KR20240119459A - A voice data processing device and operation method for automatic speech recognition model - Google Patents

Abstract

A voice data processing apparatus according to an embodiment of the present invention includes a recognition model building unit that pre-builds a multiple voice recognition model including a plurality of voice recognition models for voice recognition; a multi-recognition processor that inputs voice recognition target data into the multi-voice recognition model and obtains processing results for each of the voice recognition models; A character string similarity measurement unit that calculates character string recognition results for each speech recognition model according to the recognition results of the multiple speech recognition models, compares the character sequence recognition results, and calculates character sequence similarity between speech recognition models; and a data labeling unit configured to automatically label data of an output value using a recognition result of one or more speech recognition models among the plurality of speech recognition models, based on the character string similarity.

Description

{A voice data processing device and operation method for automatic speech recognition model}

The present invention relates to a voice data processing device and a method of operating the same. More specifically, the present invention relates to a voice data processing device for an automatic voice recognition model and a method of operating the same.

In general, automatic speech recognition (ASR), also known as speech-to-text (STT), involves the process of automatically recognizing and converting speech recordings into text.

Data for this STT can be largely viewed from an acoustic perspective and a linguistic perspective. From an acoustic point of view, environmental data such as speakers, space, and noise are mainly used, and from a linguistic point of view, language data for modeling vocabulary, context, and grammar are mainly used.

In addition, STT is largely divided into an offline learning stage that creates a recognition network model from voice/language data and an online exploration stage that recognizes the voice uttered by the user. The STT engine uses prior knowledge of voice and language data to recognize voice signals from voice signals. When outputting text information, the STT algorithm is also called a decoder in terms of interpretation. In the decoding stage, the input feature vector is compared with the model using the Acoustic Model, Language Model, and Pronunciation Lexicon, which are the results of the learning stage, and the word string is finally determined by scoring. .

Acoustic modeling is the process of generating representative patterns from the acoustic characteristics of pronunciation according to the phonological environment of the language using a probability model, and language modeling is the process of statistically learning the grammar system for usability issues of the language, such as vocabulary selection and sentence-level syntactic structure. As a process, in order to build a pronunciation dictionary, a phoneme conversion (Grapheme-to-Phoneme) implementation process that converts text phonetically is necessary, and pronunciation conversion rules targeting standard pronunciation are not enough to determine the dialect or user's speaking habits and tone. In some cases, it is difficult to reflect various patterns, so it is necessary to build a separate dictionary.

Accordingly, most of the acoustic models built are made up of probabilistic statistical methods or neural network-based deep learning learning methods, and the STT performance of individual models is improved according to the research of each model and the characteristics of each model.

Meanwhile, STT is being used as one of the natural language processing technologies for chatbot functions that have recently been implemented with various APIs, and is widely used in car navigation, kiosks, interactive IVR, virtual assistants, smart homes, etc. Chatbot technology is implemented by mixing TTS, STT, and AI processing technology.

However, the recognition accuracy of individual models for automatic speech recognition is still limited to about 90%, and recognition errors still occur. In particular, there is a problem with large variations in recognition accuracy due to household noise, inaccurate pronunciation characteristics of the elderly and children, etc.

In addition, in the case of an automatic speech recognition model in which words to be recognized are designated, they are converted to commonly used words in the grammar correction process, so the meaning of the actual speaker may be distorted, especially when recognizing words with similar pronunciations.

In particular, when using an automatic speech recognition model with a limited number of words that can be recognized, correction of incorrectly recognized words or letters is impossible, and as a result, it is impossible to overcome the decrease in recognition accuracy in real life. This automatic speech recognition method There is a need to improve recognition accuracy through fundamental changes to .

The present invention was developed to solve the problems described above, by processing the speech recognition prediction results for each model using multiple artificial intelligence speech recognition models into alphabet separation, and performing automatic data labeling according to the similarity of the separated alphabet strings. , Even if the number of words that can be recognized is limited, the recognition accuracy can be corrected, and in particular, it is a voice for automatic speech recognition model that can minimize the deviation of recognition accuracy due to everyday noise, inaccurate pronunciation characteristics of the elderly and children, etc. The purpose is to provide a data processing device and its operation method.

