KR101398639B1 - Method and apparatus for speech registration - Google Patents

Abstract

A speech recognition apparatus and method are provided. A speech recognition apparatus according to an embodiment of the present invention includes a first phoneme recognition unit for generating a phonetic lattice from a feature vector of an input speech signal, a language model correction unit for correcting the speech lattice using a phoneme rule, And a second phoneme recognition unit for recognizing speech from the feature vector using the corrected speech grid.

Speech recognition, voice registration, language model, acoustic model, language model correction, real time speech recognition.

Description

[0001] The present invention relates to a method and apparatus for speech recognition,

The present invention relates to a speech recognition apparatus and method, and more particularly, to a method and apparatus for two-pass speech recognition based on language model correction.

In general, an intelligent robot is a robot that can judge and act on its own, and it is a technology convergence system that performs human functions while sharing human interactions with real space, assisting housework labor of housewives, Robot, and the like. On the other hand, it is effective to use voice, which is one of the most natural and easy interface for a person, for interaction between a human and an intelligent robot. In particular, in order to increase the convenience of the user, it is required to command and operate the cleaning robot or the entertainment robot used in the home by voice.

In order to perform such a function, basically, a function of registering and recognizing the user's voice in real time is needed. Real-time speech processing means that the conversation between a person and a robot does not have time to actually wait for the user to feel the time that the robot processes voice, such as a dialog between a person and a person. Therefore, the output from the robot should be output as soon as the user's voice ends.

Conventional speech recognition apparatuses can be used only for words registered in advance on a word basis, so that the speech recognition function is limited according to purposes and processing for OOV (Out Of Vocabulary) is required. Therefore, as one of the methods for receiving unlimited words, a method of performing phoneme recognition using a phoneme-unit language model and an acoustic model without a phonetic dictionary was used.

However, the conventional speech recognition apparatus and method have the following problems.

First, when phoneme recognition is performed using a monophone-based acoustic model and a phoneme-based n-gram language model, phoneme recognition can be performed in real time, but recognition performance is significantly degraded. On the other hand, when a triphone-based acoustic model and a phoneme-based n-gram language model are used (hereinafter referred to as a prior art A), recognition performance improves, There was a problem that was difficult to perform.

In addition, the conventional two-pass speech recognition method (hereinafter referred to as the conventional technique B) first performs fast pass decoding by applying a simple language model such as unigram and bigram, The search space of the language model is reduced and the probability is re-adjusted to perform the second pass decoding again. This can reduce the amount of computation and improve the accuracy of voice recognition to a certain extent, but the problem of high recognition rate of voice still exists.

It is an object of the present invention to provide a dual-path speech recognition apparatus and method based on language model correction, whereby unlimited words can be registered in real time, And to provide a speech recognition apparatus and method capable of improving the accuracy of recognition.

The technical problem of the present invention is not limited to those mentioned above, and another technical problem which is not mentioned can be clearly understood by those skilled in the art from the following description.

In order to achieve the above object, a speech recognition apparatus according to an embodiment of the present invention includes a first phoneme recognition unit for generating a phonetic lattice from a feature vector of an input speech signal, And a second phoneme recognition unit for recognizing speech from the feature vector using the corrected speech grid.

According to another aspect of the present invention, there is provided a speech recognition method including generating a speech grid from a feature vector of an input speech signal, correcting the speech grid using a phoneme rule, And recognizing speech from the feature vector using the corrected speech grating.

According to the speech recognition apparatus and method of the present invention, one or more of the following effects can be obtained.

First, unlimited words can be registered and recognized unlike existing technologies that only register words existing in the built-in pronunciation dictionary.

Second, it is possible to provide a speech recognition apparatus and method which can register and recognize new words in real time through performance enhancement due to dual-path speech recognition based on language model correction, and improve the accuracy of speech recognition.

Third, in order to enhance user's convenience, it is possible to eliminate unnecessary operation of inputting tools such as a keyboard and unnecessary proximity to a home appliance for immediate and smooth interaction with a user by incorporating voice recognition into a robot and a home appliance .

