KR100427243B1 - Method and apparatus for analysing a pitch, method and system for discriminating a corporal punishment, and computer readable medium storing a program thereof - Google Patents

Abstract

Disclosed are a pitch analysis method and apparatus for analyzing a pitch of speech, a filamentous discriminant method and system, and a computer-readable recording medium incorporating a filamentous discriminant program. The analog-to-digital converter converts the voice signal into a digital voice signal, the pure voice signal extractor performs low pass filtering on the digitally converted voice signal, and the time domain analyzer extracts a signal above a predetermined threshold to analyze the waveform of the pure voice signal. To output the time domain analysis information. The analysis unit outputs the first analysis information through the pitch analysis on the time domain analysis information, outputs the second analysis information through the formant analysis, outputs the third analysis information through the frequency domain analysis, and displays and features grams. The factor extractor displays the first to third analysis information, extracts the feature factor, and the feature factor fusion unit fuses the extracted feature factor and outputs it, and the filamentous constitution analyzer checks the output feature factor and corresponds to the corresponding speech signal. Output Sasang Constitution Analysis Information. Accordingly, the acoustic feature of the voice input through the microphone can be extracted, and the ideological constitution of the main character of the voice can be discriminated based on the extracted feature factor.

Description

TECHNICAL AND APPARATUS FOR ANALYSING A PITCH, METHOD AND SYSTEM FOR DISCRIMINATING A CORPORAL PUNISHMENT, AND COMPUTER READABLE MEDIUM STORING A PROGRAM THEREOF}

The present invention relates to a pitch analysis and filamentous discriminant method, and more particularly, a pitch analysis method and apparatus for analyzing a pitch of speech, a filamentous discriminant method and system, and a computer incorporating a filamentous discriminant program. Relates to a readable recording medium.

In general, in digital voice signal processing technologies such as speech recognition, synthesis, and analysis, it is very important to accurately detect a pitch frequency, that is, a pitch, which is a basic technology. The fundamental frequency is very difficult to detect in a sound transitioned to a sound transition section or a noise because the sound change is severe and it is difficult to set a threshold value for each section. Therefore, if the pitch information can be accurately detected, it is possible to increase the accuracy of recognition through the formant frequency by minimizing the influence of the speaker in speech recognition, and in the case of speech synthesis, the formant frequency and By separating the constituent components and synthesizing them arbitrarily, the nature and personality can be easily changed and maintained.

In addition, in the analysis, it is possible to obtain high sound quality through accurate vocal parameters by removing the influence of the glottal and reducing the error according to the analysis by synchronizing with the pitch.

On the other hand, in general, ideological constitution refers to human constitutions presented in the ideological medicine system, which was originally invented by Mr. Jema Lee, a friend of the Joseon Dynasty. According to ideology, all people are classified into four constitutions, Sun, Taeumin, Soyangin, and Soinin, which are the basics of Sasang constitution. It is widely known not only in the field of oriental medicine but also in the private sector that the methods of treatment and the medicines used must be different.

In addition to this filamentous constitution, hot and cold constitutions can be further distinguished, and further subdividing the human constitution into so-called eleven constitutions can lead to a more precise and accurate classification of the human constitution. You will be able to cope more precisely and accurately in maintaining and treating your health.

However, it is difficult to accurately discriminate between the various constitutions of any one of these various constitutions, and there are various conventional methods of discriminating human constitutions.

In other words, the method using the vein, how to find a gold ring or silver ring on the finger, the method using a questionnaire and the like is the typical method.

The method using the climax is to find out the constitution through the human climax and to classify the constitution of a person by the intensity, length and frequency of the minute pulse. However, this method is difficult not only for the general public but also for beginner Chinese doctors, and it is known that a doctor having a novice career has an error rate of about 30 to 40%.

In addition, the method of detecting a gold ring or a silver ring on a finger is to insert a silver ring or a gold ring on a finger, that is, a thumb, middle finger, an unknown finger or a ring of the liver, rain, lungs, or god, and measure the energy of the human body in response to the constitution. That's how. The above method is also a method for discriminating the constitution by the size of the liver, spleen, lung, and kidney, and there is a problem in that it is possible to neglect the whole aspect of the person in discriminating the constitution.

In addition, the questionnaire survey method is a statistical survey method to find out the constitution by identifying various data such as the subject's physical condition, family history, personality, and preferences by using a questionnaire. Adapted to purpose, the conclusions of the survey are obtained through analysis of the collected data. Not only does this method suffer from inaccuracies, it also has some unstructured aspects that could be improved.

The present invention has been made in view of the above-described problems, and a first object of the present invention is to provide a pitch analysis method using a fast Fourier transform method.

In addition, a second object of the present invention is to provide a pitch analysis apparatus for performing the above-described pitch analysis method.

Another object of the present invention is to provide a method for discriminating filamentous constitution using voice.

In addition, a fourth object of the present invention is to provide a filamentous constitution discrimination system for performing the filamentous constitution discrimination method.

Further, a fifth object of the present invention is to provide a computer-readable recording medium incorporating the above-mentioned filamentous constitution discrimination program.

1 is a view for explaining a filamentous constitution discrimination system according to the present invention.

2 is a waveform diagram for explaining a digital audio signal output from the A / D converter.

3 is a diagram for explaining a pure voice signal output from the pure voice signal extraction unit.

4 is a diagram for describing a time domain signal output from the time domain analyzer.

5A to 5D display the shape and frequency spectrum of a general voice signal. In particular, FIGS. 5A and 5B are diagrams for explaining the shape of voiced and unvoiced sounds, and FIGS. 5C and 5D are FIGS. 5A and 5B. Is a view for explaining the relationship between formant and pitch through the frequency spectrum of each signal.

FIG. 6 is a diagram for describing the pitch analyzer of FIG. 1.

FIG. 7 is a diagram for describing a pitch analysis signal output from the pitch analyzer of FIG. 6.

FIG. 8 is a view for explaining the formant analyzer of FIG. 1.

9 is a diagram for explaining a formant analysis signal.

FIG. 10 is a diagram for explaining the frequency analyzer of FIG. 1.

11 is a simulation diagram for explaining a frequency analysis signal.

12 is a simulation diagram for explaining the three-dimensional gram analysis of FIG. 11 described above.

13 is a simulation diagram for explaining three-dimensional signal analysis.

FIG. 14 is a simulation diagram of FIG. 13 illustrating the X and Y axis directions in the Z axis direction.

15 is a view for explaining an example of the constitution identification using the constitution classification system according to the present invention.

16 is a flowchart illustrating a method for analyzing filamentous constitution according to the present invention.

17 is a flowchart illustrating a series of processes for extracting the pure voice signal of FIG.

FIG. 18 is a flowchart for explaining a digital signal processing procedure for the classification of FIG.