An apparatus according to an embodiment of the present invention for solving the problems described above includes a recognition model building unit that pre-constructs a multiple voice recognition model including a plurality of voice recognition models for voice recognition; a multi-recognition processor that inputs voice recognition target data into the multi-voice recognition model and obtains processing results for each of the voice recognition models; A character string similarity measurement unit that calculates character string recognition results for each speech recognition model according to the recognition results of the multiple speech recognition models, compares the character sequence recognition results, and calculates character sequence similarity between speech recognition models; and a data labeling unit configured to automatically label data of an output value using a recognition result of one or more speech recognition models among the plurality of speech recognition models, based on the character string similarity.

In addition, a method according to an embodiment of the present invention for solving the problems described above includes the steps of pre-constructing a multiple voice recognition model including a plurality of voice recognition models for voice recognition; Inputting voice recognition target data into the multiple voice recognition models and obtaining processing results of each of the voice recognition models; According to the recognition results of the multiple speech recognition models, calculating a consonant recognition result of each speech recognition model, comparing the consonant recognition results, and calculating consonant similarity between speech recognition models; Based on the character string similarity, correcting an output value using a recognition result of one or more voice recognition models among the plurality of voice recognition models based on a chatbot query; and outputting recognition result information corresponding to the voice recognition target data using the correction result and the character string similarity.

According to an embodiment of the present invention, the speech recognition prediction results for each model using a multi-artificial intelligence speech recognition model are separated into characters, and data labeling of the speech data is performed according to the similarity of the character strings that have been separated into characters for speech recognition. It can handle automatic labeling of data, and the recognition accuracy of each labeled voice data can be greatly improved compared to existing individual models.

Accordingly, according to an embodiment of the present invention, even when the number of words that can be recognized is limited, the recognition accuracy can be corrected, and in particular, the deviation in recognition accuracy due to living noise, inaccurate pronunciation characteristics of the elderly and children, etc. is minimized. A voice data processing device and a method of operating the same can be provided.

1 is a diagram schematically showing the entire system according to an embodiment of the present invention.
Figure 2 is a block diagram illustrating in more detail a voice data processing device according to an embodiment of the present invention.
Figure 3 is a flowchart for explaining the operation of a voice data processing device according to an embodiment of the present invention.
Figure 4 shows a data automatic labeling and correction table based on character string similarity to explain the voice recognition processing process according to an embodiment of the present invention.
Figure 5 is an exemplary diagram to explain the automatic voice labeling process according to an embodiment of the present invention.

The following merely illustrates the principles of the invention. Therefore, those skilled in the art will be able to invent various devices and methods that embody the principles of the present invention and are included in the concept and scope of the present invention, although not explicitly described or shown herein. In addition, all conditional terms and examples listed herein are, in principle, expressly intended only for the purpose of enabling the concept of the invention to be understood, and should be understood not as limiting to the examples and states specifically listed as such. do.

Additionally, it is to be understood that any detailed description reciting the principles, aspects, and embodiments of the invention, as well as specific embodiments, is intended to encompass structural and functional equivalents thereof. In addition, these equivalents should be understood to include not only currently known equivalents but also equivalents developed in the future, that is, all elements invented to perform the same function regardless of structure.

Accordingly, for example, the block diagrams herein should be understood as representing a conceptual view of an example circuit embodying the principles of the invention. Similarly, all flow diagrams, state transition diagrams, pseudocode, etc. are understood to represent various processes that can be substantially represented on a computer-readable medium and are performed by a computer or processor, whether or not the computer or processor is explicitly shown. It has to be.

Additionally, the clear use of terms such as processor, control, or similar concepts should not be construed as exclusively referring to hardware capable of executing software, and should not be construed as referring exclusively to hardware capable of executing software, including, without limitation, digital signal processor (DSP) hardware, and ROM for storing software. It should be understood as implicitly including ROM, RAM, and non-volatile memory. Other hardware for public use may also be included.

The above-described purpose, features and advantages will become clearer through the following detailed description in conjunction with the accompanying drawings, and accordingly, those skilled in the art will be able to easily implement the technical idea of the present invention. There will be. Additionally, in carrying out the present invention, if it is determined that a detailed description of known techniques related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted.

The terms used in this application are only used to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, terms such as “comprise” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but are not intended to indicate the presence of one or more other features. It should be understood that it does not exclude in advance the possibility of the existence or addition of elements, numbers, steps, operations, components, parts, or combinations thereof.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the attached drawings. In order to facilitate overall understanding when describing the present invention, the same reference numerals are used for the same components in the drawings, and duplicate descriptions of the same components are omitted.