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description of the claims.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and the manner of achieving them, will be apparent from and elucidated with reference to the embodiments described hereinafter in conjunction with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. To fully disclose the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

Herein, the term " part " used in the present embodiment means a hardware component such as software or an FPGA or an ASIC, and 'part' performs certain roles. However, 'part' is not meant to be limited to software or hardware. &Quot; to " may be configured to reside on an addressable storage medium and may be configured to play one or more processors. Thus, by way of example, 'parts' may refer to components such as software components, object-oriented software components, class components and task components, and processes, functions, , Subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. The functions provided in the components and components may be further combined with a smaller number of components and components or further components and components.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described with reference to the drawings for explaining a speech recognition apparatus and method according to embodiments of the present invention.

1 is a block diagram illustrating a configuration of a speech recognition apparatus according to an embodiment of the present invention.

The speech recognition apparatus 100 according to an embodiment of the present invention includes a specific vector extraction unit 110, a first phoneme recognition unit 120, a language model correction unit 130, and a second phoneme recognition unit 140 ).

Meanwhile, the speech recognition apparatus 100 according to an embodiment of the present invention can be used not only for speech recognition, but also for registering new speech. Hereinafter, the voice recognition function will be described as an example.

The feature vector extractor 110 divides the input speech signal 10 into unit frames and extracts feature vectors 112 corresponding to the divided frame regions.

First, when a voice signal is input to the voice recognition apparatus 100, an actual voice interval can be detected from the voice signal 10 input through voice activity detection (VAD). The feature of the speech is extracted in order to acquire information suitable for speech recognition from the speech signal with respect to the detected speech section. At this time, the frequency characteristic of the speech signal is calculated for each unit frame, and the feature vector 112 included in the speech signal is extracted. To this end, the feature vector extracting unit 110 may be provided with an analog-to-digital converter (A / D converter) for converting an analog voice signal into a digital signal. Can be divided and processed. For example, under the standardization rate of 16 kHz, the number of samples for speech recognition is 160. Such a unit frame may include at least one Hidden Markov Model state (Hidden Markov Model state).

Preferably, the feature vector extractor 110 may extract the feature vector 112 using a Mel-Frequency Cepstrum Coefficients (MFCC) feature extraction scheme. The Mel-frequency cepstrum coefficient feature extraction method can use the feature vector 112 in the form of a combination of the Mel-Wurst coefficients, the log energy, and their first- and second-order derivatives.

The feature vector extraction unit 110 extracts features of a speech signal in a unit frame region by performing linear predictive coding (LPC), LPC derived cepstrum by linear predictive coding, Perceptive Linear Prediction (PLP), Audio Model Feature Extraction, and Filter Bank.

The first phoneme recognition unit 120 generates a phonetic lattice 128 from the feature vector 112 of the input speech signal 10.

FIG. 2 is a diagram illustrating a process of generating a speech grid from a feature vector of a speech signal of a first phoneme recognition unit in the speech recognition apparatus of FIG. 1;

2, the first phoneme recognizing unit 120 recognizes a phoneme by pattern matching the feature vector 112 using the first acoustic model 122 and the language model 124, Group 126 and generate a voice grid 128 from the first phoneme string group 126. [ Phoneme recognition at this stage is called Fast-Pass Decoding.

Preferably, the first acoustic model 122 may use a monophone-based acoustic model. In general, the acoustic model can use a hidden Markov model (HMM) learned by word, syllable, and phoneme unit, and it is effective to use phoneme-based HMM for speech recognition. The phoneme unit HMM can mainly use a monophone model considering only a single phoneme and a triphone model considering a fore and aft phoneme. Since the phonetic characteristic of the reference phoneme differs according to the front and back phonemes, the recognition performance can be improved by using the triphone model. However, when the triphone model is used, the number of the models becomes three times the number of the monophonic models, and a great time is required for performing the phoneme recognition.

Accordingly, the first phoneme recognition unit 120 can perform phoneme recognition using a monophone-based acoustic model. On the other hand, the monophone-based acoustic model used for speech recognition may include 21 vowels, 19 consonants, and 2 silences, for a total of 42 phonemes.

