KR100363876B1 - A text to speech system using the characteristic vector of voice and the method thereof - Google Patents

Abstract

The present invention relates to a text-to-speech device and method using the feature vector of the voice.

Text-to-speech apparatus using the feature vector of the voice according to the present invention, the character database unit for storing the reading rule data, phoneme parameter data, character code data for the character string by a predetermined criterion; A character input unit which receives a character string by a predetermined input means; A phoneme parameter extracting unit for extracting phoneme data of characters input from a text input unit from the text database unit; A Gaussian noise generator for generating Gaussian noise as a fundamental frequency for the unvoiced sound when the extracted phoneme parameter is an unvoiced sound; An impulse train generator for generating an impulse train as a fundamental frequency for the voiced sound when the extracted phoneme parameter is a voiced sound; A global pulse generator for converting an impulse train generated by the impulse train generator into a fundamental frequency waveform of a speech signal from a human oral structure; A vocal track unit generating a speech signal for a phoneme by applying a phoneme parameter value extracted from the phoneme parameter extractor to a basic frequency value for an unvoiced or voiced sound generated by a Gaussian noise generator and a global pulse generator; And an articulator for connecting the voice signals of the phonemes generated by the vocal track unit to each other and synthesizing them into continuous voice signals.

The text-to-speech apparatus using the speech feature vector according to the present invention minimizes the database required for speech synthesis and synthesizes text into speech signals using only phoneme parameters, thereby shortening the processing time of speech synthesis in real time. By using only the effect is that the sound can be reproduced by the voice to be close to the natural sound.

Description

Text-to-speech system using the characteristic vector of voice and the method

The present invention relates to an apparatus for text-to-speech using a feature vector of speech and a method thereof. More specifically, a text-to-speech device using a feature vector of speech to extract a text code for a phoneme from a text and to synthesize and reproduce the text into a speech signal using a phoneme parameter as a speech feature vector to the extracted text code; It's about how.

Text-to-speech devices and methods are text-to-speech technology that converts sentences into speech using speech language, the most natural form of communication between humans and computers. Is an important technology that can be used across all areas of human use.

Referring to FIG. 1, a conventional apparatus for converting a speech character and a method thereof are as follows.

1 is a block diagram showing a schematic system configuration of a voice-to-text conversion system in Korean and a process of its operation.

As shown, the conventional text-to-speech conversion system includes a pre-processing unit 101 for inputting and processing a sentence, a dictionary type number / symbol / abbreviation DB 101 for extracting data required for the pre-processing process, and preprocessing. A morpheme analysis unit 103 for separating morphemes for characters, a morpheme DB 104 as a data dictionary for analyzing morphemes in the morpheme analysis unit 103, a parser unit 105 for syntax analysis, and a parsed syntax Character / phonic conversion unit 106 for converting letters into phonemes, exception phoneme DB 107 as a data dictionary for pronunciation rules for symbols or special characters during character / phony conversion, and conversion in character / phonic conversion unit 106 Rhymes for phonological rhymes. To set the duration of each speech data for the speech synthesis data generated by the speech synthesis data generation unit 108, the speech synthesis data generation unit 108 for generating the synthesis unit and the boundary information of letters, words, sentences Duration control unit 109, the fundamental frequency control unit 110 for generating and controlling a fundamental frequency for setting the fundamental frequency of the speech to be synthesized, by using the speech synthesis data, duration information, the fundamental frequency generated by the above-described components Synthesized sound generating unit 111 for synthesizing a voice signal, and synthesized unit DB 112 which stores a plurality of synthesized sounds as data so as to extract the synthesized units necessary for speech synthesis described above by the synthesized sound generating unit and use them for speech synthesis; It is configured by.

Exemplary embodiments of the related art implemented using some of these techniques are "Text / Speech Converter Using Multi-Step Input Information and Its Method," Patent Publication No. 10-0194814, "Digital Voice Reproduction of Patent Publication No. 2000-0024096. Device "," TTS system and TTS service method using the Internet "of Patent Publication No. 2000-0024318, and" Computer system having a function of outputting voice information when typing a keyboard "of JP 2000-0011449. .

