KR100298177B1 - Method for construction anti-phone model and method for utterance verification based on anti-phone medel - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

The present invention relates to a method for constructing a semi-phoneme model in a speech recognition system, a method for verifying speech using the same, and a computer-readable recording medium having recorded thereon a program for realizing the methods.

2. The technical problem to be solved by the invention

In the present invention, the context-dependent (CD) phoneme in the speech recognition system is divided into training and test use, and a similar phoneme set composed of context-independent (CI) phonemes generated therefrom is recognized through training for each CD phoneme. By building a semi-phoneme model for speech verification used to perform the rejection function, a method to construct a semi-phoneme model to reduce training data and training time and to improve the performance of the system, and a method of verifying speech using the same, and to realize the above methods To provide a computer-readable recording medium that records a program for the purpose.

3. Summary of Solution to Invention

In the method for constructing a semi-phoneme model for speech verification applied to a speech recognition system, the present invention extracts features required for speech recognition from speech data input from the outside, and divides them into context-dependent (CD) phonemes. Collecting and classifying the classified phonemes into training context-dependent phonemes and test context-dependent phonemes; A second step of constructing a context-independent (CI) phoneme model through training by collecting the training context-dependent phonemes in a context-independent (CI) phoneme unit; And a similar phoneme set by performing a phoneme recognition test based on the test context-dependent (CD) phoneme and the context-independent (CI) phoneme model, and training through a similar phoneme set for each context-dependent phoneme (CD). Including the third step of building a phoneme model.

4. Important uses of the invention

The present invention is used in a voice recognition system.

Description

METHODO FOR CONSTRUCTION ANTI-PHONE MODEL AND METHOD FOR UTTERANCE VERIFICATION BASED ON ANTI-PHONE MEDEL}

According to the present invention, a semi-phoneme model for speech verification used to perform a recognition rejection function in a speech recognition system is constructed by mixing a context-independent (CI) phoneme and a context-dependent phone (CD). In addition, the present invention relates to a method for constructing a semi-phoneme model for reducing training data and training time and improving system performance, a method for verifying speech using the same, and a computer-readable recording medium recording a program for implementing the methods.

The voice recognition system performs the voice recognition function on the assumption that only a specific word to be recognized is pre-set, so if the user speaks a word other than the word to be recognized by mistake or intentionally, the recognition result is displayed as one of the words to be recognized. Therefore, it has a problem that is recognized as a wrong word.

Therefore, when a word other than the word to be recognized is input in the speech recognition system, a function of rejecting the word without recognition is required. That is, in a speech recognition system, when a word other than a word to be recognized is input, a function of notifying a word other than a word to be recognized is required without a mistake. This is called a voice recognition rejection function.

The present invention relates to the construction of a semi-phoneme model related to a speech verification method in a method of performing a speech recognition rejection function.

The speech verification method is a method of determining which confidence (Refidence Score or Confidence Measure) values are used to determine whether to accept (Accept) or reject (Reject) the recognition result. Here, the reliability is a measure of how reliable the result is for the speech recognition result. If the reliability value is high, the recognition result should be accepted as reliable, and if it is low, it is difficult to trust the result. The recognition result should be rejected.

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

The reliability is different from the Viterbi search result. That is, the Viterbi search result number represents a simple similarity to a word or phoneme, while the reliability is a relative value of the probability that the word is spoken from other phonemes or words for the recognized phoneme or word. Means.

To determine the reliability, a phone model and an anti-phone model are required.

A phoneme model is an HMM created by training extracted phonemes by extracting phonemes actually spoken from a voice. The phoneme model is a model used in a speech recognition system based on a general HMM.

On the other hand, the half-phoneme model refers to an HMM trained using phonemes that are very similar to actual phonemes (these are called cohort sets).

As such, in the speech recognition system, a phoneme model and a semiphoneme model exist for each phoneme used.