FIG. 19 is a flowchart for explaining the formant analysis process of FIG. 18 described above. FIG.

20 is a flowchart for explaining the pitch analysis of FIG. 18 described above.

21 is a flowchart for explaining the frequency analysis of FIG. 18 described above.

FIG. 22 is a flowchart for explaining the novel formant analysis method of FIG. 18.

23 is a view for explaining an example of the questionnaire form for discriminating the ideological constitution of the voice hero in accordance with the present invention.

<Description of the symbols for the main parts of the drawings>

100: A / D converter 200: pure voice signal extraction unit

300: time domain analysis unit 400: analysis unit

410: pitch analysis unit 420: formant analysis unit

430: frequency domain analysis unit 510: three-dimensional gram display unit

520: feature factor extraction unit 600: feature factor fusion unit

700: Database 800: Constitution Analysis

Pitch analysis method according to one feature for realizing the first object of the present invention, the pitch of the voice generated due to the periodic characteristics of the gate opening and closing and used in the speech coder, speech recognition, speech conversion, etc. A pitch analysis method for analyzing, comprising: (a) designing a low pass filter and setting a loop count count; (b) checking whether the current loop number count value is larger than the set loop number counter value; (c) if it is checked in step (b) that the current number of loops is less than the preset number of loops, reading data and demodulating the read data; (d) decimating the demodulated data; (e) increasing the current loop number and removing the DC component of the decimated signal; (f) taking an FFT on the de-decimation signal from which the DC component has been removed and performing spectral prediction, and then feeding back to the step (b); (g) a loop number counter preset by the current loop number in the step (b); Extracting a feature factor if checked to be equal to or greater than a value; And (h) adjusting the display scale to display a color corresponding to the contour line.

In addition, the pitch analysis device according to one feature for realizing the second object of the present invention, generated due to the periodic characteristics of the gate opening and closing of the speech used in the speech coder, speech recognition, speech conversion, etc. A pitch analysis device for analyzing a pitch, comprising: a band pass filter for outputting only a signal included in a desired frequency bandwidth among voice signals provided from the outside; A square detector detecting a square of the reduced frequency bandwidth; A low pass filter for reducing a frequency bandwidth and outputting a signal having a reduced frequency bandwidth by outputting only a signal up to a predetermined frequency in which a pitch may exist among voice signals and removing the rest; A decimation unit configured to receive a voice signal having the reduced frequency bandwidth and decimate at a maximum decimation rate at which no aging occurs; A high pass filter removing a DC signal component included in the decimated signal and outputting a desired frequency band signal; A hanning window modeling a spectrum of the high pass filtered signal; A real-time Fourier transform unit for performing a real-time Fourier transform process on the spectrum modeled through the hanning window unit; A linear integrator for integrating the real-time Fourier transformed spectrum; And a normalizing unit configured to output a pitch analysis signal after normalizing the integrated frequency spectrum.

In addition, the filamentous constitution discrimination method according to one feature for realizing the third object of the present invention, (a) by checking whether or not the audio recording for Sasang Constitution Discrimination, recording if the voice is recorded Digitally converting and storing the speech; (b) extracting pure voice data from the stored voice data; And (c) performing a digital signal processing process for classifying filamentous constitution based on the pure voice data, and outputting information for classifying filamentous constitution discrimination.

In addition, the filamentous constitution discrimination method according to another feature for realizing the third object of the present invention includes (a) checking whether or not the voice is recorded by checking whether or not the sacrificial constitution discrimination voice is recorded. Digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; (c) checking whether the prediction is completed by the basic survey for Sasang Constitution Discrimination; (d) if it is checked in step (c) that the prediction has been completed, processing the prediction information to display the first Sasang Constitution discrimination information; (e) storing the prediction information processed in step (d) and checking whether to set the filamentous constitution classification; And (f) if the classification setting is checked in step (e), performing a digital signal processing process based on the classification, and outputting the classification information.

In addition, the ideological discrimination system according to one feature for realizing the fourth object of the present invention includes an analog-to-digital converter for converting an analog voice signal provided from a microphone into a digital voice signal; A pure voice signal extracting unit which performs low pass filtering on the digitally converted voice signal and extracts a pure voice signal by extracting a signal having a predetermined threshold or more; A time domain analyzer for analyzing the waveform of the pure voice signal and outputting time domain analysis information; An analysis unit configured to output first analysis information through pitch analysis on the time domain analysis information, output second analysis information through formant analysis, and output third analysis information through frequency domain analysis; A gram display and a feature factor extractor configured to display the first to third analysis information and extract a feature factor; A feature factor fusion unit for fusing and outputting the extracted feature factor; And a Sasang Constitution Analysis unit which checks the output feature factors and outputs Sasang Constitution analysis information corresponding to the corresponding speech signal.

In addition, a computer-readable recording medium incorporating a Sasang Constitution Discrimination Program according to one feature for realizing the fifth object of the present invention includes (a) checking whether the Sasang Constitution Discrimination Voice recording is performed by If it is checked as recorded, digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; And (c) performing a digital signal processing process for classifying filamentous constitution based on the pure voice data, and outputting information for classifying filamentous constitution discrimination.

In addition, a computer-readable recording medium incorporating a Sasang Constitution Discrimination Program according to another feature for realizing the fifth object of the present invention includes (a) checking whether the Sasang Constitution Discrimination Voice recording is If the voice is checked as being recorded, digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; (c) checking whether the prediction is completed by the basic survey for Sasang Constitution Discrimination; (d) if it is checked in step (c) that the prediction has been completed, processing the prediction information to display the first Sasang Constitution discrimination information; (e) storing the prediction information processed in step (d) and checking whether to set the filamentous constitution classification; And (f) if the classification setting is checked in step (e), performing a digital signal processing process based on the classification, and outputting the classification information.

According to such a pitch analysis method and apparatus, a filamentous discriminant method and system, and a computer-readable recording medium incorporating a filamentous discriminant program, an acoustic characteristic of a voice input through a microphone is extracted and extracted. Based on the characteristic factors, the ideological constitution of the main character can be discriminated.

Hereinafter, with reference to the accompanying drawings, it will be described in detail the present invention.

1 is a view for explaining a filamentous constitution discrimination system according to the present invention.

Referring to FIG. 1, the Sasang Constitution Discrimination System according to the present invention includes an A / D converter 100, a pure voice signal extractor 200, a time domain analyzer 300, an analyzer 400, a gram display and a feature. Factor extraction unit 500, feature factor fusion unit 600, database 700 and the constitution analysis unit 800 is included.

The A / D converter 100 digitally converts an analog type voice signal converted into an electric signal provided from the microphone 90, and converts the digital signal into a pure voice signal extractor 200 as shown in FIG. 2. To provide. In this case, the sampling rate of the digital signal is twice the maximum frequency of the signal according to Nyquist's sampling theory. In the present invention, the sampling frequency (Fs) is set to 11025 kHz and stored as a wave (WAV) file using a 16-bit digital recorder.