On the other hand, the language model 124 is used to reduce the search space by modeling human language generation patterns such as grammar and defining the recognition space, thereby reducing recognition time and calculation amount, and can use a statistical based method. The statistical based method is a method of statistically estimating a probability value of a possible word sequence from a database of speech uttered in a given situation. The language model generated through this technique has a flexible structure to accommodate sentences that do not match the grammar, but it has the disadvantage of increasing the search space. One of the statistical-based language models is the n-gram, which defines the probability of the next word from the previous n-1 words. That is, the probability for a word string can be approximated by the product of the conditional probabilities of the previous n-1 words.

Preferably, a bigram-based language model may be used as the language model 124 when registering the input speech signal 10. When the input speech signal 10 is to be recognized, You can use a unigram-based language model.

3 is a diagram showing a bigram-based language model of a phoneme unit.

The phoneme-based bigram-based language model is a language model in which the probability for the current phoneme is approximated by the product of the conditional probabilities for the previous phoneme. That is, if speech recognition is performed using a monophone-based acoustic model having n phonemes, a search space corresponding to n × n is required.

For example, if speech recognition is performed using an acoustic model having 42 phonemes, since the phonemes positioned before and after the 42 phonemes must be considered in the case of the triphone model-based acoustic model, about 70,000 (42 × 42 × 42) phonemes should be used. In addition, when using a bigram-based language model that considers probability values that can appear between two phonemes, a search space corresponding to the square power of 70,000 phonemes (70,000 × 70,000) is required. Therefore, although speech recognition performance can be improved, it is difficult to perform phoneme recognition in real time because the amount of calculation is increased and the time required for phoneme recognition is greatly increased.

Accordingly, in the speech recognition apparatus 100 according to an embodiment of the present invention, a bigram-based language model is generated using a monophone-based acoustic model in order to reduce the time for recognizing the first phoneme string group 126, , The search space can be reduced and the processing speed can be greatly reduced.

Meanwhile, the first phoneme string group 126 may be generated using a Viterbi algorithm. That is, the feature vector 112 is a pattern matching method for generating a first phoneme string group 126 by performing pattern matching using the first acoustic model 122 and the language model 124, Algorithm can be used. The Viterbi algorithm calculates the probability of all possible phoneme sequences for the input speech signal 10 by performing a state probability calculation process for all states of all phonemes for each frame of the input speech signal 10 , The first phoneme string having the highest probability (i.e., the N-best phoneme string) can be found. The number of phonemes constituting the first phoneme string group 126 having the highest probability may be a value preset in the speech recognition apparatus 100. Also, in order to increase the performance of the Viterbi algorithm, a pruning method for calculating only states having a certain probability may be used. Accordingly, the voice grid 128 can be generated using the N-best phoneme string 126 generated by the Viterbi algorithm and the pruning.

FIG. 4 is a diagram illustrating a state in which a first phoneme string is generated in a first phoneme recognition unit and a corresponding speech grid is generated. FIG.

For example, the first phoneme recognition unit 120 performs phoneme recognition on a monophone-based acoustic model 122 and a bigram-based language model 124 with respect to the input speech signal 10 called "school" The first phoneme string group 126, which has been pruned and survived through the Viterbi algorithm, may be "hagg jo", "hag G jo", "agg I oa", "h E gg jo". The first phoneme string group 126 refers to a phoneme string (N-best phoneme string) having the highest probability among all phoneme strings recognizable with respect to the input speech signal 10. Thereafter, the voice grid 128 can be generated using the first group of phoneme strings 126.

As described above, the first phoneme recognizing unit 120 generates the voice lattice 128 generated by performing the rapid phoneme recognition using the monophone-based acoustic model 122, and using the voice lattice 128, The second phoneme recognition unit 140 can reduce the search space significantly in the phoneme recognition.