For example, " About the phoneme called " "There is a phoneme model," The semitone phone model for "will exist.

For example, " "The model of the phoneme in the speech database" "Only extracts phonemes and trains the training method of HMM, and is made. And," To build a phoneme model for " You should get a similar phoneme set for ". This can be found by looking at the phoneme recognition results, by performing a phoneme recognition process. Any phonemes other than " See if you've been misunderstood and collect them Can determine a set of similar phonemes for " , , Phonemes are mainly " "If they were misunderstood, they could be called a similar phoneme set. "A semiphoneme model is created for the phonemes.

If the phoneme model and the phoneme model are generated for all the phonemes in this way, the reliability of the input voice is calculated as follows.

First, the phoneme model is searched to find the most similar phoneme.

Then, the similarity is calculated for the semitone phone model.

The final reliability can be obtained by calculating the difference between the similarity between the phoneme model and the similarity between the phoneme model and adjusting a range of the reliability value by applying a predetermined specific function.

In the related art, the semi-phoneme model uses only context-independent (CI) phones. Here, the context-independent phoneme refers to a simple phoneme that does not consider the phoneme before and after each phoneme, whereas the context-dependent phoneme considers a phenomenon in which the phoneme is influenced by the phoneme before and after the same phoneme. Say more subtle phonemes.

Generally, about 40-60 CI phonemes are used in Korean speech recognition system, and about 300 CD phonemes are used in consideration of memory and recognition time. In the general speech recognition system, CD phonemes are used to train and test. However, in case of the semiphoneme model used for speech verification, CI phoneme model is mainly used. This is because when using a semi-phoneme model of the CD phoneme unit, the number of phonemes is too large, which requires a lot of training data and training time.

Conventionally, there has been a method of performing speech recognition rejection using speech verification.

In the conventional method of performing speech recognition rejection function using the speech verification, a reliability value is also obtained for speech verification. To obtain such reliability, an HMM for phonemes and an HMM for semitones are required. That is, the difference between the HMM for the phoneme and the HMM for the semiphoneme is compared and the value is used as the reliability.

However, conventionally, only a method of using CI phonemes existed in constructing a half-phone HMM.

In addition, in the conventional phoneme recognition process, since the phoneme HMM uses CD phoneme and the half phonemeal HMM uses CI phoneme, there is a disadvantage in that performance decreases. In addition, conventionally, in order to build a semi-phoneme model using CD phonemes, since the number of CD phonemes is large, a large amount of data required for training and training with insufficient data, there is a problem in that the performance of the system is poor.

The present invention devised to solve the problems as described above, using the context-dependent (CD) phoneme in the speech recognition system, divided into training and test use, and similar phonemes composed of the context-independent (CI) phonemes generated therefrom By building a semi-phoneme model for speech verification used to perform the recognition rejection function through training for each CD phoneme as a set, a semi-phoneme model construction method for reducing training data and training time and improving system performance and A utterance verification method using the same and a computer readable recording medium having recorded thereon a program for realizing the above methods are provided.

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

2 is a flowchart illustrating an embodiment of a method for constructing a half-phoneme model according to the present invention.

Figure 3 is a flow diagram of an embodiment of a speech verification method using a semi-phoneme model construction process according to the present invention.

* Explanation of symbols for the main parts of the drawings

11; Endpoint Detector 12: Feature Extractor

13: Viterbi Explorer 14: Pronunciation Dictionary

15: CD phoneme model database 16: Ignition Verifier

17: CD Semitone Phone Model Database

In order to achieve the above object, the present invention provides a method for constructing a semi-phoneme model for speech verification applied to a speech recognition system, wherein a feature required for speech recognition is extracted from speech data input from an external device. A first step of dividing the phoneme into groups and classifying the classified phonemes into training context-dependent phonemes and test context-dependent phonemes; A second step of constructing a context-independent (CI) phoneme model through training by collecting the training context-dependent phonemes in a context-independent (CI) phoneme unit; And a similar phoneme set by performing a phoneme recognition test based on the test context-dependent (CD) phoneme and the context-independent (CI) phoneme model, and training through a similar phoneme set for each context-dependent phoneme (CD). And a third step of constructing a semitone phone model.

