), the qualities according to self-supporting or dependent morphemes, endings, and endings are defined. In the case of composition (in the example above, in the case of '+ Lee + La + Go'), the syllable composition of the ending is specified. In the case of the word spelling example (in the example above, 'wet + hard + solution + oh + li + la + go'), the word in a completely combined form is illustrated.

The speech recognition method according to an embodiment of the present invention will be described in detail with reference to the flowchart of FIG. 2.

2 is a diagram illustrating a multi-step speech recognition method based on statistical language model learning according to an embodiment of the present invention. Here, FIG. 2 shows a corpus learning process for generating a knowledge base directly used for speech recognition from a raw text corpus, and a final recognition result of a speech recognition engine using morphemes as a recognition unit using these knowledge bases. It shows the recognition process to generate a.

At this time, the morpheme analysis (S201), statistical language modeling (S202), shallow parsing (Shallow Parsing) (S203) is part of the corpus learning process, trigram search (S204), morpheme trigram search (S205), The stage of stem-mother searching (S206) belongs to the recognition process.

Morphological analysis step (S201), by performing a morpheme analysis on the text corpus (Text Corpus) consisting of the words of the Korean language, to generate a morpheme corpus (Morpheme Corpus) that is separated into morphemes. These morpheme corpus are applied directly as a unit of speech recognition. At the same time, a Morpheme Sequence consisting of only parts of speech tags of these morphemes is generated.

The statistical language modeling step S202 generates a word trigram that is a language model composed of morpheme unigrams, bigrams, and trigrams with respect to the morpheme corpus generated in the morpheme analysis step S201. At the same time, a Morpheme Trigram, which is a part-of-speech set language model composed of unigrams, bigrams, and trigrams of morpheme parts of speech, is generated with respect to the Morpheme Sequence generated in the morpheme analysis step S201. .

The important concept of shallow parsing (S203) does not require accurate semantic or part-of-speech interpretation of individual survey suffixes as in the example of "+ Lee + La + Go" described above, but merely functions as a mother in a word. In addition to the lexical interpretation, the terminology and narrative surveys are applied to the stem, which is a part that does not change when the term is utilized, and focuses on the changing form. Only part-of-speech tags.

In the parsing step (S203), a stem ending table is generated by generating a concatenated form of mothers or suffixes and actual connected mothers by inputting morpheme corpus, which is output data of the morphological analysis step (S201).

The trigram search step S204 is a typical search process of a large vocabulary speech recognition that converts a recorded voice into text using an acoustic model, a language model, a speech dictionary, and the like. Generate a first order N-best recognition candidate sequence.

The morpheme trigram search step S205 aims to re-evaluate the recognition result candidates applying Morph-syntactic Constraints. That is, a step of verifying and re-evaluating these en-best lists by performing a search based on a morpheme trigram on the first en-best list generated in the trigram search step S204. The first en-best is reevaluated and the ranking is re-adjusted by scores to generate a second en-best list.

The stem-mother based search step S206 is applied to the stem mother reference table, which is the output data of the shallow parsing step S203, to re-evaluate the second en-best list generated in the stem stem trigram search step S205. To generate the final en-best list.

Table 1 below is a table showing the statistical language processing results of the speech recognition method according to an embodiment of the present invention.