However, the above-described conventional techniques can improve the sound quality because the information of the individual is processed in multiple stages, but the amount of data of the voice is increased, which causes transmission of the voice when parameterizing the feature of the voice in communication such as the Internet. May cause problems. In addition, in extracting a parameter for an individual's rhyme, there is a problem that it is difficult to extract and process the information that can make the most of the characteristics of the individual in real time.

In addition, when the hardware is configured using PCM, there is an inconvenience in that the compression rate is low, and the user has to purchase and install the hardware, and an error occurs in texting the scanned document by scanning and texting and playing the document. There is a problem that there is a possibility. In addition, the data capacity is increased by coding the characteristics of the voice in PCM, and the TTS system is created with a voice value matched with the keyboard value input at the time of keyboard input, but the voice reproducing method is insufficiently described by those skilled in the art. There is a problem.

In case of storing data in the form of WAV or MP3 when playing voice, the amount of data to be stored for playback is increased, so processing time such as streaming or buffering is required when transmitting and playing data in communication such as Internet. Real time processing is difficult. And even in the case of transmitting only the parameter of the character there is a problem that the sound quality is poor compared to the present invention to be described below.

Accordingly, the present invention is to solve the above problems, the phoneme parameter data table using the feature vector of the preconfigured voice to convert the text document to speech by converting the frequency characteristics of the natural language into phoneme units A voice that extracts only a phoneme parameter as a feature vector (phoneme parameter) for a phoneme corresponding to a character constituting a text document, is linearly connected in the time domain and synthesized and reproduced in the time domain nonlinearly in the frequency domain. An object of the present invention is to provide a text-to-speech device using the feature vector and a method thereof.

1 is a block diagram showing the overall processing of the speech synthesis method in the prior art;

2 is a block diagram showing a preferred embodiment of a text-to-speech device using a speech feature vector according to the present invention,

3 is a flow chart showing a preferred embodiment showing a process for text-to-speech using a speech feature vector according to the present invention.

4 is a subroutine diagram illustrating a speech synthesis and output process in the above-described process of FIG. 3.

5 is a diagram illustrating an embodiment in which a text-to-speech device using a feature vector of speech according to the present invention is applied in a network environment.

* Explanation of symbols for the main parts of the drawings *

200: text-to-speech device 201: text input unit

202: phoneme parameter extraction unit 203: character database unit

204: Gaussian noise generator 205: Impulse train generator

206: global pulse generating unit 207: amplifying unit

208: vocal track section 209: articulation section

In the present invention, the frequency, duration, start time, pitch period, etc., of the phoneme as a feature vector of the speech can be simplified to simplify the construction of a large database of characters, words, etc. in the analysis and speech synthesis of the speech. Use the phoneme parameter of.

In order to use a phoneme parameter as a feature vector of a voice, a phoneme parameter for a phoneme must be extracted and databased. This is an apparatus for extracting a voice signal feature by a nonlinear method and a patent application of Patent 2000-0056532 filed by the same applicant. The method can be achieved by extracting a voice signal into voiced, unvoiced, and silent phoneme units, extracting a phoneme parameter as a voice feature vector for each phoneme, and constructing a database using the extracted phoneme parameters as data.

Using the preconfigured character database including phoneme parameter data extracted by the aforementioned method of Patent Application 2000-0056532, the code of the input character is interpreted in phoneme units and the phoneme parameter is extracted. . The parameters are organized by phonemes, and these parameters are used to implement speech reproduction by speech synthesis online or on a local system.

Hereinafter, a preferred embodiment of a text-to-speech device and a method using a feature vector (phoneme parameter) of a voice according to the present invention will be described in detail with reference to the accompanying drawings.

2 is a view showing an embodiment of a text-to-speech device using a speech feature vector according to the present invention.

As shown, the text-to-speech apparatus according to the present invention includes a text input unit 201, a phoneme parameter extraction unit 202, a text database unit 203, a switching unit S, a Gaussian noise generator 204, and an impulse trap. A phosphor generator 205, a global pulse generator 206, first and second amplifiers 207 and 207 ', a vocal track unit 208, and an articulator 209, and a vocal track unit. The first data line 210 transmits the phoneme parameter data and the second data line 211 transmits the pitch period data to the impulse train generator 205.