And, in the speech verification method applied to the voice recognition system,

A first step of constructing a semi-phoneme model by mixing context-dependent (CD) phones and context-independent (CI) phones; A second step of detecting an end point of the voice during voice input, extracting only a portion in which the voice exists, and extracting a feature required for voice recognition from the extracted voice portion; A third step of performing speech recognition on the extracted speech data by referring to a pronunciation dictionary and a context-dependent CD model; And a fourth step of performing speech verification based on the reliability obtained by referring to the context-dependent (CD) half-phoneme model and the context-dependent (CD) phoneme model for speech recognition speech data.

The present invention for achieving the above object, in order to build a semi-phoneme model, a semi-phone model construction device having a processor, extracts the features required for speech recognition from the speech data input from the outside context-dependent (CD) phoneme A first function of dividing the phoneme into groups and classifying and classifying the classified phonemes into training context-dependent phonemes and test context-dependent phonemes; A second function of collecting the context-dependent phonemes for training in context-independent (CI) phoneme units and constructing a context-independent (CI) phoneme model through training; And a similar phoneme set by performing a phoneme recognition test based on the test context-dependent (CD) phoneme and the context-independent (CI) phoneme model, and training through a similar phoneme set for each context-dependent phoneme (CD). A computer readable recording medium having recorded thereon a program for realizing a third function of constructing a half phoneme model is provided.

In addition, the present invention provides a speech verification apparatus having a processor for speech verification, comprising: a first function of constructing a semi-phoneme model by mixing a context-dependent (CD) phoneme and a context-independent (CI) phoneme; A second function of detecting an end point of the voice at the time of voice input, extracting only a portion in which the voice exists, and extracting a feature required for voice recognition from the extracted voice portion; A third function of performing speech recognition on the extracted speech data by searching for the Viterbi by referring to the pronunciation dictionary and the context-dependent phoneme model; And a program for realizing a fourth function of performing speech verification on the basis of the reliability obtained by referring to the context-dependent (CD) half-phoneme model and the context-dependent (CD) phoneme model for speech recognition speech data. Provide a readable recording medium.

When a word other than the word to be recognized is entered in the speech recognition system, a function of rejecting it without recognizing it is required. In order to implement this, a speech verification process is performed to determine the reliability of a certain voice and how to recognize or reject the voice. Therefore, in order to obtain the reliability, a phoneme model and a semiphoneme model must be constructed. In the present invention, a method of constructing a phoneme model efficiently is proposed. In the conventional method, the semi-phoneme model is constructed using only the CI phoneme. In the present invention, a combination of the CI phoneme and the CD phoneme is used, and a method of constructing the semi-phoneme model of the CD unit is proposed. By using this method, it is possible to construct an efficient CD half-phone model with little training data by making up for the shortcomings of using a semi-noise model based on CI and building a semi-phone model based on CD using only CD. This is improved.

The present invention builds a semi-phoneme model by mixing CI and CD, through which speech recognition rejection function is performed using speech verification.

Here, in constructing the semitone phone model, the present invention divides the training phone and the test phone using a CD, and then creates a semiphoneme model after making a similar phoneme set composed of CI. The features are as follows.

end. Extract the features required for speech recognition from speech data.

I. Features The extracted speech data are phoneme-divided into CD units, and collected by the same phoneme.

All. Classified phonemes are classified for training and test purposes.

la. Collect CD collections for each training CD phone by CI.

hemp. The collected training data is used to build the HMM model for the CI phoneme through the HMM training process.

bar. The test data collected in units of CDs is used as an input, and the phoneme recognition test is executed using the HMM model of the CI phonemes constructed in the above section as a reference pattern.

four. For each CD phoneme, we obtain a similar phoneme set that tells us which CI phonemes are misrecognized. (In the above test, the phoneme is composed of CIs for CD phones because the input is CD phone and the reference pattern was CI phone. Set is obtained).

character. For each CD phoneme, HMM training is performed with similar phoneme sets, and HMM model for semitones is constructed.