Processing target Data or result to be processed Raw text corpus <s> I never thought that he would appease or dream that his homeland would be Korea. </ s> <s> Korea is a democratic republic. </ s> <s> I've finished my meal. </ s> Stem stem <s> I / nc + / j he / np + / j probation / nc + ha / xsp + li / e also / ma his / mm motherland / nc + / j korea / nq + / j + li / e dream / ma Thought / nc + ha / xsp + d / e not / pv + d / e + da / e </ s> <s> korea / nq + / j demo / nc + republic / nc + / j + da / e </ s> <s> Today / nc rice / nc + / j da / nc ate / pv + / e + da / e </ s> Stem heat <s> nc + j np + j nc + xsp + e ma mm nc + j nq + j + e ma nc + xsp + e pv + e + e </ s> <s> nq + j nc + nc + j + e </ s> <s> nc nc + j nc pv + e + e </ s> Word trigram \ 1-Gram: (Unigram Probability, Word, Back-Off Weight of Word) -1.1027 <s> 0.0469 -1.6046 is 0.0109 -1.6046 Republic 0.0357 ... \ 2-Gram: (Bigram Probability, Baigram, Hundred of Baigram -Off weight) -0.1249 <s> or 0.0000 -0.1249 <s> Korea 0.0000 -0.1249 Probable 0.0000 -0.1249 Republic 2 0.0000 ... \ 3-grams: (trigram probability, trigram) -0.1761 </ s> <s> Korea -0.1761 <s> I -0.1761 <s> Today Bob -0.1761 <s> Korea -0.1761 That is probable ... Stem stem \ 1-gram: (unigram probability, word, back-off weight of the word) -1.1027 <s> -0.4536 -0.7347 e -0.8182 -0.7347 j -0.6096 ... \ 2-grams: (Bigram probability, Back-off weight of bigrams) -0.3010 <s> nc 0.3802 -0.6021 <s> nq 0.4771 -1.0000 e pv 0.6021 -0.6021 je 0.3010 ... \ 3-grams: (trigram probability, trigrams) -0.1761 <s > nc j -0.1761 <s> nc nc -0.1761 <s> nq j -0.1761 e ma nc -0.1761 e pv e ...

The target data processed in the multi-stage speech recognition method based on the statistical language model training according to the embodiment of the present invention as described above and the processing result thereof may be represented as in the example of Table 1.

Raw text corpus, separated by sentence, appends the "<s>" symbol at the beginning of the sentence and the "</ s>" symbol at the end of the sentence.

This (the raw text corpus) is morphologically interpreted and output in the same form as the morpheme corpus. At this time, the morpheme combining information of a given word is represented as a separator of '+', and the part-of-speech tag information of each morpheme is attached with a part-of-speech tag after the '/' symbol.

The morpheme column is intended to form individual sentences only from the parts of speech tag excluding the morpheme from this (morpheme corpus).

The word trigram is a language model used directly in the trigram search step S204, and the speech dictionary includes part-of-speech tag information and is used for the search.

The morpheme trigram is a language model used in the morpheme trigram search step S205.

Table 2 below is a table showing an implementation example of the stem mother reference table according to an embodiment of the present invention.

Processing unit example Processing steps sentence I didn't even think that he would be able to settle, nor that his motherland would be Korea . Text corpus Word I'm going to be a Korean Morphological analysis morpheme Apprentice + Lee + La + Go + Shallow Parsing Stem ending table ... (verb / stem) + (ri + la + high + silver / suffix column) (verb / stem) + (ri + la + high / suffix column) (verb / stem) + (ri + la / suffix column ) (Chain / Stem) + (E + Li + La + Go + Silver / Suffix Column) (Chain / Stem) + (E + Lee + La + High / Suffix Column) (Chain / Stem) + (E + Lee + D / suffix column) (term / stem) + (two / suffix) (term / stem) + (temp / suffix) (term / stem) + (silver / suffix) (verb / stem) + (high / suffix). .. Stem-mother exploration

In the morpheme analysis step (S201) described above, the result of dividing the word into a basic unit of speech recognition called morpheme by using '+' as a delimiter may be expressed as in the example of Table 2.

Separate morphological columns are divided into stems ("under probation") and endings ("Lee + La + Go +") and these stem groups (ie, prefixes, verbs, messages, etc.) and mother groups (ie, investigations). , Suffixes, and endings) will be registered in the stem ending table, which defines the type of association and association.

At this time, the mother registers not only an independent mother but also a concatenated form (suffix sequence) between the mothers, and performs scoring by directly referring to the combined form or association relationship of the word and the verb.

Table 3 below is a table showing a step-by-step recognition operation of the speech recognition method according to an embodiment of the present invention.

