KR0134707B1 - Voice synthesizer - Google Patents

Abstract

The present invention relates to a method of synthesizing a Korean character into a voice, and is a method of subdividing a voice into a diphone unit to obtain a parameter from a synthesis database and converting the voice into an LSP synthesis method. As a feature of the present invention, the classification and binding in diphonic units are effectively subdivided, and when constructing a synthetic database, a special landmark is given for each diphony. Each landmark city has different meanings according to the type of diphony, and it is configured to facilitate the adjustment of length and combination when combining synthetic units. In addition, the key to speech synthesis technology is the quality of synthesized sound. Therefore, the LF model and the residual signal modified as the sound source of the LSP synthesis filter are used to improve the naturalness and clarity.

The present invention has the effect of increasing the new information providing technology by the proliferation of information retrieval using a synthetic technology when connected to the increasing number of information and communication services.

Description

LSP Speech Synthesis Method Using Diphone Unit

1 is a basic configuration diagram of hardware to which the present invention is applied;

2 is a schematic diagram of a synthetic database used in the present invention,

Figure 3 is a combination display of the type of diphones used in the present invention,

4 is an overall processing flowchart of the speech synthesis method according to the present invention;

5 is a block diagram of the sound source model used in the present invention.

* Explanation of symbols for the main parts of the drawings

1: Character input device 2: Central processing module

3: Synthetic database 4: D / A converter

10: structure of frame unit 20: structure of parameter for each frame

30: Boundary Display

The present invention relates to a speech synthesis method of the LSP method using a synthesis parameter database composed of diphone units, and to a speech synthesis method applicable to all fields requiring a function of converting Korean characters into speech.

Voice has been used most since ancient times as a means of human communication. The delivery of information using voice does not require special education and can be achieved while doing other work. As a result of the development of computer and communication technology, it has been in the spotlight as a new information transmission medium. In addition, as the spread of computers increases, users are increasingly using information and communication service networks to exchange information with each other and search for necessary materials. Accordingly, there is a continuous need for text / voice conversion technology, a media conversion service of information. The trend is expanding.

Such speech synthesis technology can be applied to various information providing services using telephone lines, speech devices for the language-disabled, human-machine interfaces such as OA devices and home appliances. However, there are still many difficulties in producing natural synthetic sounds that are close to humans, so the products that are commercially available in Korea are considered to be in the hands. In particular, various synthesis units have been studied to produce voices close to humans using the rule synthesis method. These units are mainly composed of phonemes or several consecutive phonemes.

In general, the larger the size of the synthesis unit, the simpler the connection rule between the synthesis units and the better the sound of the synthesized sound, but the larger the number of synthesis units required to produce an unlimited number of phonemes. In addition, since the characteristics of these synthesis units are affected by phonemes, accents, sentence forms, or contexts, it is very difficult to determine not only the most efficient and appropriate synthesis units but also the number of necessary synthesis units.

Among the synthesized units under study, the diphone unit includes transition periods between phonemes in the synthesized unit, and thus, even though the combining rules are complicated, the connection between the phonemes is natural.

Therefore, the present invention implements a synthesis technique using a diphone unit, and synthesizes an unvoiced sound using a residual signal in terms of intelligibility, thereby increasing the size of the synthesized database, but more accurately expressing the unvoiced sound. The purpose is to provide a synthetic method.

In order to achieve the above object, the present invention, the character input means for receiving and transmitting the Korean characters that can be expressed in a complete form, the central processing means for receiving a character input from the character input means for performing a speech synthesis algorithm, It stores the parameters configured in units of diphony used in the synthesis algorithm, and synthesizes a database for transmitting the necessary parameters to the central processing means, and converts the digital data synthesized in the central processing means to analog to synthesize the sound In a speech synthesis method applied to a device including a digital / analog converter for outputting, alphabets, numbers, and limitations after the Hangul processing process of converting a completed character input through a character input means into an internal code of 3 bytes using a conversion table. The first step of performing the abbreviation process, A second step of generating rhyme control information through a system analysis and a breathing process and generating a phonetic phonetic code sequence by applying a phonological rule of Korean through a pronunciation rule processing process; The third step of adjusting the length of the phoneme using the symbol string and converting each syllable in the form of 3 bytes into the diphone type defined in the synthesis database according to the result of the third step A fourth step (160) and a fifth step of bringing parameters from the synthesis database using the generated indexes of the diphones, and performing energy regulation using linear interpolation and energy increment of the LSP parameters between adjacent units; And a sixth step of performing rhythm control to determine, by using economic analysis information, a fundamental frequency representing information on a sentence structure, meaning, emotion, and the like; And a seventh step of synthesizing the voice using the synthesis parameter generated through the sixth step.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention;

