KR20200129007A - Utterance verification device and method - Google Patents

Abstract

According to one embodiment of the present invention, a method for verifying utterance includes an operation of verifying utterance based on average phoneme recognition rank for utterance data and utterance transcription data. According to another embodiment of the present invention, the method may determine whether to perform utterance verification based on average phoneme recognition rank by measuring a misclassification distance for utterance data and utterance transcription data and using the measured misclassification distance.

Description

Ignition verification device and utterance verification method {UTTERANCE VERIFICATION DEVICE AND METHOD}

The following embodiments relate to a utterance verification apparatus and a utterance verification method.

In the speech recognition system, a technique for determining whether to judge the result of speech recognition as correct or incorrect is called speech verification.

The speech verification method includes feature data output from the search unit of the HMM (Hidden Markov Model) speech recognizer (e.g., anti-phoneme model) LLR (Log Likelihood Ratio) score, N-best ( N-best) LLR score, combination of LLR score, word duration, etc.), and SVM (Support Vector Machine) pattern classification method, etc. are used to determine whether or not it is misrecognized from these features.

In a speech recognition system, the purpose of speech verification is to reject out-of-vocabulary and vocabularies that may be misrecognized even in recognition target vocabularies. In this case, acceptance or rejection of speech recognition is determined by measuring the reliability of the recognized result. Reliability refers to the relative ratio of a phoneme or word, which is a recognized result when the corresponding vocabulary is uttered from a phoneme or word other than the recognized result. In general, reliability is measured from the viewpoint of verifying a statistical hypothesis.

The reliability of speech verification in speech recognition systems is mainly based on the anti-phoneme model. This is to measure the difference between the null hypothesis (H0) that the speech recognition was properly performed and the alternative hypothesis (H1) that the speech recognition was improperly performed. In speech verification, a confidence scale is obtained and compared with a predetermined threshold to determine acceptance/rejection of the speech recognition result.

As the reliability scale, the LRT (Likelihood Ratio Test)-based scale has been the most widely used, and an improvement method for modeling the alternative hypothesis through learning such as MCE (Minimum Classification Error) and MVE (Minimum Verification Error) has been proposed. In most of the current systems, a method of integrating candidate recognition results in the N-best list with various reliability measures is used instead of using a single measure.

According to an embodiment of the present invention, a speech verification based on an average phoneme recognition ranking may be performed for speech data and speech transcription data.

In addition, according to another embodiment of the present invention, the misclassification distance for the utterance data and the utterance transcription data may be measured, and whether to perform the speech verification based on the average phoneme recognition ranking may be determined using the measured misclassification distance.

Further, according to another embodiment of the present invention, speech verification may be performed on the speech data and the speech transcription data based on the average phoneme recognition ranking of speech data and speech transcription data phoneme sequences.

Further, according to another embodiment of the present invention, speech verification may be performed on speech data learned in an acoustic model and speech data having different speech styles.

According to an embodiment of the present invention, an operation of performing a speech verification based on an average phoneme recognition ranking for speech data and speech transcription data is included.

In addition, the operation of measuring the misclassification distance for the utterance data and the utterance transcription data based on the semiphone model, comparing the measured misclassification distance with a preset threshold, and verifying the average phoneme recognition ranking based on the comparison result. It may include an operation of determining whether to perform.

In addition, the speech verification method may further include an operation of determining appropriateness for the speech data and the speech transcription data based on the semiphoneme model.

In addition, the operation of determining whether to perform the average phoneme recognition ranking based speech verification may include, when the measured misclassification distance is greater than a preset threshold, the average phoneme recognition ranking basic speech verification ranking value is placed in the lowest ranking. If the measured misclassification distance is less than a preset threshold, it is possible to determine the execution of the speech verification based on the average phoneme recognition ranking in the speech data and the speech transcription data.

In addition, the average phoneme recognition ranking may be determined based on a previously learned acoustic model.

According to another embodiment of the present invention, in the speech verification apparatus, the speech verification apparatus includes a processor, and the processor performs speech verification based on an average phoneme recognition ranking for speech data and speech transcription data.

In addition, the processor measures the misclassification distance for the speech data and the utterance transcription data based on the semiphone model, compares the measured misclassification distance with a preset threshold, and ranks the average phoneme recognition according to the comparison result. You can decide whether to perform basic utterance verification.