The pure voice signal extractor 200 extracts only the pure voice signal through a technique of performing a low pass filtering on the digitally converted voice signal and extracting a signal above a predetermined threshold (or a reference threshold). As shown in FIG. 3, the extracted pure audio signal is provided to the time domain analyzer 300. Here, the reference threshold is a value obtained by multiplying a maximum value of a signal by a predetermined coefficient.

The time domain analyzer 300 analyzes the waveform of the pure voice signal provided from the pure voice signal extractor 200, and provides the analyzed time domain information to the analyzer 400 as shown in FIG. 4. In this case, the time domain analyzer 300 is configured through a high pass filter, a DC notch filter, and an x-square detection module.

The analyzer 400 includes a pitch analyzer 410, a formant analyzer 420, and a frequency domain analyzer 430, and performs pitch analysis on the time domain analysis information provided from the time domain analyzer 300. Information analyzed through pitch analysis, formant analysis, and frequency domain analysis is provided to the gram display and the feature factor extractor 500.

In more detail, the pitch analyzer 410 precisely analyzes low frequency regions of speech to extract basic speech frequencies and outputs pitch analysis information to the gram display and feature factor extractor 500. The pitch analysis information is information about the peak frequency, magnitude, and bandwidth of the pitch. In general, the pitch in a speech signal refers to the most basic frequency of the speech signal, that is, the frequency of peaks appearing on the time axis, and the term fundamental frequency generated by periodic shaking of the vocal cords. Used as a synonym. This pitch is a parameter that is very sensitive to human hearing, and is used to distinguish the speaker (main character) of a voice signal and greatly affects the tone of the voice signal.

5A to 5D display the shape and frequency spectrum of a general voice signal. In particular, FIGS. 5A and 5B are views for explaining the shapes of voiced and unvoiced speech, and FIGS. 5C and 5D. 5A and 5B are diagrams for describing a relationship between formants and pitches through the frequency spectrum of each signal.

As shown in Fig. 5C, the cogwheels are formed in the same coarse cycle as the whole, and the interval between adjacent cogwheels is the pitch frequency in the frequency domain. On the other hand, the unvoiced spectrum shown in Fig. 5D exhibits almost no pitch characteristics.

As such, pitch is an important parameter for improving high quality synthesis, speech compression, speech encoding or speech recognition rate. Studies to extract pitch include time domain analysis, frequency domain analysis, and combination domain method combining them. In general, pitch analysis uses a Fourier transformed frequency signal for the log and then inverse Fourier transform (Cepstrum), but the present invention precisely analyzes only the low frequency band. This is because the desired low frequency band can be selected and the selected low frequency band can be analyzed. The pitch analysis unit 410 will be described in more detail later with reference to FIG. 5.

In addition, the formant analyzer 420 detects an envelope waveform of the voice appearing on the frequency and outputs formant analysis information. Normally, in order to make a desired sound, a moderate amount of air is compressed in the lungs, and the shape of the previously remembered saints is formed, and each one passes through different saints, and the human ear senses and recognizes the signal. From the standpoint of signal processing, white noise is generated in the lung, and resonance occurs at the frequency corresponding to the characteristic as it passes through the saint and is canceled out. This can be seen in instruments such as pipe organs, where the resonant frequencies occurring in the vocal cords and nasal passages are called the formant frequency, or simply the formant.

In FIG. 5C, it can be seen that the overall shape of the spectrum has three peaks. In other words, one large peak is seen at approximately 700 [Hz], 1200 [Hz] and 2600 [Hz], which is the formant frequency observed in the frequency domain.

As such, the formant depends on the vocal cord's geometry and a particular voice signal can be represented by several representative forms. For example, the sound of 'a' and 'er' can be produced by changes in human vocal tracts, and the formant frequency at this time is different.

In particular, the present invention outputs formant analysis information by detecting an envelope of a voice based on a gain and a band operated by an operator using a multiple adaptive filter. The formant analysis information is information about the peak frequency, magnitude, and bandwidth of the formant. Description of the formant analyzer 420 will be described in more detail later with reference to FIG. 8.

In addition, the frequency domain analyzer 430 divides the voice signal into a predetermined band level over the entire band, and then outputs frequency domain analysis information by performing precise frequency analysis on each of the divided bands. In this case, the frequency domain analysis information includes tonal characteristics, correlations between tonal frequencies, sum of energy, energy average, average of tonal frequencies, and standard deviation.

That is, even if the pure voice recorded within the same time, there is a compound sound in which several sounds constituting the pure voice are mixed, so when the visible light passes through the prism, the rainbow is changed from red to purple according to the frequency range of the light constituting the visible light. As the color is spectroscopically, the sound is also classified into sound pressure levels for each frequency region constituting the sound by the frequency domain analyzer 430. The magnitude of the sound pressure level on the X axis and the magnitude of the sound pressure level on the Y axis by analyzing the magnitude of the sound pressure level for each component frequency is called an acoustic spectrum. The description of the frequency domain analyzer 430 will be described in more detail later with reference to FIG. 10.

The gram display and feature factor extractor 500 includes a three-dimensional gram display unit 510 and a feature factor extractor 520, and include a pitch analyzer 410, a formant analyzer 420, and a frequency domain analyzer 430. ), Pitch analysis information, formant analysis information, and frequency domain analysis information provided from each of the plurality of pieces are displayed, and the feature factors are extracted and provided to the feature factor fusion unit 600. In this case, the extracted feature factors include the peak frequency for each frequency band, the magnitude of the peak frequency for each frequency band, and the like.

The feature factor fusion unit 600 fuses the information provided from the gram display and the feature factor extractor 500, and provides the constitution analysis unit 800 with the fusion information obtained by processing the data for pattern matching or neural network input. do.

The database 700 stores the Sasang Constitution information primarily determined by the voice main character based on the contents of the predetermined questionnaire as shown in FIG. 23, and the Sasang Constitution information of the main character from the constitution analyzer 800. In response to the request, the primary constitution information is determined.

The constitution analyzer 800 determines whether the main character of the voice is a sun person, a Taeumin person, or a Soyang person, or a noise person based on the fusion information provided from the feature factor fusion unit 600 and the primary Sasang constitution information stored in the database 700. Print information about In this case, when the fusion information provided from the feature factor fusion unit 600 is information for pattern matching, the extracted feature factors are patterned, recognized, and compared to the pattern data corresponding to previously stored feature factors. Will be able to discriminate.

A / D converter 100, pure voice signal extractor 200, time domain analyzer 300, analyzer 400, gram display and feature factor extractor 500, feature factor fusion unit (described above) 600) and the constitution analyzer 800 may be implemented as a separate IC and applied to a computer system.