Next, the components of the text-to-speech apparatus using the aforementioned speech feature vector will be described in detail.

The character database unit 203 described above is a database required for the phoneme parameter extraction unit 202 to extract phoneme parameters corresponding to characters. The character database unit 203 described above for extracting phoneme parameters has a read rule table, a character code table, and a phoneme parameter table therein.

The read rule table in the above-described character database unit 203 is used by the phoneme parameter extraction unit 202 to convert symbols, numbers, and special characters into phonetic character codes. Therefore, the above-described reading rule table has data on reading rules for symbols, numbers, special characters, and the like. When the reading rules are Korean, it is preferable that the reading rules be configured to follow the Korean reading rule. An example of such a read rule would be to map the special character "%" to the string "percent". In the case of a number, "1,234" corresponds to a string of "one thousand two hundred and thirty four", and "1, 2, 3, 4" corresponds to a string of "one, two, three, four", and a comma, a period, The data has a predetermined time value for the processing rule of the silent connection portion for spacing and the like. The voices of different voices in numbers are converted by the space code. In this way, symbols, numbers, and special characters among the characters input to the phoneme parameter extraction unit 202 are given a character code value corresponding to the phoneme parameter. In addition, the above-described read rule table serves to generate a synthesis unit in the course of articulation.

The character code table of the above-described character database unit 203 is composed of characters, symbols, numbers, and characters other than symbols, numbers, and special characters converted by the phoneme parameter extraction unit 202 described above by the reading rule table described above. Table to reference when extracting the character code for. For this purpose, the above-described character code table uses standard Hangul code values as its data.

The phoneme parameter table described above is a table that stores phoneme parameters as feature vectors of phonemes corresponding to character codes for input characters extracted with reference to the above-described reading rule table and the above-described character code table as its data. It is a data table referenced by section 202 to extract phoneme parameters for each character code.

The phoneme parameter table described above is non-linear in time and frequency domain in the speaker's voice after receiving the speaker's voice several times by the patent application 2000-0056532, "Speech signal feature extraction apparatus and method thereof," The data is extracted using the frequency of the phoneme, the duration of the frequency, the start time, and the pitch period.

After the data stored in the character database unit 203 is extracted corresponding to the phoneme of each character code by the phoneme parameter extractor 203, the Gaussian noise generator 204, the impulse train generator 205, and the glow are extracted. The de-pulse generator 206, the vocal track 208, and the articulator 209 are used to synthesize speech.

The phoneme parameter extraction unit S202 described above receives the processed character data from the above-described character input unit S202. The phoneme parameter extracting unit 202 receiving the input character data converts a symbol, a number, a special character, etc. into a character code in a text form according to a reading rule in the character data transmitted using the character database unit. Next, we extract the character codes of the converted symbols, numbers, and special characters, and the character codes for the input characters, and extract phoneme parameters as feature vectors of speech corresponding to each code.

The Gaussian noise generator 204 described above generates a predetermined signal waveform as a noise signal for generating an unvoiced voice signal. When a phoneme parameter (voice feature vector) value extracted by the phoneme parameter extractor is applied to the generated signal, a voice signal of a phoneme for the unvoiced sound is generated.

The impulse train generator 205 described above generates a fundamental frequency by using a pitch period for voiced sound. This coincides with the frequency of the human vocal cords and statistically calculates the pitch period input from a plurality of speakers to be the basic pitch period value.

The above-described global pulse generator 206 distorts the impulse train train generated by the impulse train generator 205 into a fundamental frequency waveform for the voice generated in the oral structure of the human being to the fundamental frequency of the human voice. Change the impulse train train to be similar to the waveform produced by it.

The first and second amplifiers 207 and 207 'described above act on the generated Gaussian noise and the glacial pulse, respectively, and amplify them to a signal level suitable for processing in a text-to-speech apparatus using the speech feature vector of the present invention. Perform.

The vocal track unit 208 described above converts the Gaussian noise for the unvoiced sound and the glacial pulse for the voiced sound into a voice signal of a phoneme corresponding to the phoneme parameter extracted by the phoneme parameter extractor 202. That is to say, by applying the frequency characteristic, frequency duration, and frequency start time of the phoneme parameter value to unvoiced sound to the Gaussian noise as a noise, it converts to the voice signal waveform for the phoneme. By applying the phoneme parameter value of the voiced sound described above to generate a voice signal for the voiced phoneme.