The language model correcting unit 130 corrects the voice grid 128 using a phoneme rule. That is, the language model correcting unit 130 corrects the voice grid 128 generated by the first phoneme recognizing unit 120 so that the second phoneme recognizing unit 140 can perform more accurate and rapid phoneme recognition .

5 is a block diagram showing a language model correcting unit in the speech recognition apparatus of FIG.

The language model correcting unit 130 of the speech recognition apparatus 100 according to an embodiment of the present invention includes a phoneme string generating unit 131, a phoneme string selecting unit 132, a phoneme string correcting unit 133, And a voice grid generator 134.

The phoneme string generating unit 131 may generate a group of phoneme strings available from the voice grid 128 generated by the first phoneme recognizing unit 120. The phoneme string selecting unit 132 may sort the phoneme string groups generated by the phoneme string generating unit 131 according to the probability values, and then select a phoneme string group having a high probability value. The phoneme string correcting unit 133 can generate the corrected phoneme string group by converting the selected phoneme string in the phoneme string selector 132 using a phoneme rule.

The phoneme rule used to convert the phoneme string group in the phoneme string correcting unit 133 is based on a lot of experimental data for correcting the language model and instead of using a statistical technique dependent on a specific area, Knowledge-based universal rules can be applied. For example, the phonetic rules can use the standard pronunciation method of the Korean language dictionary.

Table 1 below shows examples of phoneme rules that can be applied to convert phoneme strings in a phoneme string modifier of a language model modifier.

number Rule name Phoneme column Calibrated phoneme string Rule 1 Remove redundant consonants Consonant 1-consonant 2-consonant 3 Consonant 1-consonant 2
Consonant 1-consonant 3
Consonant 2-consonant 3 Rule 2 Remove the end phoneme of a double consonant Consonant 1-consonant 2-silence Consonant 1-silence
Consonant 2-silence Rule 3 Removal of ending phonemes of consonant consonants Phoneme 1-consonant 1-silence Phoneme 1-silence (sil) Rule 4 Removing the first phoneme of a double consonant Silence - Consonants 1 - Consonant 2 Silence - consonant 1
Silence - Consonant 2 Rule 5 Correction of short high vowels {/,,,,, - {consonants} {/,,,,,,,,,,,,,,,,,,,,,, Rule 6 Tone correction {/ L /} - {/ l /} {/ C /} - {/ l /}
{/ L /} - {/ c /}
{/ L /} - {/ l /} Rule 7 Non-acoustic correction of obstructed sound {/ ㅁ, ㄴ, ㅇ /} - {NON} {/ F, c, a /} - {nasal}
{/ ㅁ, ㄴ, ㅇ /} - {NON} Rule 8 ㄷ - palatalization correction {V / v / v} - {vowel} - silence (sil) {/ C, § /} - {vowel}
{/ Ə, ː /} - {vowel} Rule 9 Interpretation of the term / Complementing the transcription from b, ㅁ / back {/ ㅣ, ㅁ /} - {/ ㄲ, ㄸ, ㅃ, ㅆ, ㅉ /} {/ ㅣ, ㅁ /} - {/ a, c, f,
{/ ㅣ, ㅁ /} - {/ ㄲ, ㄸ, ㅃ, ㅆ, ㅉ /} Rule 10 D - Calibration of the two-tone principle Sil - {/ ㅣ, ㅖ, ㅕ, ㅑ, ㅒ /} Sil - {/ ㅣ, ㅖ, ㅕ, ㅑ, ㅒ /}
(Sil) - {/ l /} - {/ ㅣ, ㅖ, ㅕ, ㅑ, ㅒ /} Rule 11 Correction of palatal hemisynthesis after palatal consonants /,, / / - / ㅕ / /,, / / - / ㅓ / Rule 12 "j" added / ㅣ, ㅔ. ㅐ, ㅟ, ㅚ / - / ㅓ / / ㅣ, ㅔ. ㅐ, ㅟ, ㅚ / - / ㅕ /
/ ㅣ, ㅔ. ㅐ, ㅟ, ㅚ / - / ㅓ /

For example, when the input speech signal 10 is "same", the phoneme string generated from the first phoneme recognition unit 120 may be "a". As shown in Table 1, the "ㅌ", "ㅅ", and "ㄷ" parts of the sentence are " Quot ;, "c ", or" g "-" c ". Thus, the corrected group of phonemes can be "ㄱ", "" ", and" ".