On the other hand, in speech verification for rejection of speech recognition, the present invention has the following features.

end. When the voice is input, the end point of the voice is detected and only the portion where the voice is present is selected.

I. Extract the features required for speech recognition from speech part.

All. Speech recognition is performed through the Viterbi search process by referring to the pronunciation dictionary and the CD phoneme model.

la. The reliability is obtained by referring to the CD semiphone model and the CD phoneme model.

hemp. The reliability obtained in the above paragraph is compared with a threshold value, and if the confidence value is larger than the threshold, the recognized result is accepted. If the confidence value is smaller than the threshold, the recognized result is rejected.

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

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

Since the configuration and operation of the speech recognition system to which the present invention is applied are only known techniques in the art, detailed description thereof will be omitted herein. However, the process of performing the voice recognition rejection function on the input voice will be described in more detail.

First, when the voice is input, the endpoint detector 11 finds the voice section excluding the silent section before and after the voice. Thereafter, the feature extractor 12 extracts the feature of the speech from the speech signal of the speech section found earlier.

Next, in the Viterbi searcher 13, the words most similar to the likelihood are selected by using the voice feature values for the words registered in the pronunciation dictionary 14 composed of the CD phoneme model database 15.

Subsequently, the speech verifier 16 divides the feature section into phonemes using the word selected by the Viterbi searcher 13, and then obtains a Likelihood Ratio Confidence Score using the semitone phone model. . The equation for obtaining the phoneme-like reliability δ (CM _p ) is expressed by Equation 1 below.

In Equation 1, α and β are constants (α = 1, β = 0). In addition, LLR In P (O _e / λ ^a ) is the similarity value (Likelihood) by the semi-phoneme model, P (O _e / λ ^c ) is the similarity value by the phoneme model.

Looking at the utterance verification of the word unit, the equation for obtaining the word unit similarity reliability (δ (CM _w )) is expressed as the following equation (2).

In Equation 2, N is the number of phonemes in a word. At this time, if logδ (CM _w ) rises a predetermined threshold value (Accept), if less than that, the word is rejected (Reject).

Finally, if the word is rejected, the speech verification process 16 performs the speech verification process as described above for the next candidate word.

On the other hand, in the case of recognizing a sentence, the above utterance verification process is applied in the same manner, only the grammar is added, and the sentence unit is verified.

2 is a flowchart illustrating a method for constructing a semiphoneme model for speech verification according to the present invention.

According to the present invention, unlike the method of constructing a semitone model using only CI phones, the CI phoneme and CD phoneme are used in combination. By constructing the semitone model, we solved the disadvantages of using the semiphone model of CI phoneme and the semiphone model of CD phoneme using only CD phoneme. Can be built.

In order to efficiently construct a semi-phoneme model of CD phoneme units by combining CI phonemes and CD phonemes, the present invention first collects phonemes by CD phoneme units, and then divides them into training and test cases. Here, since CD phonemes are subdivided into CI phonemes according to the surrounding environment, they can be collected in CI phoneme units.

Therefore, after gathering training data back to CI phoneme unit, HMM training process generates HMM of CI phoneme unit. Subsequently, when the CI HMM is used as a reference pattern of the phoneme recognizer and the phoneme recognition test is performed by using the CD phoneme as an input, it is possible to know which CI phonemes are misrecognized for each CD phoneme. The result is a similar phoneme set for a CD phone composed of CI phones.

This pseudo phoneme set is generated in CD phoneme units, but since its components are CI phones, it does not require as much training data as CD phones.