On the other hand, when the fusion information provided from the feature factor fusion unit 600 is data for neural network input, the neural constitution information, which is primarily determined in the database 700, is requested through the neural network that has already been learned, even if it is not separately requested. I can discriminate the ideological constitution of the main character of the voice.

FIG. 6 is a diagram for describing the pitch analyzer of FIG. 1.

Referring to FIG. 6, the pitch analyzer 410 according to the present invention includes a band pass filter (BPF) 411, a square law detection module 412, a low pass filter (LPF) 413, and a desiccant. The simulation module 414, the high pass filter (HPF) 415, the Hanning Window module 416, the real-time fast Fourier transform module 417, the linear integration module 418, and the normalization module 419. In addition, pitch analysis is performed on the time domain analysis information provided from the time domain analyzer 300, and the pitch analysis information according to the result is provided to the gram display and the feature factor extractor 500.

The band pass filter 411 reduces the frequency bandwidth and outputs a signal having a reduced frequency bandwidth by outputting only a signal included in a predetermined bandwidth for pitch analysis at a desired frequency bandwidth among voice signals and removing the rest. 412).

The square detection module 412 squares the reduced bandwidth provided from the band pass filter 411 and outputs the result to the low pass filter 413.

The low pass filter 413 reduces the frequency bandwidth and outputs a signal having the reduced frequency bandwidth to the decimation module 414 by outputting only a signal up to a predetermined frequency in which a pitch may exist among voice signals and removing the rest. to provide.

The decimation module 414 receives a voice signal having a reduced frequency bandwidth, decimates at a maximum decimation rate at which no aliasing occurs, and then provides the decimation module 414 to the high pass filter 415. The purpose of decimation is to reduce the data sample rate and to increase the resolution of the frequency. In general, frequency resolution of the same number of frequency data is determined according to the sample rate of the data.The data sample rate before this processing stage is 11,025Hz, but the sample rate is lowered by four times or more by decimation. Better than This is to precisely analyze the sample data in the frequency domain.

The high pass filter 415 completely removes the DC signal component included in the decimated signal provided from the decimation module 414 to make signal analysis much easier. In addition, the unwanted frequency bands are removed through high pass filtering, and the spectrum of the corresponding frame is modeled through the hanning window module 416 and then provided to the real-time fast Fourier transform module 417. The Hanning window is used to greatly weaken the signal component of the sidelobe during the frequency analysis. This is one of the accurate analysis of voice signal and preprocessing of noise.

The real-time fast Fourier transform module 417, preferably the real-time Fast Fourier Transform (N-FFT) module, performs a real-time Fourier transform process on the spectrum modeled through the Hanning window, and the linear integration module 418 performs Fourier transform Integrating with respect to the spectrum, the normalization module 419 normalizes the processed frequency spectrum, and then characterizes the pitch analysis signal as shown in FIG. 7 by the factor extractor 510 and the three-dimensional gram display 520. To each). The normalization module 419 prevents the same signal from being incorrectly classified according to various situations, for example, a noisy environment, by correlating the voice signal to a relative size rather than analyzing the absolute signal size. This is to ensure the objectivity of the feature factors.

In general, the method of obtaining the speech parameter includes analyzing the power spectrum of the logarithmic spectrum, that is, the cepstrum, and the method of analyzing the cepstrum, which can distinguish voiced and unvoiced sound, and linear prediction. The use of orthogonal magnetic field correlation coefficients instead of coefficients is known as the polarization coefficients (PARCOR) and the linear prediction method (LPC) that predicts the current signal by past signals. Fast Fourier Transform (FFT) method is used to reduce the calculation process using the properties of periodicity and symmetry.

Meanwhile, the band pass filter (BPF) 411, the square law detection module 412, the low pass filter (LPF) 413, the decimation module 414, and the high pass filter (HPF) described above. The 415, the Hanning Window module 416, the real-time fast Fourier transform module 417, the linear integration module 418, and the normalization module 419 may be implemented as separate ICs and applied to computer systems. will be.

FIG. 8 is a view for explaining the formant analyzer of FIG. 1.

Referring to FIG. 8, the formant analyzer 420 according to the present invention includes a hanning window module 421, a mathematical auto-regressive model calculation module 422, a transfer function extraction module 423, and a frequency. Time provided from time domain analyzer 300, including response module 424, multiple adaptive band pass filter 425, Hanning window module 426, real time FFT module 427 and normalization module 428 Formant analysis is performed on the area analysis information, and the formant analysis information is provided to the gram display and the feature factor extraction unit 500 according to the result.

The hanning window module 421 receives the time domain analysis information provided from the time domain analyzer 300 to model the spectrum, and provides the modeled signal to the mathematical autoregression model calculation module 422.

The mathematical autoregressive model calculation module 422 estimates parameters for the signal modeled by the Hanning window module 421, preferably using a Yule-Walker modeling equation, Equation 1 is as follows.

The transfer function extraction module 423 is calculated by the Yule-Walker modeling equation and extracts the transfer function of the output signal for the input signal using the estimated parameters. Here, the transfer function mentioned is equivalent to expressing the characteristics of input and output as a function in a general system.

The frequency response module 424 extracts a frequency response characteristic corresponding to the extracted transfer function provided from the transfer function extraction module 423, and extracts the frequency response characteristic from the factor extractor 510 and the three-dimensional gram display unit. 520 respectively.

On the other hand, the multi-adaptive band pass filter 425 receives the time domain analysis information provided from the time domain analyzer 300 to filter the time domain analysis information by varying the desired frequency band and gain, respectively, and then a hanning window. To module 426. At this time, the frequency band or gain may be set differently. The use of the multi-adaptive band pass filter 425 is to increase the objectification and transparency of the signal by removing the signal characteristics peculiar to the voice signal and the fixed characteristics of each voice. It is also to improve the signal-to-noise ratio for the band for intensive analysis.

The hanning window module 426 models a signal filtered through different bandwidths and gains through a hanning window and provides the modeled signal to the real-time FFT module 427. Here, as the length of the hanning window used increases, the spectrum is refined and the spectral resolution is improved. On the other hand, considering the characteristic that the length of the time axis is inversely proportional to the bandwidth in the frequency domain, the time resolution decreases, which causes a phenomenon in which the temporal change of the speech signal characteristic cannot be sufficiently modeled. There is a need for a trade off process between spectral resolution and time resolution.

The real-time FFT module 427 receives the modeled signal from the Hanning window module 426, performs a real-time Fourier transform, and then provides it to the normalization module 428. The normalization module 428 normalizes the Fourier transformed signal. After processing, the formant analysis signal is provided to the feature factor extractor 510 and the 3D gram display unit 520, respectively, as shown in FIG.