Finally, the speech grid generation unit 134 can generate the corrected speech grid 135 from the group of phoneme strings corrected by the phoneme string correction unit 133. [ Further, by applying a higher-order language model such as 2-gram or 3-gram to the corrected speech grid 135, the probability between the phoneme strings in the corrected speech grid 135 can be adjusted.

As described above, the speech grid 135 corrected by the language model correcting unit 130 can be used as a language model for phoneme recognition in the second phoneme recognizing unit 140. Therefore, after correcting an error that may occur in the fast-pass decoding process in the first phoneme recognizer 120 using a language model correction technique, the second phoneme recognition unit 140 corrects the result using a second-pass Decoding process, it is possible to perform more reliable phoneme recognition.

The second phoneme recognition unit 140 recognizes speech from the feature vector 112 using the corrected speech grid 135.

6 is a diagram illustrating a phoneme recognition process for generating an N-best phoneme string from a feature vector of a speech signal input by the second phoneme recognition unit in the speech recognition apparatus of FIG.

The second phoneme recognition unit 140 performs pattern matching on the feature vectors 112 using the second acoustic model 142 and the corrected speech grids 135 as described in the first phoneme recognition unit 120 Phonemes can be recognized. Phoneme recognition at this stage is referred to as second-pass decoding.

Preferably, the second acoustic model 142 may use a triphone-based acoustic model. As described above, by using the triphone-based acoustic model, phoneme recognition can be performed more accurately.

The recognition process is terminated by recognizing the N-best phoneme string group 144 generated by recognizing the phoneme in the second phoneme recognition unit 140 as the pronunciation of the input speech signal 10. [ The voice registration process can be performed through the phoneme recognition. At this time, the N-best phoneme sequence means phoneme strings having the highest probability value.

FIG. 7 is a graph comparing the time taken to recognize each word when recognizing the input speech signal using the speech recognition apparatus according to the prior art A, the background art B, and the speech recognition apparatus according to an embodiment of the present invention.

In the case of the prior art A, a triphone-based acoustic model having 42 phonemes is used. Since it needs to search about 70,000 phonemes, the time required for speech recognition is the most needed. In the case of the conventional technique B, phoneme recognition is performed with a monophone-based acoustic model (Fast-Pass Decoding), and phoneme recognition is performed with a language model based on the result and a triphone-based acoustic model Second-Pass Decoding) The search space of the language model is reduced and the amount of computation is reduced, which requires the least time for speech recognition. In the case of the present invention, since the voice grid 128 is generated using the monophone-based acoustic model and the phoneme recognition is performed using the language model based on the voice grid 128 corrected for the voice grid 128, It takes time.

FIG. 8 is a graph comparing the recognition rates of the respective words when recognizing the input speech signal using the speech recognition apparatus according to the prior art A, the background art B, and the speech recognition apparatus according to an embodiment of the present invention.

In FIG. 8, Word Error Rate (WER) is used to indicate the accuracy of speech recognition.

In the case of the prior art A, the lowest false recognition rate is shown by using the triphone-based acoustic model. In the case of the prior art B, the language model generated by performing the phoneme recognition with the monophone-based acoustic model and the triphone- The recognition rate is higher by performing speech recognition using the acoustic model.

However, according to the present invention, the erroneous recognition rate of the word is increased by about 14% compared to the conventional art A, but is reduced by about 18% relative to the erroneous recognition rate of the conventional technique B.

Referring to FIG. 7 and FIG. 8, it can be seen that, in the case of the prior art A, the lowest false recognition rate is obtained but the time required for speech recognition is not greatly improved. In the case of the conventional technique B, the time required for speech recognition is short, but the accuracy of speech recognition is high due to high recognition rate. Therefore, according to the speech recognition apparatus and method according to the embodiment of the present invention, the time required for speech recognition can be significantly shortened compared with the conventional art A, and the accuracy of speech recognition can be significantly improved as compared with the conventional art B.