FIG. 1 is a diagram schematically showing the entire system according to an embodiment of the present invention, and FIG. 2 is a block diagram illustrating in more detail a voice data processing device according to an embodiment of the present invention.

Referring to FIG. 2, a system according to an embodiment of the present invention may include a voice data processing device 100 and a user terminal 200.

More specifically, the voice data processing device 100 and the user terminal 200 may be connected to one or more of wired and wireless channels through a connection to a public network to transmit and receive data. The above-mentioned public network is a communication network built and managed by the state or a telecommunications carrier, and generally includes a telephone network, data network, CATV network, and mobile communication network, and provides connection services so that an unspecified number of ordinary people can access other communication networks or the Internet. .

Here, the public network network includes a local area network (LAN), a wide area network (WAN), a value added network (VAN), a personal area network (PAN), and a mobile communication network ( It can be implemented as any type of wired/wireless network, such as a mobile radio communication network or satellite communication network.

And the user terminal 200 described in this specification includes a personal computer (PC), a laptop computer, a mobile phone, a smart phone, a tablet PC, and a personal digital assistant (PDA). ), PMP (Portable Multimedia Player), etc. may be included.

In addition, the voice data processing device 100 and the user terminal 200 are not limited to the above device categories and may include server system-related devices that can be expanded by upgrading data processing, storage, and management functions.

Additionally, the voice data processing device 100 and the user terminal 200 may include respective communication modules for communicating using a protocol corresponding to each communication network.

And, the voice data processing device 100 may be a service providing device that provides a voice recognition service according to a request from the user terminal 200.

The voice data processing device 100 performs voice recognition and automatic labeling processing using the consonant string similarity of the multiple artificial intelligence voice recognition model according to an embodiment of the present invention from voice recognition target data received from the user terminal 200. It may be a service providing device that performs automatic labeling and provides voice recognition data to the user terminal 200.

In addition, the voice data processing device 100 is a voice recognition processing application that performs voice recognition processing using character string similarity of a multi-artificial intelligence voice recognition model according to an embodiment of the present invention and performs voice data correction and automatic labeling to enable the user. It may be a service providing device provided to the terminal 200.

In addition, the user terminal 200 and the voice data processing device 100 may operate as one voice recognition system, and voice recognition processing using the consonant string similarity of the multiple artificial intelligence voice recognition model according to an embodiment of the present invention Each part of the function may be distributed and processed between the user terminal 200 and the voice data processing device 100.

Therefore, in this specification, the description is focused on the configuration and operation of the voice data processing device 100, but the user terminal 200 may also have the configuration or perform the operations of the voice data processing device 100. It will be self-explanatory.

More specifically, the voice data processing device 100 according to an embodiment of the present invention includes a recognition model building unit 110, a multi-recognition processing unit 120, a character string similarity measurement unit 130, and a data labeling unit 140. ) includes.

First, the recognition model construction unit 110 pre-constructs a multiple voice recognition model including a plurality of voice recognition models for voice recognition.

Here, the multiple voice recognition model may include a plurality of voice recognition models. A plurality of voice recognition models can be classified and built in response to various currently known voice recognition methods, and can be divided into different models depending on the classification of the learning target, the category of voice learned, and the algorithm used. Can be classified and constructed.

More specifically, as shown in FIG. 2, the recognition model building unit 110 uses APIs of a plurality of voice recognition models provided for each service by company operating each voice recognition (ASR, Automatic Speech Recognition) model. By configuring a multi-processed voice recognition service process, a multi-voice recognition model can be built in advance.

For example, referring to Figure 2, the recognition model building unit 110 includes Google's Conformer service, Google's voice recognition API service, Facebook's wav2vec 2.0 service, Microsoft's Azure API service, Amazon's Transcribe API service, and IBM Multiple voice recognition models can be built in advance by configuring multiple uses of Watson Services, NEVER's Hyperclova service, SK Telecom's Adot service, and other open source-based Korean voice recognition model services using the KoSpeech Model. . The output of these multiple speech recognition models may be recognition results for each speech recognition model, and the recognition results may include word recognition results or letter recognition results composed of text data.