Further, the processor may determine appropriateness of the speech data and the speech transcription data based on the semiphoneme model.

In addition, when the measured misclassification distance is greater than a preset threshold, the processor places the average phoneme recognition ranking basic speech verification ranking value as the lowest ranking, and the measured misclassification distance is less than a preset threshold. If it is small, it is possible to determine the execution of the speech verification based on the average phoneme recognition ranking in the speech data and the speech transcription data.

In addition, the average phoneme recognition ranking may be determined based on a previously learned acoustic model.

According to an embodiment of the present invention, there is an effect of performing a speech verification based on an average phoneme recognition ranking for speech data and speech transcription data.

In addition, there is an effect of measuring misclassification distances for utterance data and utterance transcription data, and determining whether to perform basic utterance verification based on the average phoneme recognition ranking by using the measured misclassification distance.

In addition, there is an effect of performing a speech verification on the speech data and the speech transcription data based on the average phoneme recognition ranking of speech data and speech transcription data phoneme sequences.

In addition, there is an effect of performing speech verification on speech data having different speech styles from speech data learned in the acoustic model.

1 is a diagram illustrating an operation of an utterance verification device according to an exemplary embodiment.
2 is a diagram showing the configuration of an utterance verification apparatus according to an embodiment.
3 is a diagram showing the distribution of speech data-speech transcription data according to misclassification distances.
4 is a diagram showing the distribution of speech data-speech transcription data according to an average phoneme recognition ranking.
5 is a flow chart showing a method of verifying utterance according to an embodiment.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in the present specification are exemplified only for the purpose of describing the embodiments according to the concept of the present invention, and the embodiments according to the concept of the present invention are It may be implemented in various forms and is not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention can apply various changes and have various forms, the embodiments will be illustrated in the drawings and described in detail in the present specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes all changes, equivalents, or substitutes included in the spirit and scope of the present invention.

Terms such as first or second may be used to describe various elements, but the elements should not be limited by the terms. The terms are only for the purpose of distinguishing one component from other components, for example, without departing from the scope of the rights according to the concept of the present invention, the first component may be referred to as the second component, and similarly The second component may also be referred to as a first component.

When a component is referred to as being "connected" or "connected" to another component, it is understood that it may be directly connected or connected to the other component, but other components may exist in the middle. Should be. On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there is no other component in the middle. Other expressions describing the relationship between components, such as "between" and "just between" or "adjacent to" and "directly adjacent to" should be interpreted as well.

The terms used in this specification are used only to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise.

In the present specification, terms such as "comprise" or "have" are intended to designate that the described feature, number, step, action, component, part, or combination thereof is present, but one or more other features or numbers It is to be understood that the possibility of addition or presence of, steps, actions, components, parts, or combinations thereof is not preliminarily excluded.

Unless otherwise defined, all terms including technical or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this specification. Does not.

In the following description, the same identification symbols mean the same configuration, and unnecessary redundant descriptions and descriptions of known technologies will be omitted.

In an embodiment of the present invention,'communication','communication network', and'network' may be used with the same meaning. The three terms refer to wired/wireless local and wide area data transmission/reception networks capable of transmitting and receiving files between a user terminal, a terminal of other users, and a download server.

Hereinafter, the present invention will be described in detail by describing a preferred embodiment of the present invention with reference to the accompanying drawings.

1 is a diagram illustrating an operation of an utterance verification device according to an exemplary embodiment.

Referring to FIG. 1, a speech verification apparatus 100 according to an exemplary embodiment receives speech data 101 and speech transcription data 102 and generates speech verification result data 103.

The speech data 101 and the speech transcription data 102 are respectively input to the speech verification apparatus 100, and the speech data 101 and the speech transcription data 102 may be different types of data.

The speech data 101 may be data in the form of a wave (WAV).

The speech data 101 may be speech of a voice actor.

The speech transcription data 102 may be data in the form of text (TXT).

The speech transcription data 102 may be a subtitle in a plain text format.

The speech verification result data 103 may be output in the form of a score.

The speech verification result data 103 may be output only when the speech verification result is equal to or higher than a preset level (eg, when appropriate).