Finally, when HMM training is performed based on the similar phoneme set, a semiphoneme model is generated for CD phonemes.

Therefore, when constructing the semitone phone model using this method, the present invention shows better performance with less training data because the CD phone is trained by the CI phoneme, and also the semiphoneme model of CI phoneme unit is constructed. You'll get better performance than it does.

To construct a speech recognition system, it is always divided into training and testing for evaluating its performance. Training is the data needed to determine if you are well trained. Therefore, training and recognition can not write the same data. Generally, the ratio between training and test is 8: 2. This is a know-how determined by the developer in consideration of words to be recognized and development time. That is, there is no special standard.

Looking at the method for building a semi-phoneme model for the speech verification according to the present invention as described above in detail.

As shown in FIG. 2, the method for constructing a semi-phoneme model for speech verification according to the present invention extracts a feature required for speech recognition from first input speech data (201).

Subsequently, after the feature-extracted voice data are divided into phonemes by CD phoneme and collected and sorted by the same phoneme (202), the classified phonemes are divided into training and test (203).

Next, after collecting a CD set of CI phonemes by collecting the CI phonemes for the training CD phones (204), the HMM for the CI phones is constructed through the HMM training process using the collected training data (206).

Then, using the test data collected in units of CD phonemes as input and running a phoneme recognition test using the HMM for the CI phoneme as a reference pattern (207), it indicates which CI phonemes are misrecognized for each CD phoneme. A similar phoneme set is obtained (208). That is, in the phoneme recognition test, since the input is a CD phone and the reference pattern is a CI phone, a similar phoneme set consisting of CI phonemes for the CD phone can be obtained.

Finally, HMM training is performed for semi-phones by constructing similar phonemes for each CD phone (209).

3 is a flowchart illustrating a method of verifying speech using a half phoneme model construction process according to the present invention.

As shown in FIG. 3, the method of verifying speech according to the present invention using a process of constructing a semi-phoneme model of a CD phoneme unit by combining CI phonemes and CD phonemes (see FIG. 2), first, when a voice is inputted, an endpoint detector ( In step 11), an end point of the voice is detected, and only a portion in which the voice exists is extracted (301).

Then, after the feature extractor 12 extracts a feature required for speech recognition from the extracted speech part (302), the pronunciation dictionary 14 and the CD phoneme model 15 in the Viterbi searcher 13 for the extracted speech data. In operation 303, voice recognition is performed through a Viterbi search process.

Next, the speech authenticator 16 obtains the reliability by referring to the CD semitone phone model 17 and the CD phoneme model 15 in the speech verifier 16, and performs speech verification therefrom (304). At this time, the obtained reliability value is compared with a predetermined threshold (305). If the reliability value is greater than the threshold, the recognized result is accepted (306). If the reliability value is not greater than the threshold, the recognized result is rejected (307).

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

The present invention described above is capable of various substitutions, modifications, and changes without departing from the spirit of the present invention for those skilled in the art to which the present invention pertains, and the above-described embodiments and accompanying It is not limited to the drawing.

According to the present invention as described above, the semitone model can be modeled more accurately by using a combination of the CI phoneme and the CD phoneme, so that the semitone phone can be modeled more accurately, and the CD is trained by the CI. It is also possible to construct an efficient CD half-phone model, which can improve the speech verification function of the system compared to the case of using the existing semi-phone model of the CI phoneme or using only the CD phoneme. It can be effective.