Additionally, each speech recognition model may be individually configured according to an individual learning model and may include one or more neural network models that perform deep learning using a data set according to the individual learning model. For example, the neural network model is a deep learning neural network model that learns voice data and text recognition results for specific categories such as age, gender, occupation, health, etc., and uses known CNN, RNN, or CRNN methods to classify each category. As star voices and their recognition results are learned, a learning model that outputs each voice recognition result can be built.

Then, the multi-recognition processing unit 120 inputs voice recognition target data into the multi-voice recognition model and obtains processing results for each of the plurality of voice recognition models. As described above, various recognition results may be output from models built according to each classification in the recognition model building unit 110. These recognition results may include word recognition results or character recognition results consisting of text data.

And, the character string similarity measurement unit 130 calculates character string recognition results for each speech recognition model from the recognition results of the multiple speech recognition models, compares the character string recognition results, and compares the character sequence recognition results between speech recognition models. Calculate similarity.

Here, the consonant sequence similarity measurement unit 130 separates the word recognition result or letter recognition result of the speech recognition model so that consonants and vowels are arranged into independent character codes, and calculates the consonant recognition result. .

For example, the letter string similarity measuring unit 130 may measure first language recognition result data mapped to the character code of the character recognition result or one or more character codes constituting the word recognition result to the initial consonant, middle consonant, or final consonant. By converting and combining each value into the corresponding relative distance position value, the alphabetic sequence recognition result can be calculated.

Accordingly, the character string similarity measurement unit 130 calculates the similarity between the character string recognition results of each of the plurality of voice recognition models, calculates the similarity between the calculated character string recognition results, and determines the similarity. It can be calculated.

More specifically, the consonant string similarity measurement unit 130 separates the consonant characters and vowel characters of the recognition results for each word or the recognition result for each character in the order of pronunciation, such as initial consonant, middle consonant, and final consonant, based on spacing. , By converting by character code, the character codes can be separated and configured, and the character string recognition results can be obtained by connecting the separately composed character codes. For example, the voice recognition results for the Korean character 'Chat' are: As the initial consonant, middle consonant, and final consonant are each separately converted and processed based on Unicode, such as 'ㅊㅐㅅ', in the alphabet string similarity measurement unit 130, the alphabet string recognition result can be obtained.

Additionally, the letter string similarity measuring unit 130 may measure the letter string similarity between each speech recognition model by comparing the separately constructed letter string recognition results. This method of separating and comparing letter sequences makes it possible to prevent errors in the inappropriate comparison process for words with similar pronunciations in existing word similarity or letter similarity comparisons.

For example, if the first recognition model recognizes a specific voice as 'Shin Ramyun' and the second recognition model recognizes it as 'Jin Ramyun', in the word similarity-based multiple recognition method, 'Jin Ramyun' and 'Shin Ramyun' are completely Because the words are not similar, there is a problem in that it is impossible to determine where the incorrect recognition occurred in the first recognition model or the second recognition model, and it is inevitably output as a recognition failure.

Furthermore, even in the character similarity-based recognition process, 'Jin' and 'Shin' are completely different letters, so only recognition failure occurs, and there is a problem in that the system itself cannot correct this or correct the error.

On the other hand, the letter string similarity measurement unit 130 according to an embodiment of the present invention can measure the letter string similarity, and accordingly, the 'Shin Ramyun' recognition result of the first model is 'ㅅㅣㄴㄹㅏㅁㅕㄴ'. The consonants are separated into 'Jin Ramyun' in the second model, and the recognition result for 'Jin Ramyun' is separated into 'ㅈㅣㄴㄹㅏㅁㅕㄴ', and the consonant similarity of the recognition results between the first and second models is more than 70% similar. Therefore, it can be confirmed that only the first consonant has a difference in recognition, and ultimately, location information of the character code that can correct or delete this can be derived.

In addition, in order to efficiently process such alphabet string separation, the alphabet string similarity measurement unit 130 provides a first language recognition result mapped to the character code of the letter recognition result or one or more character codes constituting the word recognition result. By converting and combining the data into relative distance position values corresponding to the initial consonant, middle consonant, or final consonant, the alphabet string recognition result can be calculated.

For example, the alphabet used in Korean Unicode consists of 19 initial consonants, 21 middle consonants, and 27 final consonants, and the number of characters that can be combined is 19x21x28 (including no final consonants) = 11,172 characters. In a string composed of Hangul character codes, the distance between the final consonant, the middle consonant, the middle consonant, and the final consonant can always be configured to be constant.