The above-described articulator 209 connects the audio signals of the phonemes generated by the vocal track unit 208 to each other to generate an actual continuous voice waveform. At this time, the phoneme parameter values for the silence extracted by the phoneme parameter extractor 202 are used for sentence concatenation such as phoneme concatenation, word concatenation, comma, and spacing. In addition, the above-described read rule table of the character database unit and the speech synthesis unit data are used in the articulation unit for speech synthesis, and these data are extracted in the phoneme parameter extraction process, transmitted through the data line, and stored in a buffer as a predetermined storage means. It is stored and used when needed.

Next, an operation process of the apparatus for text-to-speech using the aforementioned speech feature vector will be described in detail with reference to the above-described elements.

First, when a predetermined string is input, the above-described text input unit 201 receives text information in phoneme units. Next, the phoneme parameter extraction unit 202 described above converts the symbols, numbers, and special characters from the input character string into the character string according to the reading rule by referring to the character database unit 203 described above. Strings that follow the converted reading rules are classified into primary, neutral, and final characters as the character components that make up each character. Each classified phoneme is extracted and converted into a character code corresponding to the phoneme using a character code table. Next, a phoneme parameter for each character string is extracted by referring to a phoneme parameter table as a phoneme parameter corresponding to a character code required in a vocal track unit and an articulator unit, which are the following processes.

The extracted phoneme parameter is composed of a parameter representing the frequency characteristic of the phoneme, a parameter representing the speech time of the phoneme, and a parameter representing the duration of the phoneme. The frequency characteristics of the phoneme described above have three basic parameters, each subdivided according to unvoiced, voiced and silent. Therefore, the extracted phoneme parameters have five parameter values, including parameters for unvoiced voice, voiced sound, and silence as frequency characteristics, parameters for duration of phonetic speech and voice start time. Here, the parameter as the frequency characteristic is a reference value for distinguishing whether the character is an unvoiced sound or a voiced sound, or a silent sound as a connection part such as a phoneme and a phoneme, a word and a word, a comma and a period.

The processing sequence of the character string determines whether unvoiced, voiced or unvoiced in the order of input phonemes using frequency characteristic data of the phoneme parameter extracted according to the input sequence of the character string.

When the input phoneme parameter value has a characteristic frequency of unvoiced sound, the Gaussian noise generator 204 is driven to generate Gaussian noise as a fundamental frequency for unvoiced sound processing. Next, when the extracted phoneme parameter is determined to be a voiced sound, the impulse train generator 205 is driven to generate a predetermined impulse train that matches the pitch period value of the voiced sound transmitted from the phoneme parameter extractor 202. Since the generated impulse trains are waveforms of periodic repetitive pulses having the pitch period value for the fundamental frequency in the statistically extracted speech signal as the period, they should be converted to have the waveform of the frequency for the human oral structure. This conversion process is performed in the global pulse generator 206. The global pulse generated by converting the impulse train in the global pulse generator 206 has a waveform of a fundamental frequency with respect to the voiced sound coming from the human oral structure. If the phoneme of the input character is silent, it is determined whether the silence is silence between phonemes, silence between letters, silence between words, and silence by commas or periods. If it is determined that there is a silence between phonemes constituting a character, the phoneme parameter extracting unit 202 includes data values of phoneme parameters for linearly connecting the phonemes before and after the silence in a frequency domain nonlinearly in a frequency domain for a predetermined time. Is transmitted to the articulation unit 209 to be used for concatenating the phonemes in the case where subsequent phonemes are connected to generate a continuous voice signal.

Next, after the unvoiced and voiced voice signals generated by the Gaussian noise generator 204 and the global pulse generator 206 are amplified to a predetermined signal level by the first and second amplifiers 207 and 207 '. When characters are sequentially input to the vocal track unit 208, the phoneme parameter extracting unit 202 converts the voice signals for the corresponding phonemes according to phoneme parameter values corresponding to the unvoiced and voiced sounds. . The voice signal for each phoneme is transmitted to the articulator 209 together with the connection data for each connection by silence, and by the read rule configured in the character database part in the articulator and the silence as the connection part of each phoneme The connection is synthesized, converted into a continuous voice signal, and output as voice by a predetermined output means.