Processing order Processing result N-best candidate sequence example Trigram navigation Morphological (word) unit N-best candidate sequence 1) Motherland / nc yi / j korea / nq yi / j li / e dreams / ma thought / nc ha / xsp zi / e not / pv i / e da / e 2) home country / nc i / j port / nq / J li / e dreams / ma thoughts / nc ha / xsp ji / e can't / pv was / e da / e 3) home country / nc yi / j korea / nq nc yi / j li / e dreams / ma Think / nc ha / xsp ji / e maternal / nc da / j 4) motherland / nc ga / j surrender / nc ha / xs ga / j li / e dream / ma think / nc ha / xsp ji / e can pv was / e die / e Stemming Trigram Navigation N-Best Candidate Revaluation 1) Motherland / nc yi / j korea / nq yi / j li / e dreams / ma thought / nc ha / xsp zi / e not / pv i / e da / e 2) home country / nc i / j port / nq This / j li / e dream / ma thought / nc ha / xsp ji / e i can't / pv was / e da / e 3) homeland / nc i / j korea / nq i / j li / e dream / ma thought / nc ha / xsp edge / e womb / nc die / j 4) homeland / nc Yi / j surrender / nc ha / xs Yi / j li / e dreams / ma thought / nc ha / xsp ji / e hardship / pv Was / e die / e Stem-mother exploration N-Best Candidate Revaluation 1) Motherland / nc + / j korea / nq + / j + li / e dream / ma thought / nc + ha / xsp + ji / e + not / pv + / e + c / e 2) Motherland / nc + i / j port / nq + This / j + li / e dream / ma thought / nc + ha / xsp + ji / e + not / pv + was / e + da / e 3) Motherland / nc + / j korea / nq + / j + ri / e dream / ma thought 4) Motherland / nc + surrenders / j surrender / nc + ha / xs? / j + li / e dreams / ma thought / nc + ha / xsp + ji / e + Nails / pv + / e + multi / e Final N-Best Candidate 1) I never thought my motherland would be Korea 2) I never thought my motherland was a port 3) I never thought my motherland would be Korea 4) I never thought my motherland would surrender

In the multi-speech speech recognition method based on statistical language model learning according to the embodiment of the present invention as described above, the recognition operation for the example of uttering that "the home country is not dreaming of Korea" is shown in the example of Table 3. As shown.

An example of a general speech recognition engine using a morpheme as a recognition unit is a form in which a recognized unit morpheme and its part-of-speech tag are attached, as in the trigram search shown in Table 3.

The example in Table 3 is an example in which the speech recognition engine generated up to four candidates for valid recognition columns, the first showing the highest score, and the candidate columns from 1) to 4) are presented in order. to be.

The next step, the morpheme trigram search step S205, is a morpheme trigram whether it is a valid morpheme column for this (that is, a morpheme (word) unit en-best candidate sequence generated in the trigram search step S204). It is determined by referring to.

In this case, the morphological arrangement of candidate strings 1) and 2) follows the array of trained morpheme trigrams that are exactly as shown in the morphological corpus, but in the case of candidate string 3), "G / e matrix / nc da / j". It can be seen that the "e nc j" of is far from the learned stemgram, and in the case of candidate 4), "nc xs je" of "yield / nc ha / xs is / e" We can see that there is a distance from the learned stemgram. In this case, in the morpheme trigram search step S205, the score is adjusted downward by applying a penalty to the candidate strings 3) and 4).

Finally, in the stem-mother searching step (S206), the coupling structure of the stem and the mother is evaluated. In particular, by referring to the connection rule and the stem ending reference table for the words and phrases, as shown in the example of Table 3 presented in the stem-mother searching step (S206), it is possible to create a coupling structure between the stem and the stem.

Also. Automatically detects the result of a combination or incomplete combination that violates the stem ending table, such as "yield / nc + / xs? / J +? / E" in candidate column 4). The score is reduced by designating the part with a symbol such as' and penalizing it.

As described above, the multi-stage speech recognition method according to an embodiment of the present invention performs a morpheme trigram search based on morpheme sequences as a first step in constructing a large vocabulary speech recognition system based on a statistical language model. The second step of searching for morpheme part-of-speech trigrams based on parts of speech is defined, and stemming is defined as a stem-mother combination structure, which is a concept of sub-elements of a word, and the reference table is knowledge-based. Follow the first step. In particular, by simply reconstructing and utilizing basic information generated in the morphological language model learning process, it is possible to efficiently reevaluate the results of speech recognition and to obtain performance improvement.