1 is a hardware structural diagram to which the synthesis technique of the present invention is applied, and as shown in the drawing, a character input device 1, a central processing module 2, a synthesis database 3, and digital / analog (D / A). ) With a transducer 4.

The character input device 1 receives the Korean characters that can be represented in the KS5601 complete type and transmits them to the central processing module 2. The central processing module 2 is a part where the software for executing the algorithm of the present invention is loaded and executed. The synthesizing database 3 is a parameter database composed of units of diphony used in the synthesizing algorithm, recorded in the storage device, and responsible for transmitting necessary parameters to the central processing module 2. The D / A converter 4 is a part that converts the synthesized digital data into analog and plays the synthesized sound to the outside.

Table 1 is a table showing the classification according to the type of diphones used in the present invention.

In order to synthesize the voice, each syllable is divided into the form of a die phone as shown in the table by referring to the preceding syllables and the following syllables. Each diphony generated is used to fetch parameters from the synthetic database, and is also used to combine diphony units.

FIG. 2 is a diagram showing a database structure of a diphone unit composed of synthetic parameters, and each diphone is a portion representing a composition parameter part composed of frame units and section information for length control and connection between adjacent diphones. Consists of. One die phone consists of a series of frame shapes 10 having a constant length and each frame is divided into four kinds of composition parameters 20.

The pitch indicates whether the current frame is unvoiced or voiced. If this value is 0, it indicates that the current frame is unvoiced, so that the speech is synthesized using a residual signal during synthesis. Since it is a voiced sound, it indicates to synthesize it with a modified LF model. The energy boost is a value that determines the amplitude of the synthesized sound and is used as the gain of the LSP synthesis filter. There are 12 LSP parameters, P and Q, which are used as the coefficients of the synthesis filter represented by the 12th-order all-poll model.

A residual signal is a parameter used as a sound source when synthesizing a consonant, and can produce a synthetic consonant that is almost similar to the original sound. The part 30 drawn on the top part connects the die-phones 10 of the frame unit to synthesize a continuous voice, so that a special landmark time, i.e., the start of the phone, the boundary of the phoneme, the connection point, the end, etc. Defines up to five, depending on the type of diphony. Detailed landmark time by type of diphony is as follows.

The number circled below corresponds to the number in landmark hour 30 of FIG. 2 and does not indicate landmark time not in use.

-Diphon type 1)

① Location where the transition period from silence to vowel begins

② The position where the change from silence to vowel ends and stable vowel begins

③ The end of the stable section of the vowel before the change from collection to silence

④ Location where the change section from vowels to mute ends

-Diphon type 2)

① Location where the transition period from silent to half vowel begins

② Position where stable vowel starts after change from half vowel to vowel

③ The end of the stable section of the vowel before the change from collection to silence

④ Location where the change section from vowels to mute ends

-Diphon type 3)

① Location where the transition period from silence to initial consonants begins

② The position where the change from the silent to the vowel ends and the stable vowel starts.

③ The end of the stable section of the vowel before the change from collection to silence

④ Location where the change section from vowels to mute ends

⑤ The location where the unvoiced sound ends when the initial consonant sound is unvoiced

-Diphon type 4)

① Location where the transition period from silence to initial consonants begins

② The position where stable vowels begin after the change from silence to consonants to half vowels ends.

③ The end of the stable section of the vowel before the change from collection to silence

④ Location where the change section from vowels to mute ends

⑤ The location where the unvoiced sound ends when the initial consonant sound is unvoiced

-Diphon type 5)

① Location where the transition period from silence to vowel begins

② The position where the change from silence to vowel ends and the stable vowel starts.

③ The end of the stable section of the vowel before the change from the vowel to the final consonant begins

④ The location where the transition period from the vowels to the final consonants ends.