Accordingly, the alphabet string similarity measurement unit 130 converts the voice recognition result into Unicode, and calculates the relative distance between the initial consonant, middle consonant, and final consonant for each character of the string converted to Unicode to the starting position of Unicode (' The alphabet string recognition result can be constructed by calculating and connecting each according to the distance from 'a').

For example, in the case of '? The distance between initial consonants up to the initial consonant 'ㅊ' is calculated to be 14, so when each relative distance location information is constructed as a result of consonant string recognition, a separate consonant called 'ㅊㅞㅅ' is formed by converting and connecting the character codes mapped therefrom. A recognition result may be obtained.

In the case of Korean character recognition, it can be performed as follows by configuring the alphabet string composition code.

*********************************************

//Configure initial consonant mapping character code

chut = '#'

//Configure neutral mapping character code

ga = 'ㅏㅐㅑㅒㅓㅔㅕㅖㅗㅘㅙㅚㅛㅜㅝㅞㅟㅠㅡㅢㅣ#'

//Configure final consonant mapping character code

ggut = '#'

//Set Unicode start position

BASE = 0xAC00

//Input voice recognition result

query = '?X'

//Composition of alphabetic string recognition results based on relative position information, *ord() function is a function that converts the target character value to ASCII number.

code = ord(query) - BASE

//Extract final relative location information

jongsung = code % 28

//Extract neutral relative location information

jungsung = ((code-jongsung) // 28) % 21

//Extract relative position information of initial consonant

chosung = ((code - jongsung) // 28) // 21

//Output character code mapped to relative location information

print(chut[chosung], ga[jungsung], ggut[jongsung])

*********************************************

The result of executing the above code may be as follows.

>>> ㅊ ㅞ ㅅ

Here, the chut variable (=initial sound), ga variable (=middle sound), and ggut variable (=final sound) are variables commonly used in alphabet combinations.

Meanwhile, the letter string similarity measurement unit 130 can measure the letter string similarity between voice recognition models by comparing the obtained letter string recognition results for each multiple speech recognition model. Here, the ratio method of the known SequenceMatcher algorithm can be used to measure string similarity, and accordingly, the string similarity between speech recognition models can be derived using a function that calculates each similarity. This function is a function that compares character strings and outputs the similarity. For example, in the case of manana, which is similar to banana, it can output a result such as 83% similarity.

Here, the alphabetic sequence similarity can be calculated between each speech recognition model and the speech recognition model, and the final alphabet string similarity can be determined by combining the alphabetical sequence similarities calculated between each speech recognition model. This may be composed of the average value, median value, or various statistical values calculated by combining the alphabet string similarity values calculated between each speech recognition model.

Then, the data labeling unit 140 identifies the consonant string similarity for each word of the voice recognition target data, and when one or more words whose consonant string similarity for each word is within a threshold condition is identified, the voice recognition target data is based on the data. The output value of can be corrected, data automatic labeling can be processed, and the processing results can be stored or transmitted to the user terminal 200.

Here, the data labeling unit 140 can correct the data automatic labeling information of the word corresponding to the threshold condition by using the recognition results for each speech recognition model of the word whose letter string similarity corresponds to the threshold condition. there is. For example, the data labeling unit 140 may amplify the speech recognition target data and then delete the portion of the speech waveform corresponding to the word whose consonant string similarity meets the threshold condition.

Accordingly, the processing result of the data labeling unit 140 may be stored and managed in a separate database (not shown) as voice recognition data as automatically labeled voice data, or provided to the user terminal 200. .

Here, the data labeling and correction processing of the speech data determines whether the recognition result for each speech recognition model of the identified word matches the threshold condition of the alphabet string similarity for each word, using pre-configured alphabet string similarity-based data. Depending on the application to the automatic labeling and correction table, it can be performed optionally. The automatic data labeling and correction table based on alphabetical string similarity and the operation of the data labeling unit 140 accordingly will be described in more detail later with reference to FIGS. 4 and 5.

In addition, the voice data with automatically labeled recognition data output from the data labeling unit 140 may include model identification information corresponding to each of one or more voice recognition models, and may include recognition obtained from any model among the multiple recognition models. It is possible to check whether it is a result, and this is fed back to the recognition model building unit 110 so that it can be used to update learning data and improve reliability.

Figure 3 is a flowchart for explaining the operation of a voice data processing device according to an embodiment of the present invention.