The processing in the articulator 209 will be described in more detail as follows.

The parameters of the characters inputted into primary, neutral, and final characters are phonological phenomena such as soft, consonant, consonant, retrograde, and forward motion during speech reproduction. As a result of the difference in the connection time of, the speech signal for the phoneme parameter for the text can be solved by reproducing the voice as input by adjusting the duration of each time and the connection time between the phoneme and the phoneme. For example, if the sound is composed of a flat sound, a hard sound, and a rhythm such as “ㄷ” and “ㄸ”, it is made by adjusting the duration of each phoneme. The length of the phoneme is longer than the phoneme in the back, and the phoneme of the phoneme is synthesized close to the first character of the next letter when articulation is made so that the synthesized sound close to natural language is output.

In the case of soft sound, for example, "flying", "b ㅏ ㄹ ㅏ", such as the sound is made naturally by reproducing the articulation according to the input sequence, so no separate consideration is required.

In other words, the phonological phenomenon that occurs when we speak is generated by the duration and the start time of the phoneme. Therefore, it is a method of reproducing a voice that conforms to the reading rule by reproducing a voice by tuning input parameters in order. Therefore, when a character is reproduced as a voice, the interval between the phonemes, the words, the sentences, and the like between the phonemes can be changed differently, so that the sound can be reproduced to be close to the natural language.

This process is processed in the order of the characters to be input, and processing of unvoiced and voiced sound must also be switched by the switching unit S in order to be processed in the order of input of the phonemes.

The above-mentioned text-to-speech apparatus uses only phoneme parameters when reproducing text by voice. When playing or synthesizing a voice, voiced sound uses an impulse train and voiced sound uses Gaussian noise (noise). This is for reproducing or synthesizing the voice by inputting to the vocal track generator 208. At this time, in order to bring the voice closer to the natural language, Gaussian noise or impulse train, which is the original signal of voiced sound and unvoiced sound, is selected by switching, and the switching time or duration is set by the silent duration data among preset voice characteristic data. That is another reason. Therefore, using the text-to-speech device using the phoneme parameter, it is possible to find the phoneme to be reproduced in the established database, or to prevent the loss of voice with fast transmission and low bandwidth occupancy when transmitting through a communication network such as the Internet. There is an advantage, and can provide a voice playback service in real time.

Figure 3 is a flow chart showing a preferred embodiment showing a process for the text-to-speech method using a speech feature vector (phoneme parameter) according to the present invention.

The process of the text-to-speech method according to the present invention first registers the feature data of the voice to be output. The process of registering the feature data of the voice is in accordance with the previously filed Patent No. 10-2000- 0056532 as described above (S301). Next, a character to be converted into a voice is input from a keyboard or predetermined text editing software. The received character is divided into a string of a predetermined length for data processing. The divided string is converted into a general string, that is, a Hangul string according to the Hangul reading rule in the case of Hangul, according to a reading rule. When a symbol / number / special character is converted into a Hangul character string, the input character string is extracted with reference to a character code table composed of Korean standard character codes (S304). When the character code for the character is extracted, the phoneme parameter value for the character code is extracted by referring to the phoneme parameter table in which the preconfigured character code and the phoneme parameter are stored in association (S305). From the extracted phoneme parameter value, unvoiced sound, voiced sound, and unvoiced sound are nonlinearly separated in the frequency domain according to the zero crossing ratio of frequency among the phoneme parameters, and then voice signals for each unvoiced sound, voiced sound, and unvoiced sound are generated and output. S306). Next, it is determined whether the character string to be converted to voice is continuously input (S307). If the character string is continuously input as a result of the determination, the operation is repeated from the above-described step S303, and the process is terminated if there is no input character string.

4 is a subroutine diagram illustrating a speech synthesis and output process of step S306 in the above-described process of FIG. 3.