The acoustic model 104 can be used in a decoder after learning a sound unit by deep learning technology and making it knowledgeable.

The acoustic model 104 may be modeled based on the HMM (Hidden Markov Model), which is a probability statistics method, but is not limited thereto.

The acoustic model 104 may be a pre-trained model, but is not limited thereto.

The acoustic model 104 may be used to classify the speech data 101 into phoneme units.

The acoustic model 104 may be the basis of the semiphoneme model.

An average phoneme recognition ranking may be determined based on the acoustic model 104.

2 is a diagram showing the configuration of a speech verification device according to an embodiment, FIG. 3 is a diagram showing the distribution of speech data-speech transcription data according to misclassification distances, and FIG. 4 is a utterance according to average phoneme recognition ranking It is a diagram showing the distribution of data-speech transcription data.

Referring to FIG. 2, the speech verification apparatus 100 includes a processor 110, an interface module 120, and a database 130.

The processor 110, the interface module 120, and the database 130 constituting the speech verification apparatus 100 are connected, and data can be transmitted to each other.

The processor 110 may include a memory (not shown) in which programs are stored.

The processor 110 may execute programs stored in the included memory (not shown).

The processor 110 may be connected to a memory (not shown) in which programs are stored.

The processor 110 may execute programs stored in the connected memory (not shown).

In this case, an operation program (eg, OS) for operating the speech verification device 100 may be stored in the memory (not shown).

The processor 110 may execute a program for managing information on the speech verification apparatus 100.

The processor 110 may execute a program for managing the operation of the speech verification apparatus 100.

The processor 110 may execute a program for managing the operation of the interface module 120.

The processor 110 may control the operation of the speech verification apparatus 100 by using the user's input received by the interface module 120.

Ⅰ) Forced alignment

Referring to FIGS. 1 and 2, the processor 110 may extract voice features from speech data 101 in 25 ms intervals every 10 ms.

The processor 110 may extract speech features from the speech data 101 using speech feature extraction techniques such as Mel Frequency Cepstral Coefficient (MFCC) and Filter BANK (FBANK).

The processor 110 may convert the speech transcription data 102 into a phone-sequence using a pronunciation dictionary (not shown). At this time, the pronunciation dictionary includes a word and a phone sequence corresponding to the pronunciation of the word.

The processor 110 applies a forced alignment to the speech features extracted from the speech data 101 and the phoneme string converted from the speech transcription data 102 to determine each phoneme of the phoneme sequence and the speech of the corresponding speech data 101. The signal section can be determined.

Ii) Basic speech verification of ANTI-PHONEME MODEL

The processor 110 may measure a misclassification distance for the speech data 101 and the speech transcription data 102 based on the semiphoneme model. In this case, the semiphoneme model may be based on the acoustic model 104.

The processor 110 may measure inappropriateness of the speech data 101 and the speech transcription data 102 using the measured misclassification distance.

The processor 110 may determine appropriateness for the speech data 101 and the speech transcription data 102 using the measured misclassification distance.

If the measured value of the misclassification distance is smaller than a preset value, the processor 110 may determine that between the speech data 101 and the speech transcription data 102 is appropriate.

If the measured value of the misclassification distance is greater than a preset value, the processor 110 may determine that between the speech data 101 and the speech transcription data 102 is not appropriate.

The processor 110 may perform speech verification on the speech data 101 and the speech transcription data 102 using the measured misclassification distance.

That is, the processor 110 measures the misclassification distance between the speech feature extracted from the utterance data 101 and the phoneme sequence converted from the utterance transcription data 102, and uses the measured misclassification distance. ) And the utterance transcript data 102 may be determined, and utterance verification may be performed using the determined appropriateness.

At this time, the more appropriate between the speech data 101 and the speech transcription data 102, the smaller the value of the misclassification distance.

The processor 110 may use the following [Equation 1] to measure the misclassification distance.

는 발화 데이터(u)와 발화 전사문 데이터(t) 사이의 음소인식 로그-우도(log-likelihood)를 의미하며,

는 발화 데이터(u)가 임의의 발화 전사문 데이터로 인식될 로그-우도(즉, 반음소 모델)를 나타낸다. here,

Denotes the phoneme recognition log-likelihood between speech data (u) and speech transcription data (t),

Denotes a log-likelihood (ie, a semi-phone model) in which the speech data u will be recognized as arbitrary speech transcription data.

