KR20200011176A - Korean speech recognition system with pronunciation classification based on vowel pronunciation - Google Patents

Abstract

The present invention relates to a Korean speech recognition and conversion system with pronunciation classification focusing on vowel pronunciation. The Korean speech recognition and conversion system recognizes one syllable focusing on vowel and converts one character including at least one consonant and vowel focusing on vowel. By providing a method to classify vowels in a method different from an existing vowel classification method and to provide the speech pronunciation characteristics focusing on vowel pronunciation, the Korean speech recognition and conversion system may recognize speech with excellent recognition rate and convert speech with natural pronunciation.

Description

Korean speech recognition system with pronunciation classification based on vowel pronunciation}

The present invention relates to a pronunciation classification method centering on vowel pronunciation and a Korean speech recognition system in which the classification method is introduced.

Here, background art is provided with respect to the present disclosure, and these do not necessarily mean known art.

Voice is the most basic means of communication among the various means that humans use to communicate with each other. The voice recognition technology converts a person's voice into text to be recognized by an electronic device such as a computer, a telephone, a mobile phone, and the like. A lot of research has been done with respect to speech recognition and speech conversion for a long time, and recently, speech recognition and speech conversion technologies have been commercialized in various fields.

In general, speech recognition is performed as follows. When voice is input through the microphone, the input voice signal is converted into a digital signal. After extracting the speech feature from the converted speech signal, recognition is made by comparison with a code book stored in the memory, and the result is processed. Here, the codebook is a collection of code words that are reference models used for recognition determination. In general, a speech signal is encoded into a code word after sampling and quantization.

On the other hand, in the case of speech conversion, a sound database for the phonemes is constructed and connected to generate continuous speech, and the natural speech is synthesized by adjusting the size, length, and height of the speech. That is, when text is input, the grammatical structure of the input document is analyzed in the language processing step, and then a rhyme like human reading is generated by the analyzed document structure, and the basic units of the stored voice DB are collected according to the generated rhyme. The waveform synthesis step is generated.

The speech recognition and speech conversion techniques such as the unique speech features of the language, in particular, pronunciation features are considered in consideration of the higher recognition rate of speech recognition, and the natural voice conversion is possible, the conventional Korean speech recognition technology is a unique feature of Korean pronunciation Korean voice recognition technology is developed by introducing foreign technology as it is, without recognition, and there is a problem that the recognition rate decreases and sounds awkward during voice conversion.

The present invention has been made to solve the above problems and provides a method for assigning a phonetic pronunciation feature according to Korean phonetic pronunciation features, and using Korean phonetic recognition data classified using the phonetic pronunciation features. It is intended to be used for voice conversion technology.

However, the objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

The present invention provides a Korean speech recognition system in which pronunciation classification based on vowel pronunciation is introduced.

The pronunciation classification based on the vowel pronunciation includes a first classification method of classifying the vowel into three series vowels including the first series vowel, the second series vowel, and the third series vowel according to sound. It is done.

The first series vowel includes ㅏ, ㅏ, ㅗ, TT, ㅡ,, ㅐ, ㅔ, and the second series vowel includes ㅑ, ㅕ, ㅛ, ㅠ, ㅒ, ㅖ, and the third Series vowels are characterized by including ㅘ, ㅙ, ㅞ, ㅚ, ㅝ, ㅟ, ㅢ.

In addition, the pronunciation classification based on the vowel pronunciation includes a second classification method of giving two pronunciation shapes including a first pronunciation shape and a second pronunciation shape for each vowel.

In addition, the pronunciation shape is characterized by using the first series vowels and the second series vowels.

In addition, in the second classification method, the first series vowel and the third series vowel are given two pronunciation shapes using the first series vowel, and the second series vowel and the first series vowel; It is characterized by giving two pronunciation shapes using the second series of vowels.

In addition, in the second classification method, the first series vowels are assigned the same first series vowels as the first pronunciation shape and the second pronunciation vowels, and the second series vowels are assigned the second series vowels as the first pronunciation vowels. And a first series vowel as a second pronunciation shape, and a third series vowel as a first pronunciation vowel and a second pronunciation vowel as a second pronunciation vowel.