As shown, the speech synthesis and output process receives the phoneme parameters extracted in step S305 (S461) and then classifies them as unvoiced, voiced and silent (S462) by the zero crossing ratio in the frequency characteristics of the respective phoneme parameters. Each parameter is switched in the order in which it is entered to perform the next process. First, when the above-mentioned S462 input voice is a voiced sound, an impulse train corresponding to the pitch period is generated using the pitch period among the phoneme parameter values (S463). Since the generated impulse train is a continuous pulse train having a predetermined period corresponding to the frequency of the human vocal cords, which is statistically extracted, it is converted into a glutal pulse train having a characteristic of a frequency output from the human vocal cords (S464). Next, when the phoneme parameter value input in step S462 described above is classified as an unvoiced sound, Gaussian noise as a fundamental frequency for the unvoiced sound is generated (S465). The generated glotal pulse and voiced phoneme parameter values and Gaussian noise and unvoiced phoneme parameter values are converted to the frequency waveform of the voice signal corresponding to the phoneme by the phoneme parameter. The audio signal is converted (S466). Next, the generated Gaussian noise, the phoneme parameter value for the input unvoiced voice, the generated glotal pulse and the phoneme parameter value of the input voiced sound are transmitted to the vocal track part. The Gaussian noise transmitted to the vocal track part is transmitted to the vocal track part. The voice parameters of the unvoiced sound, that is, the frequency characteristics, the frequency duration, and the start time of the unvoiced sound, are converted into voice signals for the phonemes corresponding to the unvoiced sound. The audio signal is converted into a phoneme (S466). In operation S402, the signals classified as silent are converted into different connection signals according to phonemes and phonemes, words and words, sentences connected, commas, and periods. The connection between the phoneme and the phoneme removes the disconnection between the phonemes by non-linearly connecting the end of the frequency of the phoneme before the position where the silence is to be inserted and the frequency of the phoneme located behind it for a predetermined time interval. A silent signal is generated and converted into connection setting data so as to connect words and words, between sentences and sentences, commas, periods, and the like at different time intervals by setting pauses (S467). The voice signal for the phoneme corresponding to the voiced sound and the unvoiced sound is transmitted to the tuning unit 209 together with the connection setting data of the silent signal and synthesized into the voice signal. The synthesis process is the same as the connection method in the silent data extraction process of step S467 (S468). Next, after reproducing and outputting the voice signal synthesized by the predetermined voice output means as voice (S469), the process returns to the process of FIG. 3 to continue the next process.

5 is a diagram showing an application example using a text-to-speech device using a speech feature vector according to the present invention.

The illustrated embodiment is a block diagram of a network environment for providing a voice output service through a communication network such as the Internet, and the processing thereof is as follows.

When the user needs to log in using the terminal 500 configured in the user environment, the user accesses through authentication information such as voice and text (S501 and S502). The connected user selects a file such as a voice output service, that is, a book, a magazine, or a newspaper, through a search engine configured in the server 510 (S503). The server 510 receiving the user's service request transmits phoneme parameters including frequency information, start time information, and duration information of phonemes constituting each text among the text information of the corresponding file to the terminal 500 configured in the user environment. send. The phoneme parameter transmitted from the server 510 is converted into voice by the text-to-speech device using the voice feature vector according to the present invention in the terminal 500 configured in the user environment, thereby providing voice services for various documents online. It becomes possible.

Next, a patent " apparatus for extracting features of a speech signal using a nonlinear method and a method thereof " of the previously applied application No. 2000-0056532 used in the present invention will be schematically explained.

The wavelets used in the use of the dynamic wavelet transform described in the patent application No. 2000-0056532 are applied to a haar wavelet for nonlinear processing of voiced sound and a spline for nonlinear processing of unvoiced sound. Spline Wavelet is used to classify human voice into voiced, unvoiced and silent, and then the frequency characteristics, start time, and duration of each phoneme are voice characteristic data.

The present invention is not limited to the above-described embodiments of the present invention, and various changes can be made without departing from the spirit of the present invention.

The apparatus for text-to-speech using the speech feature vector according to the present invention has the effect of minimizing the database required for speech synthesis and shortening the processing time of speech synthesis in real time by synthesizing text into speech signals using only phoneme parameters. .

In addition, by synthesizing and reproducing text into voice using only phoneme parameters, there is an effect of making the reproduced voice close to natural sound.