Therefore, according to the speech recognition apparatus and method according to an embodiment of the present invention, the time required for speech recognition can be reduced by reducing the search space using the dual path speech recognition method, and the speech can be recognized in real time. In addition, the accuracy of speech recognition can be improved by correcting the language model.

A speech recognition method according to an embodiment of the present invention will now be described.

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

First, the input speech signal 10 is divided into unit frames, and feature vectors 112 corresponding to the divided frame regions can be extracted (S201). The first phoneme recognition unit 120 recognizes phonemes by pattern matching the feature vectors 112 using a first acoustic model 122, that is, a monophone-based acoustic model and a language model, (Step S202). Then, the voice grid 128 can be generated from the first phoneme string group 126 (S203). Fast-Pass Decoding in the first phoneme recognition unit 120 can reduce the time for performing phoneme recognition by using a monophone-based acoustic model.

Next, in the language model correcting unit 130, the speech grid 128 generated by the first phoneme recognizing unit 120 can be corrected. That is, a phoneme string group is generated from the speech grating 128 generated from the first phoneme recognition unit 120 (S204), and the generated phoneme string group is sorted according to the probability value, (S205).

Next, the selected phoneme string group is converted to generate a corrected phoneme string group (S206). At this time, the same rules as in Fig. 6 can be applied as described above. Then, the corrected speech grid 135 can be generated from the corrected group of phonemes (S207). The language model corrector 130 generates the corrected speech grid 135 and uses the generated speech grid 135 for the phoneme recognition in the second phoneme recognition unit 140, thereby reducing the time required for speech recognition.

Finally, the second phoneme recognition unit 140 performs the pattern matching using the second acoustic model 142, that is, the language model based on the triphone-based acoustic model and the corrected voice grid 135, So that the phoneme can be recognized (S208). The time required for speech recognition can be reduced by performing the phoneme recognition using the speech grid 135 corrected by the language model correction unit 130 and the recognition rate can be increased by using the triphone-based acoustic model.

The speech recognition apparatus 100 and method according to an embodiment of the present invention can be applied not only to an intelligent robot system but also to all fields using a speech interface. This can be applied to portable devices or household appliances to perform commands and operations using voice. By identifying the speaker through the voice of a registered speaker, the use of the device by a person other than the registered user is restricted, . In addition, OOV (out-of-vocabulary), which is a problem in large-capacity speech recognition for natural conversation between people and machines, can be efficiently processed through unlimited speech recognition, thereby enabling a more natural interaction between a device and a person .

It will be understood by those skilled in the art that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. The scope of the present invention is defined by the appended claims rather than the foregoing detailed description, and all changes or modifications derived from the meaning and scope of the claims and the equivalents thereof are included in the scope of the present invention Should be interpreted.

1 is a block diagram illustrating a configuration of a speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating a process of generating a speech grid from a feature vector of a speech signal of a first phoneme recognition unit in the speech recognition apparatus of FIG. 1;

3 is a diagram showing a bigram-based language model of a phoneme unit.

FIG. 4 is a diagram illustrating a state in which a first phoneme string is generated in a first phoneme recognition unit and a corresponding speech grid is generated. FIG.

5 is a block diagram showing a language model correcting unit in the speech recognition apparatus of FIG.

6 is a diagram illustrating a phoneme recognition process for generating an N-best phoneme string from a feature vector of a speech signal input by the second phoneme recognition unit in the speech recognition apparatus of FIG.

FIG. 7 is a graph comparing time taken for recognition of an input speech signal when recognizing an input speech signal using the speech recognition apparatus according to the prior art A, the background art B, and the embodiment of the present invention.

FIG. 8 is a graph comparing the word recognition rates of the input speech signals when recognizing input speech signals using the speech recognition apparatus according to the prior art A, the related art B, and an embodiment of the present invention.