In addition, the pronunciation classification based on the vowel pronunciation includes a third classification method for applying stress to any one of the first and second pronunciation shapes given for each vowel. .

In the third classification method, in the case of the first series vowel and the second series vowel, the accent is applied to the first pronunciation shape among the two pronunciation shapes, and in the case of the third series vowel, among the two pronunciation shapes. It is characterized by giving stress to the second pronunciation pattern.

In addition, the Korean voice recognition system is characterized in that the system for converting the input voice to the corresponding text using a voice recognition database stored pronunciation classification centered on the vowel pronunciation.

In addition, the Korean voice recognition system includes a voice input unit for receiving a voice; An A / D converter for converting a voice input through the voice input unit into a digital voice signal; A feature extraction unit for extracting a pronunciation feature from the digital voice signal converted by the A / D converter; And a determination unit for determining which character corresponds to the pronunciation classification extracted from the feature extraction unit by comparing with the pronunciation classification stored in the speech recognition database.

The present invention provides a method of classifying a vowel and assigning a voice pronunciation feature centering on the vowel pronunciation by a method different from the existing method of classifying a vowel, and providing a voice pronunciation database configured according to the given voice pronunciation feature. And by being able to be applied to the voice conversion technology it is possible to provide an effect capable of speech recognition with a good recognition rate when converting the voice to the text and a natural pronunciation when converting the text to the voice.

Prior to describing the present invention in detail below, it is understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention, which is limited only by the scope of the appended claims. shall. All technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art unless otherwise indicated.

Throughout this specification and claims, unless otherwise indicated, the termcomprise, comprises, and configure means to include the referenced article, step, or group of articles, and step, and any other article It is not meant to exclude a stage or group of things or a group of stages.

On the other hand, various embodiments of the present invention can be combined with any other embodiment unless clearly indicated to the contrary. Any feature indicated as particularly preferred or advantageous may be combined with any other feature and features indicated as preferred or advantageous. Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention and the effects thereof.

Unlike other foreign languages, especially English, Korean pronunciation is not pronounced only by consonants, but it is possible to pronounce consonants only when it meets vowels. Although vowels are the criterion of Korean syllable pronunciation, there is a problem in that the recognition rate is lowered and awkward speech is provided because the Korean vowel classification is not considered in applying speech recognition or speech conversion techniques.

The present invention provides a speech pronunciation database constructed according to these characteristics by classifying vowels and assigning vocal pronunciation features centering on vowel pronunciations, and applying them to speech recognition and speech conversion techniques. It provides speech recognition with recognition rate and voice conversion with natural pronunciation.

First, the method of classifying the existing vowels is a short vowel (ㅏ, ㅔ, ㅓ, ㅐ, ㅗ, ㅚ, ㅜ, ㅟ, ㅡ, ㅣ) that is fixed because the lips or tongue are fixed when pronounced and the shape of the lips when pronounced. It is divided into two vowels (ㅑ, ㅒ, ㅕ, ㅖ, ㅘ, ㅙ, ㅛ, ㅝ, ㅞ, ㅠ, ㅢ) that the tongue changes. In addition, the short vowels may be classified into legend vowels and rear vowels depending on whether the position of the highest point of the tongue is in the front or the rear when pronounced as shown in Table 1, and whether the position of the tongue is high, middle, or low. Therefore, it can be classified into high vowel (closed vowel), middle vowel and low vowel (open vowel), and is divided into round vowels and flat vowels according to whether the lips make a sound or flat sounds.

Peak position of the tongue Legend collection Husserl Collection The height of the tongue / the shape of the lips Normal Original Normal Original High vowels ㅣ ㅟ ㅡ TT Heavy vowels ㅔ ㅚ ㅓ ㅗ Low vowel ㅐ ㅏ

As mentioned above, the vowel classification system is not suitable for speech recognition and speech conversion because the vowel is classified, but the sound cannot be accurately separated because there is no classification system according to the pronunciation of the vowel.

In general, the recognition of Korean sounds means syllables that combine consonants and vowels, that is, one syllable when at least one consonant and vowels are provided. I want to recognize it as the center.