Referring to FIG. 3, the voice data processing apparatus 100 according to an embodiment of the present invention builds in advance a multiple voice recognition model including a plurality of voice recognition models for voice recognition (S101).

Thereafter, the voice data processing apparatus 100 inputs the voice recognition target data into a multiple voice recognition model including a plurality of voice recognition models for voice recognition, and obtains processing results of each of the voice recognition models. (S103).

Then, the speech data processing apparatus 100 processes the recognition results of the speech recognition models into character separation for each word, compares the resulting character separation processing, and measures the similarity of character strings for each word (S105).

Thereafter, the voice data processing apparatus 100 processes automatic data labeling of the voice recognition target data based on the letter string similarity for each word (S107).

Figure 4 shows a data automatic labeling and correction table based on character string similarity for explaining the voice recognition processing process according to an embodiment of the present invention, and Figure 5 is an example for explaining the automatic voice labeling process according to an embodiment of the present invention. It's a degree.

4 and 5, referring to FIG. 4, the voice data processing apparatus 100 according to an embodiment of the present invention identifies words whose consonant string similarity for each word corresponds to a threshold condition. Correction of recognition results can be selectively performed by applying the matching recognition results for each speech recognition model to a pre-configured alphabet string similarity-based data automatic labeling and correction table.

Figure 4 is an example of such a data automatic labeling and correction table based on alphabet string similarity. The data labeling unit 140 selects words to be automatically labeled according to the alphabet string similarity measurement result of the recognition result for each speech recognition model (ASR). can be decided.

Referring to FIG. 4, first, if the recognition results of all models match the consonant sequence similarities, the recognition results can be automatically labeled without additional correction. However, if there is one or more words whose alphabetic similarity is measured to be within a critical condition (for example, 2 out of 3 matches but 1 does not match, which is less than 100%), the word is excluded from the recognition target. Exception processing such as exclusion may occur.

In this way, referring to FIG. 4, when recognizing a user's voice, multiple voice recognition models are utilized, and by comparing the recognition results between the multiple voice recognition models according to character string similarity, it is possible to check whether there are errors in the recognition results. You can. In addition, the voice data processing apparatus 100 can determine in advance where and how many such errors exist and perform appropriate correction or exception processing corresponding thereto, thereby producing accurate recognition results. In addition, as the recognition results are used as learning data for each model, accuracy can be continuously improved based on feedback.

And, referring to Figure 5, if there is one or more words whose letter string similarity meets the threshold condition, such as a word with a similarity of less than 100%, exception processing such as excluding the word from recognition may be performed, In the case of voice data for which the exception processing is performed, the data labeling unit 140 may amplify the voice waveform and then perform deletion processing on the corresponding voice portion based on the portion without the waveform (spaced portion). In this case, the ability to distinguish between spaced parts without waveforms is increased, and the waveform parts that are not properly recognized and may cause errors are deleted, thereby improving the labeling accuracy of recognition data.

As a result, referring to Figures 4 and 5, if the recognition result based on the consonant similarity between all multi-speech recognition models in the voice 'It is delicious to put tuna in Jin Ramyun' is 100%, the result is 'If you put tuna in Jin Ramyun, it is delicious' ' is mapped as is, but 'Jin Ramyun' is recognized as 'Shin Ramyun', and if a word with a letter sequence similarity of less than 100% is generated, the word is deleted from the labeling target and 'It is delicious with tuna' is automatically changed. Labeled, and after amplification of the voice data, only the portion corresponding to the voice section is deleted, so the automatically labeled data can be mapped to the corrected voice data and stored and managed.

The method according to the present invention described above can be produced as a program to be executed on a computer and stored in a computer-readable recording medium. Examples of computer-readable recording media include ROM, RAM, CD-ROM, and magnetic tape. , floppy disks, optical data storage devices, etc., and also includes those implemented in the form of carrier waves (for example, transmission via the Internet).

The computer-readable recording medium is distributed in a computer system connected to a network, so that computer-readable code can be stored and executed in a distributed manner. And, functional programs, codes, and code segments for implementing the method can be easily deduced by programmers in the technical field to which the present invention pertains.

In addition, although preferred embodiments of the present invention have been shown and described above, the present invention is not limited to the specific embodiments described above, and the technical field to which the invention pertains without departing from the gist of the present invention as claimed in the claims. Of course, various modifications can be made by those skilled in the art, and these modifications should not be understood individually from the technical idea or perspective of the present invention.