In addition, the embodiment of the present invention is not implemented only through the above-described apparatus and / or method, but through a program for realizing a function corresponding to the configuration of the embodiment of the present invention, a recording medium on which the program is recorded, and the like. Such implementations may be readily implemented by those skilled in the art from the description of the above-described embodiments.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements of those skilled in the art using the basic concepts of the present invention defined in the following claims are also provided. It belongs to the scope of rights.

1 is a schematic flowchart of a speech recognition method according to an exemplary embodiment of the present invention.

2 is a flowchart illustrating a speech recognition method in detail according to an embodiment of the present invention.

Claims (15)

In the method of recognizing speech, A corpus learning process to generate a knowledge base for speech recognition from raw text corpus, and Recognition process for generating a final recognition result of the speech recognition engine using the morpheme as a recognition unit using the knowledge base Speech recognition method comprising a. The method of claim 1, The corpus learning process, Generating a morpheme corpus and a morpheme column by performing a morpheme analysis on the raw text corpus, Generating a word trigram and a morpheme trigram by performing statistical language modeling on the morpheme corpus, and Constructing a linkage form of mother or suffixes and actual linked type endings by inputting the morpheme corpus into a stem mother reference table Speech recognition method comprising a. The method of claim 2, The morpheme corpus is A speech recognition method in which the words of the raw text corpus are separated into morphemes. The method of claim 3, The morphological heat, Speech recognition method for forming a separate sentence only from the part of speech tag excluding morpheme from the morpheme corpus. The method of claim 2, The word trigram, Speech recognition method generated by the language model consisting of morpheme unigram, bigram, trigram for the morpheme corpus. The method of claim 2, The morpheme trigram, And a part-of-speech set language model consisting of unigrams, bigrams, and trigrams of morpheme parts of speech for the morpheme sequence. The method of claim 1, The recognition process, Performing a trigram search on the spoken speech to generate a plurality of primary en-best recognition candidate sequences; Generating a second en-best list by performing a search based on a morpheme trigram on the first en-best recognition candidate sequence; and Generating a final en-best list by performing a re-evaluation on the secondary en-best list by applying a stem mother reference table Speech recognition method comprising a. The method of claim 7, wherein The trigram search, A speech recognition method for converting recorded speech into text using acoustic models, language models, and speech dictionaries. The method of claim 7, wherein Generating the secondary en-best list, And a second en-best list is generated by re-adjusting the ranks by scores by verifying and re-evaluating the first en-best recognition candidate sequence. The method of claim 7, wherein The search based on the morpheme trigram is Speech recognition method for re-evaluating recognition result candidates applying morphological-syntax constraints. In the method of recognizing spoken voice, Generating a plurality of first order n-best candidate sequences by performing a word trigram search on the spoken voice; Generating a secondary en-best candidate sequence by performing a morpheme trigram search on the primary en-best candidate sequence, and Generating a final en-best candidate sequence by performing a stem-mother search on the secondary en-best candidate sequence Speech recognition method comprising a. The method of claim 11, Generating the secondary en-best list, And re-evaluating the first en-best candidate sequence using a part-of-speech tag sequence of a recognition candidate sequence by applying a morpheme trigram. The method of claim 11, Generating the final en-best list, Constructing a stem mother reference table in the stem-mother search, and Re-evaluating the secondary en-best candidate sequence based on the constructed stem mother reference table Speech recognition method comprising a. The method of claim 13, The step of constructing the stem mother reference table, A speech recognition method that constructs a stem mother reference table by analyzing a lexical syntactic stem and connection structure of a mother and a database of combining rules of the mother. The method of claim 13, The step of constructing the stem mother reference table, Separating words into stems and endings, Grouping the mothers as mother fever without separating them anymore Defining the mother join string as a database, and Steps to Define Database Fields in the Organ Mother Table Speech recognition method comprising a.

Images