Meanwhile, the Hanning window module 421 described above, the mathematical auto-regressive model calculation module 422, the transfer function extraction module 423, the frequency response module 424, and the multiple adaptive band pass filter 425. ), The Hanning window module 426, the real-time FFT module 427, and the normalization module 428 may be implemented as separate ICs and applied to computer systems.

Referring to Figure 9, it can be seen that the waveform of six peaks, this peak means the formant.

Of course, in FIG. 8, the analysis of the formant using a Hanning Window, which is superior in modeling a spectrum than a Rectangular Window, has been made, but a Hamming Window, a Parzen Window, It can also be implemented using a Bartlet Window.

FIG. 10 is a diagram for explaining the frequency analyzer of FIG. 1.

Referring to FIG. 10, the frequency analyzer 430 may include a square detection module 431, a low pass filter 432, a decimation module 433, a high pass filter 434, a data overlap and frequency shifting module 435, Frequency domain analysis information calculated through a frequency analysis process on the time domain analysis information provided from the time domain analyzer 300, including the band selection module 436, the interpolation module 437, and the normalization module 438. It is provided to the feature factor extraction unit 510 and the three-dimensional gram display unit 520, respectively.

The square detection module 431 squares and processes the time domain analysis information provided from the time domain analyzer 300 to the low pass filter 432, and the low pass filter 432 outputs the squared time domain analysis information. Low pass filtering provides only the low band information to the decimation module 433.

The decimation module 433 receives the voice signal having the reduced frequency bandwidth from the low pass filter 432 to decimate at the maximum decimation ratio without aliasing and then provides the decimation module to the high pass filter 434. do. The purpose of decimation is to reduce the data sample rate and to increase the frequency resolution. In other words, the frequency of the same number of frequency data is determined according to the sample rate of the data. The data sample rate before the decimation processing stage is 11,025 Hz, but the data sample rate is lowered by 1/4 or more through the decimation processing stage. The frequency resolution is better than four times better. This is to precisely analyze the sample data in the frequency domain.

The high pass filter 434 performs high pass filtering on the decimated time domain analysis information to provide only the high band information to the data overlap and frequency shifting module 435.

The data overlap and frequency shifting module 435 overlaps the data in the time domain so as not to cause an aliasing in the frequency domain, shifts the frequency corresponding to the overlapped data, and provides the band selection module 436. Here, frequency shifting is for precisely analyzing a signal of a desired band. In other words, in order to zoom the decimated analysis signal, low pass filtering (a band of 1 ㎑ or less) must be performed, and a band to be precisely analyzed may be in a high band (1 ㎑ or more band). Therefore, after shifting the signal in the high band to the low band, the signal can be enlarged through low pass filtering for easier analysis. In addition, overlapping the data is to minimize the noise signal component caused by the data breakage during the conversion to the frequency domain and to accelerate the continuity of the signal.

The band selection module 436 selects a desired bandwidth among the shifted frequencies, provides a signal of the selected band to the interpolation module 437, and the interpolation module 437 performs an interpolation process on the selected bandwidth and then normalizes the module. In operation 438, the normalization module 438 normalizes the interpolated signal and provides it to the feature factor extractor 510 and the 3D gram display 520, respectively. The interpolation process performs data smoothing using cubic spline interpolation to provide more accurate information to the feature factor extractor 510 and the 3D gram display 520. More than 10 times more data is calculated and interpolated between the actual data and the data, so that it is easy to extract precise information and smoothly display or display according to the data smoothing.

Meanwhile, the square detection module 431, the low pass filter 432, the decimation module 433, the high pass filter 434, the data overlap and frequency shifting module 435, the band selection module 436, The interpolation module 437 and the normalization module 438 may be implemented as separate ICs and applied to computer systems.

11 is a simulation diagram for explaining a frequency analysis signal.

Referring to FIG. 11, the X-axis is represented by the frequency domain, the Y-axis is represented by the amplitude (or energy level) in decibels (dB), and three tonals on the frequency range from 0 to 1 Hz, that is, 0.25 Hz. It can be seen that the tonal is in the region of 0.5 ms and 0.75 ms. The tonal frequency may be used to identify the correlation between the tonal frequency and the tonal frequency.

Of course, although only the frequency ranges from 0 to 1 kHz are shown in the drawings, the frequency ranges from 1 GHz to 2.1 GHz, 2.1 GHz to 3.1 GHz, 3.1 kHz to 4.1 GHz, and 4.1 kHz to 5 GHz can also be shown. There will be a correlation between the tonal frequency and the tonal frequency in each frequency domain.

12 is a simulation diagram for explaining the three-dimensional gram analysis of FIG. 11 described above.

Referring to FIG. 12, the above-described FIG. 11 is observed from the viewpoint of amplitude. In particular, the X-axis is observed as a frequency domain in units of [Hz] and the Y-axis is observed in a time domain in units of [sec]. Drawing. As can be seen, the frequency line can be traced at a tonal frequency around 0.25 Hz. More specifically, frequency line tracking is performed to calculate the frequency variation over time of the tonal frequency as data information, wherein a Kalman filter or an alpha filter is applied to estimate the data. To track the optimal frequency line.

13 is a simulation diagram for explaining three-dimensional signal analysis.

Referring to Fig. 13, the X-axis direction shows a frequency range from 0 to 4.5 kHz, the Y-axis direction shows a time domain from 0 to 300, and the Z-axis direction shows an energy level of 0 to 12 ranges. While showing the signal analysis waveform in three-dimensional space.

FIG. 14 is a simulation view in which the X and Y axis directions of FIG. 13 are observed in the Z axis direction. In particular, FIG. 13 illustrates a frequency domain in the X axis direction and a time domain in the Y axis direction in the Z axis direction. In this case, although not illustrated in the Z-axis direction, it can be confirmed that the color representation is sufficiently possible.

15 is a view for explaining an example of the constitution identification using the constitution classification system according to the present invention.

Referring to FIG. 15, the feature factor extractor 510 compares, analyzes, and compares a peak frequency detected after noise cancellation for each frequency band, a calculated average energy for each band, and a standard deviation calculated after detecting a fundamental frequency from the feature factor extractor 510. Judging whether the hero of the voice is Sunin, Taeininjin, Soyangin, or Noisein and outputs.

That is, when it is determined that a voice signal is detected in a high frequency band, it is classified as a solar phosphorus and output. In this case, voices classified as sun people generally have high voice, clear voice, and roundness, and are known to harmonize with 'sang' sound.

In addition, when it is determined that the sum of energy over the entire band of the voice signal is high, it is classified and output as Taeumin. At this time, voices classified as Taeumin are generally known to have heavy, muddy, fuzzy characteristics, and harmonize with 'bow' sounds.