Claims (14)

a recognition model building unit that pre-builds a multiple voice recognition model including a plurality of voice recognition models for voice recognition;
a multi-recognition processor that inputs voice recognition target data into the multi-voice recognition model and obtains processing results for each of the voice recognition models;
A character string similarity measurement unit that calculates character string recognition results for each speech recognition model according to the recognition results of the multiple speech recognition models, compares the character sequence recognition results, and calculates character sequence similarity between speech recognition models; and
Based on the character string similarity, a data labeling unit that processes automatic data labeling of the output value using the recognition result of one or more speech recognition models among the plurality of speech recognition models.
Voice data processing device. According to paragraph 1,
The alphabet string similarity measurement unit,
Separating the word recognition results or letter recognition results of the speech recognition model so that consonants and vowels are listed as independent character codes to calculate the consonant string recognition results.
Voice data processing device. According to paragraph 2,
The alphabet string similarity measurement unit,
Converting and combining the first language recognition result data mapped to the character code of the character recognition result or one or more character codes constituting the word recognition result into relative distance position values corresponding to the initial consonant, middle consonant, or final consonant, respectively, Calculating the result of character string recognition
Voice data processing device. According to paragraph 3,
The alphabet string similarity measurement unit,
Calculating the similarity between the character string recognition results of each of the plurality of voice recognition models, and calculating the similarity between the calculated character string recognition results to calculate the similarity.
Voice data processing device. According to paragraph 1,
The data labeling unit,
Identifying the string similarity for each word of the voice recognition target data, and correcting the data automatic labeling information of the output value when one or more words whose string similarity for each word corresponds to a threshold condition are identified.
Voice data processing device. According to clause 5,
The data labeling unit,
Using the recognition results for each speech recognition model for words whose letter string similarity corresponds to the threshold condition, the data automatic labeling information of the word corresponding to the threshold condition is corrected.
Voice data processing device. According to clause 6,
The data labeling unit,
After amplifying the voice recognition target data, the portion of the voice waveform corresponding to the word whose alphabet string similarity meets a threshold condition is deleted and processed.
Voice data processing device. Pre-building a multi-speech recognition model including a plurality of speech recognition models for speech recognition;
Inputting voice recognition target data into the multiple voice recognition models and obtaining processing results of each of the voice recognition models;
According to the recognition results of the multiple speech recognition models, calculating a consonant recognition result of each speech recognition model, comparing the consonant recognition results, and calculating consonant similarity between speech recognition models; and
Based on the character string similarity, a data labeling step of automatically labeling data of an output value using a recognition result of one or more speech recognition models among the plurality of speech recognition models.
Method of operating a voice data processing device. According to clause 8,
The step of calculating the alphabetical sequence similarity is,
Separating the word recognition results or letter recognition results of the speech recognition model so that consonants and vowels are listed as independent character codes to calculate the consonant string recognition results.
Method of operating a voice data processing device. According to clause 9,
The step of calculating the alphabetical sequence similarity is,
Converting and combining the first language recognition result data mapped to the character code of the character recognition result or one or more character codes constituting the word recognition result into relative distance position values corresponding to the initial consonant, middle consonant, or final consonant, respectively, Including the step of calculating a character string recognition result.
Method of operating a voice data processing device. According to clause 10,
The step of calculating the alphabetical sequence similarity is,
Comprising the step of calculating the similarity between the character string recognition results of each of the plurality of voice recognition models, calculating the similarity between the calculated character string recognition results, and calculating the similarity of the character string.
Method of operating a voice data processing device. According to clause 8,
The data labeling step is,
Identifying consonant string similarity for each word of the voice recognition target data, and when one or more words whose consonant string similarity for each word corresponds to a threshold condition are identified, correcting data automatic labeling information of the output value.
Method of operating a voice data processing device. According to clause 12,
The data labeling step is,
Comprising the step of correcting the data automatic labeling information of the word corresponding to the threshold condition, using the recognition results for each speech recognition model of the word whose letter string similarity corresponds to the threshold condition.
Method of operating a voice data processing device. According to clause 13,
The data labeling step is,
After amplifying the speech recognition target data, deleting a portion of the speech waveform corresponding to a word whose alphabet string similarity corresponds to a threshold condition.
Method of operating a voice data processing device.