Korean has a vowel sound. The vowel center means that without the vowel, the sound of the consonant and the consonant are not heard. Consonants have sounds, but consonants themselves cannot be pronounced as a single sound. This is a unique pronunciation characteristic of Korean, which has a great difference from the unique pronunciation characteristic of English. For example, the English consonants P and F have sound effects on their own, such as the Korean consonants P, F, and Korean consonants P.

 Accordingly, the present invention provides a classification method using vowels as syllables based on the pronunciation characteristics of Korean. In other words, in constructing a phonetic pronunciation database used for Korean speech recognition and voice conversion, a classification method is used to classify syllables based on vowels rather than classifying existing syllables with consonants and vowels combined.

In addition, according to the present invention, the vowels are first classified into three series of A series, Y series, and W series according to sound. Next, two kinds of pronunciation shapes are assigned to all the vowels classified as follows, and accentuation is given to at least one of the two pronunciation shapes to give a pronunciation form. That is, vowel pronunciation is primarily classified according to sound, and secondly, by two kinds of pronunciation shapes, and thirdly, by phonological forms, so that pronunciation can be distinguished with a clearer and higher recognition rate.

The vowel classification method according to the present invention is summarized as shown in Table 2 below.

sound Pronunciation shape Pronunciation form
(Pronounced pronunciation) First pronunciation Second pronunciation ㅏ A series
(ㅏ [a] series) ㅏ ㅏ First pronunciation shape ㅓ ㅓ ㅓ First pronunciation shape ㅗ ㅗ ㅗ First pronunciation shape TT TT TT First pronunciation shape ㅡ ㅡ ㅡ First pronunciation shape ㅣ ㅣ ㅣ First pronunciation shape ㅐ ㅐ ㅐ First pronunciation shape ㅔ ㅔ ㅔ First pronunciation shape ㅑ J series
(Ja [ja] series) ㅑ ㅏ First pronunciation shape ㅕ ㅕ ㅓ First pronunciation shape ㅛ ㅛ ㅗ First pronunciation shape ㅠ ㅠ TT First pronunciation shape ㅒ ㅒ ㅐ First pronunciation shape ㅖ ㅖ ㅔ First pronunciation shape ㅘ W series
(ㅘ [wa] series) ㅗ ㅏ Second pronunciation shape ㅙ ㅗ ㅐ Second pronunciation shape ㅞ TT ㅔ Second pronunciation shape ㅚ ㅗ ㅔ Second pronunciation shape ㅝ TT ㅓ Second pronunciation shape ㅟ TT ㅣ Second pronunciation shape ㅢ ㅡ ㅣ Second pronunciation shape

As shown in Table 2, 모음, ㅓ, ㅗ,, ㅣ, ㅐ, ㅔ of the vowels are classified as A series, and ㅑ, ㅕ, ㅛ, ㅠ, ㅒ, 로 as J series, ㅘ, ㅙ , ㅞ, ㅚ, ㅝ, ㅟ, ㅢ are classified into W series based on sound.

Next, two phonetic shapes are given to all vowels without distinguishing between the previously classified short vowels and double vowels. As shown in Table 2, the first pronunciation shape and the second pronunciation shape are given to all vowels. The vowels given by the phonetic shape use the A series and J series collections, except for the W series collections.

In this case, two phonetic shapes are assigned to the A-series and the W-series vowels using the A-series vowels, and two pronunciation shapes are assigned to the J-series vowels using the A-series and J-series vowels.

More specifically, in the case of the A series vowel, the same A series vowel is assigned to the first pronunciation shape and the second pronunciation shape. In addition, in the case of the J series vowel, the J series vowel is given as the first pronunciation shape and the A series vowel is given as the second pronunciation shape. In addition, in the case of the W series vowels, different A series vowels are assigned to the first pronunciation shape and the second pronunciation shape.

According to the classification method according to the present invention, even in the case of a vowel previously classified as a single vowel, two phonetic shapes are given. For example, in the case of 되는 previously classified as a short vowel, [ㅔ, ㅔ] is classified as [ㅗ, ㅔ].