If it is determined that the sum of energy over the entire band of the voice signal is low, it is classified as a person and outputted. In this case, voices classified as Soyangin are generally light, low, and withdrawal, and harmonize with the 'chi' sound.

In addition, when it is determined that the standard deviation of the fundamental frequency of the audio signal is large, the noise is classified and output as noise noise. At this time, the voice classified as a noise person is generally activated, has a characteristic of being smooth and flat, and harmonizes with the 'right' sound. Of course, by correcting the filamentous constitution of the voice main character (Fixing) may output any one of the sun person, Taeumin person, Soyangin person, noise person, or may be output as a probability corresponding to each of the classified filamentous constitution.

16 is a flowchart illustrating a method for analyzing filamentous constitution according to the present invention.

Referring to Fig. 16, first, it is checked whether or not a voice is recorded (step S10), and when the voice is checked as being recorded, the voice is converted into a digital signal because it is an analog signal (step S20).

The digitally converted voice is then stored (step S30). At this time, it may be saved automatically or manually by an operator's operation.

Subsequently, not only the main character's voice but also the surrounding noise components are recorded in the stored voice data, thereby extracting only the pure voice data from which the noise is removed (step S40). The pure voice data extraction step will be described in more detail later with reference to FIG. 16.

Then, it is checked whether the prediction is completed by the basic survey (step S50). In this case, the basic survey may be referred to as a survey of a certain questionnaire, which may be made by a type directly written by the main character of the recorded voice. have.

When it is checked in step S50 that the prediction by the basic survey is completed, the prediction information is processed to display the constitution discrimination (step S60).

Then, the predicted information is stored (step S70), and it is checked whether or not classification is set (step S80).

Subsequently, the digital signal processing process according to the classification is performed, and information for classifying the constitution is displayed (step S90). The digital signal processing process will be described in more detail later with reference to FIG. 18.

Then, it is checked whether or not to end (step S100), and if it is checked to end, it ends.

17 is a flowchart illustrating a series of processes for extracting the pure voice signal of FIG.

Referring to FIG. 17, first, low pass filtering is performed on an analog-digital converted voice (step S410). In general, since the human voice is in the range of approximately 0 to 5 kHz, the frequency domain beyond 5 kHz is filtered out to reduce the load required for data processing.

Subsequently, since the low-pass filtered voice signal includes ambient noise in addition to the voice of the voice main character, only the pure voice signal from which the noise is removed is extracted (step S420).

Subsequently, a new file including the pure voice signal is stored (step S430). In this case, the stored file type may be a wave type.

Subsequently, the pure audio signal is displayed in the form as shown in FIG. 3 and then terminates (step S440).

FIG. 18 is a flowchart for explaining a digital signal processing procedure for the classification of FIG.

Referring to FIG. 18, a formant analysis process is first performed, a 2-D formant graph is displayed (step S910), and a formant feature factor is stored (step S920). Here, the description of the formant analysis processing will be described in more detail with reference to FIG. 19 described later.

Subsequently, the pitch processing is performed, the 3-D pitch gram is displayed (step S930), and the pitch feature factors are stored (step S940). Here, the description of the pitch processing will be described in more detail with reference to FIG. 20 to be described later.

Then, frequency analysis processing is performed, and a 3-D frequency gram is displayed (step S950). Here, the description of the frequency analysis processing will be described in more detail with reference to FIG. 21 described later.

Next, a new formant analysis process is performed (step S960), and the frequency characteristic factor is stored (step S970). Here, the description of the new formant analysis process will be described in more detail in FIG. 22 described later.

Subsequently, the feature factors are fused using pattern matching or a neural network, and further, the main character of the voice is fused with primary discrimination data for the ideology of the main character of the voice based on the data created in the form of a questionnaire as shown in FIG. 23 (step S980). ).

Here, the pattern matching and the neural network input the feature factor extracted from the accurate input information, that is, the voice information of the person whose constitution is completely determined, and the determined constitution as an output to calculate a pattern capable of objectifying and identifying the constitution. The calculated pattern is applied again by correcting the error compared to the output, and this process is repeated numerous times to obtain an optimal pattern that minimizes the error between the calculated pattern and the output.

Similarly, since neural networks repeat the learning according to the perceptron learning theory, we model the efficient classification neural network by obtaining the optimal connection strength, that is, the weight. Therefore, the pattern matching and the neural network made in step S980 have already been learned.

Then, the predictive value of the constitution discrimination is displayed (step S990). In this case, the displayed constitution discrimination prediction value may be displayed by pointing to the most likely Sasang constitution among the four Sasang constitutions, or may be displayed as a probability for several Sasang constitutions.

FIG. 19 is a flowchart for explaining the formant analysis process of FIG. 18 described above. FIG.

Referring to FIG. 19, first, a hamming window is designed for formant analysis to be performed, and a spectrum of a signal is modeled using the designed hamming window (step S9110).

Next, the modeled voice is modeled through a mathematical auto-regressive model (step S9120), and a transition function of the voice track is calculated (step S9130).

Next, the frequency response characteristic of the audio track is calculated (step S9140), and after detecting the peak point for the frequency bandwidth based on the calculated frequency response characteristic (step S9150), the feature factor is extracted (step S9160).

20 is a flowchart for explaining the pitch analysis of FIG. 18 described above.

Referring to Fig. 20, a low pass filter is first designed (step S9310), and it is checked whether the current loop number count value is larger than the preset loop number counter value (step S9320). Here, the number of loops is set in consideration of the frequency resolution of the frequency band to be analyzed.

If it is checked in step S9320 that the current loop number is smaller than the preset loop number, the data is read (step S9325) and the read data is demodulated (step S9330).

Then, a decimation process is performed on the demodulated data (step S9335). The decimation performed at this time is to reduce the sample rate on the time axis.

Then, the current loop number is increased (step S9340), and the DC notch high pass filter is passed to remove the DC component of the decimated signal (step S9345).

Next, the fast Fourier transform (FFT) is taken to perform spectral prediction (step S9350), and then fed back to step S9320.

On the other hand, if it is checked in step S9320 that the current loop number is equal to or larger than the preset loop number counter value, the feature factor is extracted (step S9360). That is, the magnitude and bandwidth of the fundamental frequency corresponding to the time axis of each frame are obtained first, and the average fundamental frequency (here, the fundamental frequency = pitch frequency) is obtained by frequency line tracking for the entire frame, and the magnitude and frequency bandwidth, and Find the frequency standard deviation.

Subsequently, the display scale may be further adjusted to smoothly display the color corresponding to the contour line (step S9365).

As such, the pitch generated as a periodic characteristic of the opening and closing of a vocal cord due to the pronunciation of a person, and analyzed as one of the important parameters used in the speech modeling process, varies in the speech coder (or vocoder, voice codec), speech recognition, speech conversion, etc. Will be able to be used.