는 발화검증 장치(100)로 입력된 발화 데이터(u)가 임의의 발화 전사문 데이터로 인식될 확률이 발화검증 장치(100)로 입력된 발화 전사문 데이터로 인식될 확률보다 얼마나 큰지를 나타낸다.In other words,

Denotes how much greater the probability that the utterance data u input to the utterance verification device 100 is recognized as arbitrary utterance transcription data is greater than the probability of being recognized as utterance transcript data input to the utterance verification device 100.

를 계산할 수 있다.The processor 110 uses [Equation 2] to

Can be calculated.

는 발화 전사문 데이터(t)의 음소열 중 i 번째 음소를 나타내며,

는 강제음소정렬 결과에서 i 번째 음소에 해당하는 음성 자질을 나타낸다.

는 음성 자질

가 음소

로 인식될 확률이며, N은 발화 전사문 데이터(t)의 전체 음소 수이다. 즉,

는 각 음소가 음향모델에 의해 인식될 로그-우도의 평균을 나타낸다.here,

Represents the i-th phoneme of the phoneme sequence of the speech transcription data (t),

Represents the voice quality corresponding to the i-th phoneme in the forced phoneme sorting result.

Voice qualities

Autumn phoneme

Is the probability of being recognized as, and N is the total number of phonemes in the spoken transcription data t. In other words,

Represents the average of log-likelihood that each phoneme will be recognized by the acoustic model.

를 계산할 수 있다.The processor 110 uses [Equation 3] to

Can be calculated.

는 i 번째 음소

가 임의의 다른 음소로 인식될 로그-우도를 나타내는 반음소 모델이다.here,

Is the i-th phoneme

Is a semiphoneme model representing log-likelihood to be recognized as any other phoneme.

는 음향모델을 통해 직접 구할 수 없으므로, 입력 받은 음소 이외의 음소의 인식확률을 통해 간접적으로 계산된다.However, the semiphone model

Is not directly obtained through the acoustic model, so it is indirectly calculated through the recognition probability of phonemes other than the received phoneme.

를 제외한 나머지 음소의 인식 로그-우도 값의 평균을 계산할 수 있다.The processor 110 uses [Equation 4] to determine the i-th phoneme in the entire phoneme set P.

It is possible to calculate the average of the recognized log-likelihood values of the other phonemes except for.

Iii) Basic speech verification of average phoneme recognition ranking

Referring to FIG. 3, FIG. 3(a) is a graph showing the misclassification distance for speech data-speech transcription data in a speech style similar to the acoustic model training data, and FIG. 3(b) This is a graph showing the misclassification distance for speech data of speech style-speech transcription data.

3(a), in the graph of FIG. 3(a), suitable speech data-speech transcription data 200 and inappropriate speech data-speech transcription data 210 are represented, and acoustic model training data and For utterance data of a similar utterance style, suitable utterance data-utterance transcription data 200 and inappropriate utterance data-utterance transcription data 210 are distinguished at a point 220 with a misclassification distance of -1.5, and suitable utterance data It can be seen that the portion 230 where the -speech transcription data 200 and the inappropriate speech data-speech transcription data 210 overlap is very small.

Referring to FIG. 3(b), in the graph of FIG. 3(b), suitable utterance data-utterance transcription data 240 and inappropriate utterance data-utterance transcription data 250 are represented,

For speech data of speech styles different from the acoustic model training data, appropriate speech data-speech transcription data 240 and inadequate speech data-speech transcription data 250 are distinguished at a point 260 with a misclassification distance of -1.1. It can be seen that the portion 270 where the appropriate speech data-speech transcription data 240 and the inappropriate speech data-speech transcription data 250 overlap is very large as compared to FIG. 3(a).

3(a) and 3(b), it can be seen that the performance of the semiphone model basic speech verification decreases for speech data of speech styles different from the acoustic model training data.

Referring to FIG. 4, FIG. 4(a) is a graph showing the average phoneme recognition ranking for speech data-speech transcription data in a speech style similar to the acoustic model training data, and FIG. 4(b) is a graph showing the acoustic model training data and It is a graph showing the average phoneme recognition ranking for speech data-speech transcription data of different speech styles.