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

Description of the Related Art

100: Speech recognition device

110: Feature vector extracting unit

120: First phoneme recognition unit

130: Language Model Correction

140: Second phoneme recognition section

Claims (24)

A first phoneme recognition unit for generating a phonetic lattice from the feature vector of the input speech signal; A language model corrector for correcting the speech grid using a phoneme rule; And And a second phoneme recognition unit for recognizing speech from the feature vector using the corrected speech grid. The method according to claim 1, And a feature vector extracting unit for dividing the input speech signal into unit frames and extracting the feature vectors corresponding to the divided frame regions. 3. The method of claim 2, Wherein the feature vector extracting unit comprises: And a Mel-Frequency Cepstrum Coefficients feature extraction unit for extracting the feature vector using a Mel-Frequency Cepstrum Coefficients feature extraction method. The method according to claim 1, Wherein the first phoneme recognition unit comprises: Wherein the feature vectors are pattern-matched using a first acoustic model and a language model to generate a first phoneme string group by recognizing phonemes and generate the speech lattice from the first phoneme string group. 5. The method of claim 4, Wherein the first acoustic model is a monophone-based acoustic model. 5. The method of claim 4, Wherein the language model is a bigram-based language model in units of phonemes when registering the input speech signal. 5. The method of claim 4, Wherein the language model is an unigram-based language model for recognizing the input speech signal by word unit. 5. The method of claim 4, Wherein the first phoneme string group is generated using a Viterbi algorithm. The method according to claim 1, The language model correcting unit, A phoneme sequence generating unit for generating a phoneme string group from the speech grid; A phoneme string selector for selecting a phoneme string group having a high probability value after sorting the generated phoneme string group according to a probability value; A phoneme column correcting unit for converting the selected phoneme string group using the phoneme rule to generate a corrected phoneme string group; And And a speech grid generator for generating the corrected speech grid from the corrected group of phoneme strings. 10. The method of claim 1 or 9, Wherein the phoneme rule is based on knowledge of phonetics or phonology. The method according to claim 1, Wherein the second phoneme recognition unit comprises: Wherein the feature vector is pattern-matched using the second acoustic model and the corrected speech grating to recognize a phoneme. 12. The method of claim 11, And the second acoustic model is a triphone based acoustic model. Generating a speech lattice from feature vectors of the input speech signal; Correcting the speech grid using a phoneme rule; And And recognizing speech from the feature vector using the corrected speech grid. 14. The method of claim 13, Further comprising dividing the input speech signal into unit frames and extracting the feature vectors corresponding to the divided frame regions. 15. The method of claim 14, Wherein the extracting of the feature vector comprises: And extracting the feature vector using a mel-frequency cepstral coefficient feature extraction method. 14. The method of claim 13, Wherein the step of generating the speech grating comprises: Generating a first phoneme string group by recognizing a sound by pattern matching the feature vectors using a first acoustic model and a language model; And And generating the speech grid from the first phoneme string group. 17. The method of claim 16, Wherein the first acoustic model is a monophone acoustic model. 17. The method of claim 16, Wherein the language model is a bigram-based language model of the phoneme when the input speech signal is to be registered. 17. The method of claim 16, Wherein the language model is an unigram-based language model for recognizing the input speech signal by word unit. 17. The method of claim 16, Wherein the first phoneme string group is generated using a Viterbi algorithm. 14. The method of claim 13, The step of correcting the speech grating comprises: Generating a phoneme string from the speech grid; Selecting a phoneme string group having a high probability value after sorting the generated phoneme string group according to a probability value; Converting the selected phoneme string group using the phoneme rule to generate a corrected phoneme string group; And And generating the corrected speech grid from the corrected group of phoneme strings. 22. The method according to claim 13 or 21, Wherein the phoneme rule is based on knowledge of phonetics or phonology. 14. The method of claim 13, Wherein recognizing speech from the feature vector comprises: Wherein the feature vector is pattern-matched using a second acoustic model and the corrected speech grating to recognize a phoneme. 24. The method of claim 23, And the second acoustic model is a triphone-based acoustic model.

Images