21 is a flowchart for explaining the frequency analysis of FIG. 18 described above.

Referring to FIG. 21, first, a parameter to be processed in the frequency analysis is basically set (step S9510). The parameters set at this time are sample rate, FFT number, window size, window type, FFT overlap rate, gain for each filter, and bandwidth of filter.

Subsequently, the voice signal frequency is divided into predetermined frequency bands, and then filters corresponding to each of the divided frequency bands are designed (step S9515). For example, assuming that the audio signal is 5 kHz, 5 band band pass filters are designed with 1 kHz as one unit. At this time, the gain is different for each frequency band. Because the amplitude of the frequency is high at the low end of the voice signal and the amplitude of the frequency is low at the high end, the gain is lowered at the low end and the gain is increased at the high end to compensate for this, thereby obtaining uniform information for each frequency band.

Then, the loop count counter value is set to '1' (step S9520), and it is checked whether the current loop count count value is larger than the preset loop count counter value (step S9525). Here, the number of loops is set in consideration of the frequency resolution of the frequency band to be analyzed.

If it is checked in step S9525 that the current number of loops is smaller than the preset number of loops, the input data is partially overlapped to read the data (step S9530), and the read data is subjected to FFT processing and then frequency shifted ( Step S9535). Through this process, part of the time data input for each frame is overlapped to configure each frame, thereby enabling accurate frequency analysis for each frame. In this case, frequency shifting is performed to perform precision decimation for each band, thereby increasing frequency resolution for each band, thereby enabling accurate frequency analysis.

Subsequently, after performing interpolation processing for smoothing the frequency data analyzed for each frame in order to extract more precise tonal features (step S9540), the counter value is increased to '1' (step S9545), and then to step S9525. Feedback.

On the other hand, when it is checked in step S9525 that the current number of loops is equal to or greater than the preset number of loops, each frequency bandwidth for frequency analysis is selected (step S9550), and the energy spectrum is extracted for the selected frequency bandwidth. At the same time, the extracted energy spectrum is displayed (step S9555).

Subsequently, the average energy sum, peak frequency, magnitude and bandwidths for each frequency bandwidth are detected (step S9560).

Subsequently, the average and standard deviation of each of the tonals, the number of tonals, and the correlations between the tonals are examined (step S9565), and then the frequencies analyzed for each band are normalized (step S9570).

Then, the display scale is adjusted (step S9575), and the frequency line is traced to find the average or standard deviation of the tonal frequencies for each frame (step S9575). At this time, the display scale to be adjusted is for the observer who observes the screen to more easily observe the frequency characteristic. For example, as shown in FIGS. 12 and 13, as the energy level of the signal increases, the color of the red series is increased. To be displayed. In addition, the frequency line to be traced is a line connecting peak values, as shown in FIG. 12.

FIG. 22 is a flowchart for explaining the novel formant analysis method of FIG. 18.

Referring to FIG. 22, first, time domain analysis information is filtered with a specific gain for a specific frequency band using an adaptive multi-band pass filter (step S9610). The filtering by specific frequency bands removes the signal characteristics unique to the voice signal and the fixed features of the pronunciation to improve the objectification and transparency of the signal, and also improves the signal-to-noise ratio for the band for intensive analysis. to be.

Subsequently, the signal filtered by the multi-band is filtered using a Hamming Window (step S9615), a mathematical auto-regressive model is obtained, and the parameters are estimated (step S9620). The transfer function of the audio track is obtained (step S9625), and the frequency response of the audio track is obtained (step S9630). The steps described in the above steps S9615 to S9630 are general formant analysis processing.

Next, the frequency, amplitude, and bandwidth (Bandwidth or bin) of the peak formant frequencies are found (step S9635), and normalization is performed on the remaining formants except the highest formant (step S9640).

Subsequently, the number of formants is found for each of the bands, energy is found through the integration process for the found formants (step S9645), and the correlation between the formants is examined for each band (step S9650). ).

In the above, only the filamentous sifting discrimination method has been described. However, the method may be used as a single program by storing it on various recording media that can be read by a computer such as a floppy diskette or a CD-ROM. For example, a Sasang Constitution Discrimination Program may be embedded in a computer installed in an oriental medicine clinic or a Chinese medicine room to record voices of patients requiring Sasang Constitutional Diagnosis. Of course, it is obvious that the Sasang Constitution Discrimination Information can be provided not only through a oriental medicine clinic or a Chinese medicine room, but also through a kiosk in which the Sasang Constitution Discrimination Program according to the present invention is incorporated into a bank or a station with a high flow of people.

As another example, the voices of the clients who want to discriminate Sasang constitution through the Internet service network may be recorded and analyzed in the Internet service server, and the analyzed Sasang Constitution information may be provided to the corresponding client side through the Internet network or a wired / wireless network. will be.

Although described above with reference to the embodiments, those skilled in the art can be variously modified and changed within the scope of the present invention without departing from the spirit and scope of the invention described in the claims below. I can understand.

As described above, according to the present invention, the voice input through the microphone is stored as a data file in the storage device of the computer through an analog-to-digital conversion process, the pitch analysis method, to extract the acoustic characteristics of the voice from the stored data, Analyze using formant and frequency domain analysis, and use the analyzed voice to discriminate the ideological constitution of the main character of the voice.

Of course, such a voice analysis may be actively used as a key corresponding to the locking device, and may be used as a means of authentication or payment by identifying a specific person in cyberspace.

In addition, based on the contents of the questionnaire for discriminating ideological constitution written by the main character of the voice, the ideological constitution of the main character of the voice is primarily discriminated, and the ideological constitution of the main character of the voice is discriminated secondly by using the voice as described above. It is possible to add the reliability of the discrimination since the ideological constitution of the main character can be finally discriminated based on the ideological constitution information that is discriminated first and the ideological constitution information discriminated second.