Referring to FIG. 4(a), in the graph of FIG. 4(a), suitable speech data-speech transcription data 300 and inappropriate speech data-speech transcription data 310 are represented, and the acoustic model training data and For speech data of similar speech styles, suitable speech data-speech transcription data 300 and inappropriate speech data-speech transcription data 310 are distinguished from the point 320 where the average phoneme recognition rank is 4, and suitable speech data It can be seen that the -speech transcription data 300 and the inadequate speech data-speech transcription data 310 have almost no overlapping part (330).

Referring to FIG. 4(b), in the graph of FIG. 4(b), suitable utterance data-utterance transcription data 340 and inappropriate utterance data-utterance transcription data 350 are represented,

For speech data of speech styles different from the acoustic model training data, appropriate speech data-speech transcription data 340 and inappropriate speech data-speech transcription data 350 are distinguished at the point 360 where the average phoneme recognition rank is 6 It can be seen that the portion 370 where the appropriate speech data-speech transcription data 340 and the inappropriate speech data-speech transcription data 350 overlap is very small.

4(a) and 4(b), it can be seen that even for speech data of speech styles different from the acoustic model training data, the decline in performance of the basic speech verification by the average phoneme recognition ranking is very small.

The results shown in FIGS. 3 and 4 are analyzed because the probability of recognizing a suitable phoneme for speech data having characteristics different from the speech style learned in the acoustic model is generally lowered. However, contrary to the probability of falling phoneme recognition, the ranking of phoneme recognition does not change significantly.

The processor 110 may perform speech verification based on the average phoneme recognition ranking.

The processor 110 may use the following [Equation 5] to measure the average phoneme recognition ranking.

는 평균 음소인식 순위이고,

는 발화 전사문 데이터(t)의 음소열 중 i 번째 음소를 나타내며,

는 강제음소정렬 결과에서 i 번째 음소에 해당하는 음성 자질을 나타낸다. N은 발화 전사문 데이터(t)의 전체 음소수이다.

는 음성 자질

에 대하여 음소

의 인식순위를 가리킨다. here,

Is the average phoneme recognition ranking,

Represents the i-th phoneme of the phoneme sequence of the speech transcription data (t),

Represents the voice quality corresponding to the i-th phoneme in the forced phoneme sorting result. N is the total phoneme number of the speech transcription data t.

Voice qualities

About phoneme

Refers to the ranking of recognition.

값이 작을수록) 발화 데이터와 발화 전사문 데이터 사이가 적절한 것으로 판단할 수 있다.The processor 110, the higher the average phoneme recognition priority of the speech transcription data phoneme sequence (that is,

The smaller the value, the more appropriate between the speech data and the speech transcription data.

Iv) Stage 2 ignition verification

In order to solve the problem that occurs in speech verification based on a semiphoneme model for speech data of speech styles different from the acoustic model training data, the processor 110 performs speech verification based on the semiphoneme model and an average phoneme recognition ranking. You can combine based utterance verification.

When it is determined that the utterance data and the utterance transcript data have a certain level of relevance by the semiphone model, the processor 110 performs utterance verification based on the semiphoneme model and the speech verification based on the average phoneme recognition ranking. Combined two-step utterance verification can be performed.

The processor 110 may use the following [Equation 6] to combine speech verification based on a semiphoneme model and speech verification based on an average phoneme recognition ranking.

가 임계치

보다 크거나 같은 경우, 순위 값을 가장 낮은 순위로 배치한다. 이때, 가장 낮은 순위는 전체 음소 집합 P의 크기로 설정할 수 있다. 이때, 상기 임계치

는 발화 스타일에 따라 변할 수 있다.The processor 110 is a misclassification distance measured between the speech data (u) and the speech transcription data (t) in [Equation 6]

Fall threshold

If greater than or equal to, the rank value is placed in the lowest rank. In this case, the lowest priority may be set as the size of the entire phoneme set P. At this time, the threshold

Can change depending on the style of speech.