Claims (7)

In the semi-phoneme model construction method for speech verification applied to speech recognition system, Extracts the features required for speech recognition from the voice data input from the outside, divides them into contextual phoneme (CD) phonemes, and classifies and classifies the classified phonemes into training contextual phonemes and test contextual phonemes (CD). First step divided into; A second step of constructing a context-independent (CI) phoneme model through training by collecting the training context-dependent phonemes in a context-independent (CI) phoneme unit; And A phoneme recognition test is performed on the basis of the test context-dependent (CD) phoneme and the context-independent (CI) phoneme model to obtain a similar phoneme set. The third step in building a phoneme model Half phoneme model construction method in speech recognition system comprising a. The method of claim 1, The second step, A fourth step of collecting a phoneme set of context-independent (CI) phonemes by gathering the training context-independent (CI) phones by context-independent (CI) phones; And The fifth step of constructing a hidden Markov model for context-independent (CI) phonemes through the Hidden Markov Model (HMM) training with the collected training data. Half phoneme model construction method in speech recognition system comprising a. The method according to claim 1 or 2, The third step, By using the test data collected in units of the context dependent (CD) phonemes as input and executing the phoneme recognition test using the hidden Markov model (HMM) for the context independent (CI) phonemes as a reference pattern, each context dependency ( CD) a sixth step of obtaining a similar phoneme set indicating which context-independent (CI) phones are misrecognized for the phoneme; And Seventh step of constructing hidden Markov models for semitones through training hidden Markov models with similar phoneme sets for each context-dependent (CD) phoneme Half phoneme model construction method in speech recognition system comprising a. In the speech verification method applied to the speech recognition system, A first step of constructing a semi-phoneme model by mixing context-dependent (CD) phones and context-independent (CI) phones; A second step of detecting an end point of the voice during voice input, extracting only a portion in which the voice exists, and extracting a feature required for voice recognition from the extracted voice portion; A third step of performing speech recognition on the extracted speech data by referring to a pronunciation dictionary and a context-dependent CD model; And A fourth step of performing speech verification on speech recognized speech data based on the reliability obtained by referring to the CD-based phoneme model and the CD-based phoneme model Speech verification method in a voice recognition system comprising a. The method of claim 4, wherein The first step is, Extracts the features required for speech recognition from the voice data input from the outside, divides them into contextual phoneme (CD) phonemes, and classifies and classifies the classified phonemes into training contextual phonemes and test contextual phonemes (CD). A fifth step divided into; A sixth step of constructing a context-independent (CI) phoneme model through training by collecting the training context-dependent phoneme in a context-independent (CI) phoneme unit; And A similar phoneme set is obtained by performing a phoneme recognition test based on the test context-dependent (CD) phoneme and the context-independent phoneme (CI) phoneme model, and training through the similar phoneme set for each context-dependent phoneme (CD) phoneme. 7th Step to Build a Semitone Phone Model Speech verification method in a voice recognition system comprising a. In order to build a semi-phoneme model, in a semi-phone model building apparatus having a processor, Extracts the features required for speech recognition from the voice data input from the outside, divides them into contextual phoneme (CD) phonemes, and classifies and classifies the classified phonemes into training contextual phonemes and test contextual phonemes (CD). A first function divided into; A second function of collecting the context-dependent phonemes for training in context-independent (CI) phoneme units and constructing a context-independent (CI) phoneme model through training; And A phoneme recognition test is performed on the basis of the test context-dependent (CD) phoneme and the context-independent (CI) phoneme model to obtain a similar phoneme set. Third function to build phoneme model A computer-readable recording medium having recorded thereon a program for realizing this. In order to verify speech, a speech verification apparatus having a processor, A first function of constructing a semi-phoneme model by mixing a context-dependent (CD) phoneme and a context-independent (CI) phoneme; A second function of detecting an end point of the voice at the time of voice input, extracting only a portion in which the voice exists, and extracting a feature required for voice recognition from the extracted voice portion; A third function of performing speech recognition on the extracted speech data by searching for the Viterbi by referring to the pronunciation dictionary and the context-dependent phoneme model; And A fourth function of performing speech verification on speech recognized speech data based on the reliability obtained by referring to the context-dependent (CD) half-phoneme model and the context-dependent (CD) phoneme model A computer-readable recording medium having recorded thereon a program for realizing this.