The inventors of the present invention found that the vowels affect the consonants and the bass sounds, respectively, when the vowels are sounded, thereby recognizing that the vowels are the standard sounds of Korean syllables. The recognition probability was increased and natural pronunciation was made possible.

According to the present invention, when a letter is pronounced, it may be divided into two sounds. That is, in the case of one letter consisting of consonants + vowels + consonants, it is divided into two sounds, [consonants + vowels (first pronunciation)] and [vowels (second pronunciation) + support]. One letter is divided into two sounds, [consonant + vowel (first pronunciation)] and [vowel (second pronunciation)].

Finally, the accent is given to any one of the two pronunciation shapes. That is, it distinguishes which sounds are pronounced more strongly from two sounds in one letter, that is, [consonants + vowels (first pronunciation)] and [vowels (second pronunciation) + (support)].

As shown in Table 2, in the case of the A-series vowels and the J-series vowels, stress is given to the first pronunciation shape among the two pronunciation shapes, and in the case of the W-series vowels, the stress is applied to the second pronunciation shape among the two pronunciation shapes. In addition, the recognition probability was increased and natural pronunciation was possible.

For convenience of description, the sequence, the first pronunciation shape, and the second pronunciation shape are described in order in square brackets, and double quotation marks are displayed on the accented pronunciation shapes among the first and second pronunciation shapes. For example, ㅚ is expressed as [W, ㅗ, “ㅔ”].

For example, the pronunciation of the word 'eye' can be divided into two sounds around the vowel TT. First, the first pronunciation shape of the vowel TT and the consonant b are combined, and the [nu] sound is pronounced first, and the [pronounce] sound is pronounced after being combined with the second pronunciation form and the support b of the vowel TT. That is, a vowel sound is pronounced for a long time (double pronunciation), and a letter is pronounced under the influence of consonants and buzzing sounds. At this time, the TT is pronounced with emphasis on the first pronunciation since it is an A series vowel.

As another example, the pronunciation of the word 'Guam' is the first pronunciation of the vowel 과 and the consonant a combined with 모음, the vowel [. These are combined and the sound of [cancer] is followed. At this time, ㅘ is pronounced with emphasis on the second pronunciation because it is a W-series vowel.

When it is applied to speech recognition and speech conversion according to the classification system for Korean pronunciation, it greatly affects the recognition rate and speech conversion pronunciation of speech recognition. The present invention provides a Korean voice method classification system centering on the vowel pronunciation as described above, so that voice recognition can be performed at a higher recognition rate when the voice is recognized and converted into a character, and when the character is changed into a voice, Enable voice conversion.

The present invention can provide a speech pronunciation database classified by the above classification method, and can provide a speech recognition system and a speech conversion system using the speech pronunciation database.

In the case of a voice recognition system, for example, when a voice is input, the input voice is converted into a digital voice signal, a pronunciation feature is extracted from the converted digital voice signal, and then compared with a voice pronunciation database according to the present invention. It can be applied to speech recognition technology by determining the text.

More specifically, the speech recognition system 10 includes a speech input unit 11, an A / D converter 12, a feature extractor 13, a speech recognition database 14, and a discriminator 15. The speech recognition database 14 includes and stores a speech pronunciation database according to the present invention.

The voice input unit 11 receives a user's voice, for example, as a micro. The A / D converter 12 converts an analog voice signal input through the voice input unit 11 into a digital voice signal. The feature extractor 13 extracts pronunciation feature data from the converted speech signal. The determination unit 15 compares the pronunciation feature data extracted by the feature extraction unit 13 with the voice pronunciation database 14a stored in the voice recognition database 14 to determine which character corresponds.

At this time, the determination unit 15 may classify the digital voice signal according to the Korean voice pronunciation classification centered on the vowel pronunciation of the present invention and compare the database with the database to determine the corresponding character, and further, through the classification system according to the present invention. Excellent recognition rate can be provided.

In the case of a voice conversion system, for example, when a character is input, the voice corresponding to the input character may be extracted from the voice pronunciation database according to the present invention, thereby generating a voice for the corresponding character, and applied to a voice conversion technology.