In addition, when the text-to-speech device and the method according to the present invention are applied on a network, the network traffic required can be significantly reduced by transmitting only the voice parameters necessary for voice synthesis, and streaming is performed on the terminal side even when playing with sound reproduction. In addition, there is an effect of allowing sound reproduction to be performed in real time without buffering.

Claims (9)

A character database unit for storing read rule data, phoneme parameter data, and character code data for a character string based on a predetermined criterion; A character input unit which receives a character string by a predetermined input means; A phoneme parameter extraction unit for extracting phoneme data of a character input from the character input unit from the character database unit; A Gaussian noise generating unit generating Gaussian noise as a fundamental frequency for the unvoiced sound when the extracted phoneme parameter is an unvoiced sound; An impulse train generator for generating an impulse string as a fundamental frequency for the voiced sound when the extracted phoneme parameter is a voiced sound; A global pulse generator for converting an impulse train generated by the impulse train generator into a fundamental frequency waveform of a voice signal from a human oral structure; A vocal track unit for generating a voice signal for a phoneme by applying a phoneme parameter value extracted from the phoneme parameter extractor to a basic frequency value for the unvoiced or voiced sound generated by the Gaussian noise generator and the global pulse generator; And And an articulator for connecting the voice signals of the phonemes generated by the vocal track unit to each other and synthesizing the voice signals into continuous voice signals. The method of claim 1, wherein the character database unit, A read rule table having read rule data for symbols, numbers and special characters; A character code table having data of initial, neutral, and final character codes for the input characters; And And a phoneme parameter table having, as data, a phoneme parameter value for a phoneme corresponding to each character code value configured in the text code table. The phoneme parameter table of claim 2, Text-to-speech device using a speech feature vector comprising a basic frequency value for each phoneme, a duration value of the phoneme and a start time value of the phoneme. The method of claim 3, wherein the fundamental frequency value, An apparatus for text-to-speech using a speech feature vector, wherein each phoneme has a frequency value for any one of unvoiced, voiced or unvoiced. The method of claim 4, wherein the phoneme parameter value for the silence, When the phoneme and the phoneme are connected in the speech synthesis process, the frequency of the unvoiced sound located in front of the phoneme where the silence is located and the frequency of the phoneme located behind the silence are linearly connected in the time domain in a nonlinear manner in the frequency domain. A text-to-speech device using a feature vector of speech, characterized in that the speech is synthesized. The method of claim 4, wherein the phoneme parameter value for the silence, In the case of a silence corresponding to any one of letters, letters, words, words, sentences, sentences, commas, or periods, the voice feature is characterized in that the voice data is processed as an idle time for each corresponding predetermined time interval. Character-to-speech apparatus using a feature vector of a speech using a vector. The method of claim 1, wherein the articulation portion, Phonetic character conversion device using a speech feature vector, characterized in that to implement the phonological phenomenon in the articulation process using only a phoneme parameter value. A first step of receiving a phoneme parameter data, read rule data, and character code data of a voice to construct a character database; A second step of receiving a string to be synthesized as a voice; A third step of converting symbols, numbers, and special characters from the input string into strings according to reading rules, classifying the converted strings into initial, neutral, and final strings according to reading rules, and extracting a character code for the corresponding characters; Extracting a phoneme parameter corresponding to the extracted character code; And And a fifth step of synthesizing and reproducing the speech using the extracted phoneme parameters. The method of claim 8, wherein the fifth step, A sixth step of receiving the extracted phoneme parameter; A seventh step of determining whether the input phoneme parameter is for unvoiced sound, voiced sound, or silent sound; If it is determined that the sound is unvoiced in the seventh step, Gaussian noise is generated, and if it is determined that the sound is sound, the impulse train is generated and converted into a global pulse train. An eighth step of extracting and transmitting a phoneme parameter to a connection portion of the corresponding silence when it is determined that the sound is silent; A ninth step of generating a voice signal corresponding to the phoneme in the order of the input phonemes by receiving a result signal for each of the voiced sound, the unvoiced sound, or the silent sound processed by the eighth step and the corresponding phoneme parameters; And And a tenth step of synthesizing each of the voice signals generated in the ninth step sequentially using the silent data, synthesizing the voice signals into continuous voice signals, and outputting the voices by a predetermined voice output means. Text-to-speech method using feature vector of speech.