-Diphone type 6)

① Position where the stability section of the front bar starts

② The stability section of the front vowel ends before the transition period from the front vowel to the rear vowel begins.

I position

③ The position where the change section from the front vowel to the back vowel ends and the stable section of the back vowel begins.

④ Position where the stability section of the rear vowel ends

-Dipon type 7)

① Position where the stable section of the collection starts

② The end of the stable section of the vowel before the change from vowels to half vowels

③ Location where the change interval from vowels to half vowels ends

-Dipon type 8)

① Position where the stable section of the collection starts

② The end of the stable section of the vowel before the change from vowels to consonants begins

③ Location where the change interval from the vowels to the consonants ends

-Dipon type 9)

① Position where the change section from the half vowel to the vowel begins

② Position where the vowel stability section starts after the change period from the half vowel to the vowel ends

③ Position where the stable section of the collection ends

-Dipon type 10)

① Location where consonants begin

② The position where the vowel stable section starts after the transition period from consonant to vowel

③ Position where the stable section of the collection ends

④ Location where the change section from vowels to mute ends

⑤ Location where the unvoiced sound ends

-Dipon type 11)

① Location where consonants begin

② The position where the vowel stability section starts after the change period from consonant to half vowel to half vowel to vowel ends

③ Position where the stable section of the collection ends

④ Location where the change section from vowels to mute ends

⑤ Location where the unvoiced sound ends when the consonant sound is unvoiced

-Diphon type 12)

① Position where the stable section of the collection starts

② The end of the stable section of the vowel before the change from the vowel to the final consonant

③ Position where the change section from the vowel to the final consonant ends and the stable section of the final consonant begins

④ Position where the stable section of the final consonant ends

-Diphon type 13)

① Position where the stable section of the final consonant begins

② The position where the stable section of the final consonant ends before the change of the initial consonant from the final consonant

-Diphon type 14)

① Position where the stable section of the collection starts

② The end of the stable section of the vowel before the change from the vowel to / ㄹ /

③ Position where the stable section of / ㄹ / begins after the change section from the collection to / ㄹ / ends.

④ / ㄹ / position where the stable section ends

-Diphon type 15)

Location where the stable section of ① / ㄹ / starts

② The position where the stable section of / ㄹ / ends before the transition period from / ㄹ / to vowel begins

The position where the stability section of the vowel starts after the change section from ③ / ㄹ / to the vowel

④ Position where the stable section of the collection ends

FIG. 3 shows possible coupling states for each type of diphone used in the present invention, and has a great influence on the complexity of the coupling rule and the naturalness of the synthesized sound depending on how the combined units are combined with each other. In the drawings, the black part indicates that the bond is possible, and the letter L indicates the compound unit on the left when combining two diphons, and the unit R on the right.

4 is a flowchart illustrating an overall process of a text / voice conversion method according to the present invention, which will be described in detail with reference to the accompanying drawings.

The process of converting a text into a voice is largely composed of a language processing process 200 and a speech synthesis process 300 for synthesizing an actual voice.

The input complete character is passed through the Hangul process 110, which converts it into a 3-byte internal code using a conversion table, and then passes through the alphabet, numbers, and limited abbreviation process 120. Next, rhyme control information is generated in the boundary analysis and breathing process 130. In the pronunciation rule processing 140, which is the end of the language processing process, a phonetic rule sequence of a phonetic form is generated by applying Korean phonological rules.

In the speech synthesis process, the phoneme length control 150 is first performed using this symbol string. The length control is syllable based on the boundary analysis information of words, phrases, phrases, and sentences, and the minimum duration and intrinsic duration of the phoneme. Determined in phoneme order. Next, each syllable composed of three bytes is converted into a diphone type defined in the synthesis database (160). Then, using the generated indexes of the diphony (170) from the synthesized database from the index, and performs the energy control using linear interpolation and energy additive value of the LSP parameter between adjacent units. The rhyme control unit 180 determines the fundamental frequency expressing information on the structure, meaning, emotion, etc. of the sentence by using boundary analysis information. In the present invention, the second function P (t) is implemented.