가 임계치

보다 크거나 같은 경우, 발화 데이터(u)와 발화 전사문 데이터(t) 사이의 오분류 거리

를 발화 데이터(u)와 발화 전사문 데이터(t) 사이의 발화검증의 척도로 사용할 수 있다. The processor 110 is a misclassification distance measured between the speech data (u) and the speech transcription data (t) in [Equation 6]

Fall threshold

If greater than or equal to, the misclassification distance between the speech data (u) and the speech transcription data (t)

Can be used as a measure of utterance verification between the utterance data u and the utterance transcription data t.

가 임계치

보다 크거나 같은 경우, 반음소 모델을 기초로 발화검증을 수행하여 발화검증 결과 데이터를 출력할 수 있다.The processor 110 is a misclassification distance measured between the speech data (u) and the speech transcription data (t) in [Equation 6]

Fall threshold

If it is greater than or equal to, the speech verification result data may be output by performing speech verification based on the semiphoneme model.

가 임계치

보다 작은 경우, 발화 데이터(u)와 발화 전사문 데이터(t) 사이의 평균 음소인식 순위

를 발화 데이터(u)와 발화 전사문 데이터(t) 사이의 발화검증의 척도로 사용할 수 있다. 이때, 상기 임계치

는 발화 스타일에 따라 변할 수 있다.The processor 110 is a misclassification distance measured between the speech data (u) and the speech transcription data (t) in [Equation 6]

Fall threshold

If smaller, the average phoneme recognition rank between the speech data (u) and the speech transcription data (t)

Can be used as a measure of utterance verification between the utterance data u and the utterance transcription data t. At this time, the threshold

May change depending on the style of speech.

가 임계치

보다 작은 경우, 평균 음소인식 순위를 기초로 발화검증을 수행하여 발화검증 결과 데이터를 출력할 수 있다.The processor 110 is a misclassification distance measured between the speech data (u) and the speech transcription data (t) in [Equation 6]

Fall threshold

If it is smaller than that, the speech verification result data may be output by performing speech verification based on the average phoneme recognition ranking.

의 인식순위가 높을수록(즉, 값이 작을수록) 발화 데이터(u)와 발화 전사문 데이터(t) 사이가 적절하다고 결정할 수 있다.The processor 110 is

It can be determined that the higher the recognition rank of (i.e., the smaller the value) is, the appropriate between the speech data u and the speech transcription data t.

The interface module 120 may receive the speech data 101 and the speech transcription data 102 from outside the speech verification apparatus 100.

The interface module 120 may output the speech verification result data 103.

The interface module 120 may be connected to the acoustic model 104 to transmit and receive data.

The interface module 120 may receive an input from a user of the speech verification apparatus 100.

The database 130 may store the misclassification distance measured by the processor 110.

The database 130 may store data received by the interface module 120 from the acoustic model 104.

The database 130 may store the speech data 101 and the speech transcription data 102 received by the interface module 120.

The database 130 applies a forced alignment to each phoneme of the phoneme sequence by applying a forced alignment to the phoneme sequence converted from the speech features extracted from the speech data 101 determined by the processor 110 and the speech transcription data 102. The audio signal section of the data 101 can be stored.

The database 130 may store speech verification result data 103.

The database 130 may store values calculated by the processor 110.

The term'device or module' used herein denotes a logical structural unit, and it is obvious to those skilled in the art that the present invention is not necessarily a physically classified component.

5 is a flow chart showing a method of verifying utterance according to an embodiment.

Referring to FIG. 5, in the utterance verification method according to an embodiment, the utterance verification apparatus acquires utterance data and utterance transcription data (400).

The speech data may be data in a wave (WAV) form, and the speech transcription data may be data in a text (TXT) form.

The speech verification device determines a speech signal section (410).

At this time, the speech verification device may determine each phoneme of the phoneme sequence and the speech signal section of the corresponding speech data 101 by applying a forced alignment to the speech features extracted from the speech data and the phoneme string converted from the speech transcription data. have.

The speech verification device measures the misclassification distance based on the semiphoneme model (420).

In this case, the speech verification apparatus may measure a misclassification distance for speech data and speech transcription data based on the semiphoneme model.

The speech verification device determines the appropriateness of the speech data and the speech transcription data based on the semiphoneme model (430).

In this case, the utterance verification device may determine appropriateness of the utterance data and the utterance transcription data by using the measured misclassification distance.

The ignition verification device compares the misclassification distance and the threshold value (440).

In this case, the threshold may be changed according to the speech style.