More specifically, the voice conversion system 20 includes a text analysis unit 21, a voice conversion database 22, a voice synthesis unit 23, and a voice output unit 24. The voice conversion database 22 includes and stores a voice pronunciation database according to the present invention.

The voice conversion database 22 includes information defining voice data corresponding to each of the text data, and information defining feature data such as phonetic phonology, pronunciation, and stress on the voice data. The classification system according to

The text analysis unit 21 receives text data and performs morphological analysis / tagging, syntax analysis and semantic analysis, and syllable / phoneme conversion using the voice conversion database 22.

At this time, the voice synthesizer 23 receives the text feature data extracted from the text analyzer, generates a synthesized sound using voice data corresponding to the text data stored in the voice conversion database 22, and outputs the voice output unit. (24) For example, a voice may be output through a speaker, and a more natural Korean pronunciation may be output through a classification system according to the present invention.

Features, structures, effects, and the like illustrated in the above-described embodiments may be combined or modified with respect to other embodiments by those skilled in the art to which the embodiments belong. Therefore, it should be interpreted that the contents related to such a combination and modification are included in the scope of the present invention.

Claims (11)

In Korean,
Korean speech recognition system introduced pronunciation classification centered on vowel pronunciation, characterized in that one syllable is speech recognition centered on vowel.
The method of claim 1,
The pronunciation classification based on the vowel pronunciation includes a first classification method of classifying a vowel into three series vowels including a first series vowel, a second series vowel, and a third series vowel according to sound. Korean speech recognition system.
The method of claim 2,
The first series vowels include ㅏ, ㅓ, ㅗ, TT, ㅡ, ㅣ, ㅐ, ㅔ,
The second series vowels include ㅑ, ㅕ, ㅛ, ㅠ, ㅒ, ㅖ,
And the third series vowels include 한국어, ㅙ, ㅞ, ㅚ, ㅝ, ㅟ, ㅢ.
The method of claim 3,
The pronunciation classification based on the vowel pronunciation includes a second classification method for giving two pronunciation shapes including a first pronunciation shape and a second pronunciation shape for each vowel.
The method of claim 4, wherein
And the pronunciation shape is given by using the first series vowel and the second series vowel.
The method of claim 5,
The second classification method,
In the case of the first series vowel and the third series vowel, two phonetic shapes are given by using the first series vowel,
In the case of the second series vowel, the Korean speech recognition system is configured to give two pronunciation shapes using the first series vowel and the second series vowel.
The method of claim 5,
The second classification method,
In the case of the first series vowel, the same first series vowel is assigned to the first pronunciation shape and the second pronunciation shape.
In the case of the second series vowel, the second series vowel is assigned to the first pronunciation shape and the first series vowel is assigned to the second pronunciation shape.
In the case of the third series vowels, the Korean speech recognition system of claim 1, wherein the first series vowels are different from each other in the first pronunciation vowel and the second pronunciation vowel.
The method of claim 4, wherein
The pronunciation classification centering on the vowel pronunciation includes a third classification method for applying stress to one of the first and second pronunciation shapes given to each vowel. Voice recognition system.
The method of claim 8,
The third classification method,
In the case of the first series vowels and the second series vowels, stress is applied to the first pronunciation vowel among the two pronunciation vowels,
In the case of the third series vowels, the Korean voice recognition system of claim 1, wherein stress is applied to the second phonetic shape among the two phonetic shapes.
The method according to any one of claims 1 to 9,
The Korean speech recognition system is a Korean speech recognition system, characterized in that for converting the input voice into a corresponding character using a speech recognition database stored pronunciation classification centered on the vowel pronunciation.
The method of claim 10,
The Korean voice recognition system
A voice input unit for receiving a voice;
An A / D converter for converting a voice input through the voice input unit into a digital voice signal;
A feature extraction unit for extracting a pronunciation feature from the digital voice signal converted by the A / D converter; And
And a discriminating unit for determining which character corresponds to the pronunciation classification stored in the speech recognition database by comparing the pronunciation feature extracted from the feature extraction unit.