P (t) = Pb- (Pb-Pa) * ((Tb-t) / (Tb-Ta)) ** 2, Ta = tTb

P (t) = Pc, Tb = tTc

P (t) = Pb- (Pb-Pd) * ((t-Tc) / (Td-Tc)) ** 2, Tc = tTd

At this time, Pa, Pb, Pc, and Pd are constants.

The LSP synthesis 190 at the bottom of the figure is a process of synthesizing speech using the synthesis parameters generated through the above processes 110-180 and is represented by a 12th order all-poll LSP filter and digital. The synthesized synthesized sound is synthesized using a D / A converter 4 that converts the synthesized sound into an analog.

FIG. 5 is a diagram of a sound source used in the LSP synthesizer 190 of FIG. 6, which shows unvoiced sound when this value is 0 by using the pitch information of the generated synthesis parameter. Use it to generate synthesized sound with intelligibility close to the original sound. If there is pitch information, the frame to be synthesized is voiced sound, so the LF model similar to the human vocal model was modified to fit Korean and used as a sound source.

Therefore, the present invention configured and carried out as described above can obtain the following special effects.

First, all Korean characters consisting of complete codes can be synthesized.

Second, up to five landmarks were made according to the characteristics of the diphony units to facilitate the connection of the synthesized units and the adjustment of the phoneme length.

Third, a more natural and clear synthesized sound can be obtained by using the modified LF model and the residual signal in the sound source model.

Fourth, the user can listen to the desired information by voice so that they can do other tasks at the same time.

Fifth, it is possible to synthesize text information by connecting to the information communication network currently in service.

Sixth, it can be used in the synthesis of welfare services and automatic interpretation technology for the disabled.

Claims (6)

Character input means (1) for receiving and transmitting Korean characters that can be expressed in a complete form, and central processing means (2) for receiving a character input from the character input means (1) and performing a speech synthesis algorithm; A synthetic database (3) for storing the parameters configured in units of the diphony used for the algorithm, and transmitting the necessary parameters to the central processing means (2), and digital data synthesized by the central processing means (2). In the speech synthesis method applied to a device including a digital-to-analog converter (4) for converting to analog and outputting synthesized sound to the outside, A first step of performing alphabetic, numeric and limited abbreviation processing after the Hangul processing process of converting a completed character input through the character input means 1 into an internal code of 3 bytes using a conversion table; A second step of generating rhyme control information through boundary analysis and breathing process and generating a phonetic phonetic code sequence by applying phonological rules of Korean through pronunciation rule processing; A third step of first adjusting the phoneme length by using the symbol string generated in the second step; A fourth step of converting each syllable having a three-byte form into a diphone type defined in a synthesis database according to a result of performing the third step; A fifth step of bringing the parameters from the synthesis database using the generated indexes of the diphones and performing energy adjustment using linear interpolation and energy weights of the LSP parameters between adjacent units; A sixth step of performing rhythm control to determine a fundamental frequency representing information on a sentence structure, meaning, emotion, etc. using boundary analysis information; And a seventh step of synthesizing the voice using the synthesis parameter generated through the sixth step. The method of claim 1, wherein the length adjustment in the third step is based on the boundary analysis information of words, phrases, clauses, and sentences, and the minimum duration and intrinsic duration of the phoneme for each phonetic phonetic sequence received. To determine the order of syllables and phonemes. According to claim 1, wherein the rhyme control in the sixth step, P (t) = Pb- (Pb-Pa) * ((Tb-t) / (Tb-Ta)) ** 2, Ta = tTb P (t) = Pc, Tb = tTc P (t) = Pb- (Pb-Pd) * ((t-Tc) / (Td-Tc)) ** 2, Tc = tTd And Pa, Pb, Pc, and Pd are implemented using a quadratic function P (t) which is a constant. 2. The speech synthesis in the seventh step is performed by using a twelfth all-poll LSP filter and a D / A converter 4 that converts synthesized speech expressed in digital into analog. Speech synthesis method characterized in that. The speech synthesis method of claim 1, wherein the parameter from the synthesis data base comprises a pitch, an energy weight, an LSP parameter, and a residual signal. 5. The speech synthesis method according to claim 4, wherein the speech synthesis in the seventh step is performed using an LF model modified as a sound source and a residual signal.