The speech verification device performs speech verification based on the average phoneme recognition ranking (450).

At this time, when the misclassification distance is greater than a preset threshold, the speech verification device may place the average phoneme recognition ranking basic speech verification ranking value at the lowest priority, and when the misclassification distance is less than a preset threshold, the speech data And, the speech verification based on the average phoneme recognition rank may be performed on the speech transcription data.

The speech verification device outputs speech verification result data (460).

In this case, the utterance verification device may output the utterance verification result only when the utterance verification result data exceeds a preset level (eg, when appropriate).

In the above, even if all the constituent elements constituting the embodiments of the present invention have been described as being combined into one or operating in combination, the present invention is not necessarily limited to these embodiments. That is, within the scope of the object of the present invention, all the constituent elements may be selectively combined and operated with at least one.

In addition, although all of the components may be implemented as one independent hardware, a program module that performs some or all functions combined in one or more hardware by selectively combining some or all of the components. It may be implemented as a computer program having Codes and code segments constituting the computer program may be easily inferred by those skilled in the art of the present invention.

Such a computer program is stored in a computer-readable storage medium, and is read and executed by a computer, thereby implementing an embodiment of the present invention. The computer program storage medium may include a magnetic recording medium, an optical recording medium, and the like.

In addition, the terms such as "include", "consist of" or "have" described above mean that the corresponding component may be embedded unless otherwise stated, excluding other components Rather, it should be interpreted as being able to further include other components.

All terms, including technical or scientific terms, unless otherwise defined, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms generally used, such as terms defined in the dictionary, should be interpreted as being consistent with the meaning in the context of the related technology, and are not interpreted as ideal or excessively formal meanings unless explicitly defined in the present invention.

The above description is merely illustrative of the technical idea of the present invention, and those of ordinary skill in the art to which the present invention pertains will be able to make various modifications and variations without departing from the essential characteristics of the present invention.

Accordingly, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention, but to explain the technical idea, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present invention.

100...Ignition verification device

Claims (11)

The operation of performing basic speech verification on the average phoneme recognition ranking for speech data and speech transcription data
The utterance verification method comprising a. The method of claim 1,
The firing verification method,
Measuring the misclassification distance for the speech data and the speech transcription data based on a semiphone model; And
An operation of comparing the measured misclassification distance with a preset threshold, and determining whether to perform basic speech verification based on the average phoneme recognition ranking according to the comparison result
The utterance verification method further comprising a. The method of claim 2,
The firing verification method,
Determining appropriateness of the speech data and the speech transcription data based on the semiphone model
The utterance verification method further comprising. The method of claim 2,
The operation of determining whether to perform speech verification based on the average phoneme recognition rank,
When the measured misclassification distance is greater than a preset threshold, the average phoneme recognition ranking basic speech verification ranking value is placed in the lowest ranking,
When the measured misclassification distance is less than a preset threshold, the speech verification method for determining the execution of the speech verification based on the average phoneme recognition ranking in the speech data and the speech transcription data. The method of claim 1,
The average phoneme recognition ranking is,
A speech verification method that is determined based on the acoustic model learned in advance. In the firing verification device,
The speech verification apparatus includes a processor,
The processor,
A speech verification device that performs basic speech verification on the average phoneme recognition ranking for speech data and speech transcription data. The method of claim 6,
The processor,
Measure the misclassification distance for speech data and speech transcription data based on the semiphoneme model,
A speech verification device that compares the measured misclassification distance with a preset threshold, and determines whether to perform basic speech verification based on an average phoneme recognition ranking according to a comparison result. The method of claim 7,
The processor,
A speech verification device that determines appropriateness for the speech data and the speech transcription data based on the semiphoneme model. The method of claim 7,
The processor,
When the measured misclassification distance is greater than a preset threshold, the average phoneme recognition ranking basic speech verification ranking value is placed in the lowest ranking,
When the measured misclassification distance is less than a preset threshold, the speech verification device determines to perform speech verification based on the average phoneme recognition ranking in the speech data and the speech transcription data. The method of claim 6,
The average phoneme recognition ranking is,
A speech verification device that determines based on a pre-learned acoustic model. A computer-readable recording medium storing a program for performing the method of any one of claims 1 to 5.

Images