Claims (22)

In the pitch analysis method for analyzing the pitch of speech generated due to the periodic characteristics of the gate opening and closing and used in the speech coder, speech recognition, speech conversion, etc. (a) designing a low pass filter and setting a loop count counter value; (b) checking whether the current loop number count value is larger than the set loop number counter value; (c) if it is checked in step (b) that the current number of loops is less than the preset number of loops, reading data and demodulating the read data; (d) decimating the demodulated data; (e) increasing the current loop number and removing the DC component of the decimated signal; (f) taking an FFT on the decimation signal from which the DC component has been removed and performing spectral prediction, and then feeding back to step (b); (g) extracting a feature factor when it is checked in step (b) that the current loop number is equal to or greater than a preset loop number counter value; And (h) adjusting the display scale to display a color corresponding to the contour line. In the pitch analysis device for analyzing the pitch of the voice generated due to the periodic characteristics of the gate opening and closing and used in the speech coder, speech recognition, speech conversion, etc., A band pass filter for outputting only a signal included in a desired frequency bandwidth among voice signals provided from the outside; A square detector detecting a square of the reduced frequency bandwidth; A low pass filter for reducing a frequency bandwidth and outputting a signal having a reduced frequency bandwidth by outputting only a signal up to a predetermined frequency in which a pitch may exist among voice signals and removing the rest; A decimation unit configured to receive a voice signal having the reduced frequency bandwidth and decimate at a maximum decimation rate at which no aging occurs; A high pass filter removing a DC signal component included in the decimated signal and outputting a desired frequency band signal; A hanning window modeling a spectrum of the high pass filtered signal; A real-time Fourier transform unit for performing a real-time Fourier transform process on the spectrum modeled through the hanning window unit; A linear integrator for integrating the real-time Fourier transformed spectrum; And And a normalization unit configured to output a pitch analysis signal after normalizing the integrated frequency spectrum. (a) checking whether or not the sacrificial constitution discriminating voice is recorded, and if the voice is checked as being recorded, digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; And and (c) performing a digital signal processing process for classifying filamentous constitution on the basis of the pure voice data, and outputting the classified information for classifying filamentous constitution. The method of claim 3, wherein step (c) comprises: (c-1) performing a formant analysis process on the extracted pure speech signal and displaying a 2-D formant graph; (c-2) storing the formant feature factor; (c-3) performing a pitch process on the extracted pure speech signal and displaying a 3-D pitch gram; (c-4) storing the pitch feature factors; (c-5) performing frequency analysis on the extracted pure speech signal and displaying a 3-D frequency gram; (c-6) performing a new formant analysis process; (c-7) storing the frequency feature factor; (c-8) fusing the formant feature, the pitch feature, and the frequency feature by using pattern matching or a neural network; And (c-8) displaying the predictive value of the constitution discrimination based on the fused feature factors. The method of claim 4, wherein step (c-1) comprises: (c-11) designing a hamming window and modeling a spectrum of the signal using the hamming window; (c-12) modeling the modeled speech through a mathematical autoregression model and calculating a transition function of the speech track; (c-13) calculating frequency response characteristics of the voice track; And (c-14) detecting a peak point for a frequency bandwidth based on the calculated frequency response characteristic, and extracting a feature factor based on the peak point. The method of claim 4, wherein step (c-3) comprises: (c-31) designing a low pass filter and setting a loop count counter value; (c-32) checking whether the current loop number count value is larger than the set loop number counter value; (c-33) if it is checked in step (c-32) that the current number of loops is less than the preset number of loops, reading data and demodulating the read data; (c-34) decimating the demodulated data; (c-35) increasing the current loop number and removing the DC component of the decimated signal; (c-36) taking an FFT on the decimation signal from which the DC component has been removed and performing spectral prediction, and then feeding back to the step (c-32); (c-37) extracting a feature factor when it is checked in step (c-32) that the current loop number is equal to or greater than a preset loop number counter value; And and (c-38) adjusting the display scale to display a color corresponding to the contour line. The method of claim 4, wherein step (c-5) comprises: (c-51) setting a parameter to be processed in the frequency analysis; (c-52) dividing the voice signal frequency into a predetermined frequency band, designing filters corresponding to each of the divided frequency bands, and setting the number of loops to '1'; (c-53) checking whether the current loop count count value is larger than a preset loop count counter value; (c-54) if it is checked in the step (c-53) that the current number of loops is smaller than the preset number of loops, overlapping the input data and reading data; (c-55) performing frequency shifting on the read data after performing FFT processing; (c-56) performing interpolation for smoothing the frequency data analyzed for each frame; (c-57) incrementing the counter value by '1' and feeding back to the step (c-53); (c-58) selecting each frequency bandwidth for frequency analysis when it is checked in step (c-53) that the current number of loops is equal to or greater than a preset number of loops; (c-59) extracting an energy spectrum for the selected frequency bandwidth and displaying the extracted energy spectrum; (c-60) detecting average energy sum, peak frequency, magnitude and bandwidths per frequency bandwidth; (c-61) examining the mean and standard deviation of each of the tonals, the number of tonals, and the correlations between the tonals; (c-62) normalizing the analyzed frequency for each band; And and (c-63) adjusting a display scale and tracing a frequency line to calculate an average and a standard deviation of the tonal frequencies of each frame. The method of claim 4, wherein step (c-6) (c-6a) filtering the time domain analysis information with a specific gain for each specific frequency band; (c-6b) filtering a signal filtered for each of the multiple bands using a Hamming window; (c-6c) estimating a parameter by obtaining a mathematical autoregression model of the filtered speech signal; (c-6d) obtaining a transfer function of the voice track and obtaining a frequency response of the voice track; (c-6e) finding the frequency, amplitude, and bandwidth of the peak formant frequencies, and performing normalization on the remaining formants except the first formant, which is the highest; (c-6f) finding the number of formants for each of the bands and obtaining energy through an integration process for the found formants; And (c-6g) examining the correlation between the formants for each of the bands filamentous constitution discrimination method comprising the. (a) checking whether or not the sacrificial constitution discriminating voice is recorded, and if the voice is checked as being recorded, digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; (c) checking whether the prediction is completed by the basic survey for Sasang Constitution Discrimination; (d) if it is checked in step (c) that the prediction has been completed, processing the prediction information to display the first Sasang Constitution discrimination information; (e) storing the first filamentous constitution discrimination information processed in step (d), and checking whether to set the filamentous constitution classification; And (f) if it is checked in the step (e) that the filamentous constitution classification setting is obtained, the second filamentous constitution discrimination information is obtained through a digital signal processing process on the time domain analysis signal obtained by analyzing the pure voice data in the time domain; And outputting third filamentous constitution discrimination information through fusion of the first filamentous constitution discrimination information and the second filamentous constitution discrimination information. The method of claim 9, wherein step (b) comprises: (b-1) performing low pass filtering on the analog-digital converted voice; (b-2) extracting a pure voice signal by removing noise from the low pass filtered signal; (b-3) storing the pure voice signal as a new file including the pure voice signal; And and (b-4) displaying the pure audio signal. The method of claim 9, wherein step (f) comprises: (f-1) performing a formant analysis process on the time domain analysis signal and displaying a 2-D formant graph; (f-2) storing the formant feature factor; (f-3) performing a pitch processing on the time domain analysis signal and displaying a 3-D pitch gram; (f-4) storing the pitch feature factors; (f-5) performing a frequency analysis process on the time domain analysis signal and displaying a 3-D frequency gram; (f-6) performing a new formant analysis process; (f-7) storing the frequency feature factor; (f-8) fusing the formant feature, the pitch feature and the frequency feature by using pattern matching or a neural network; And (f-8) displaying the third filamentous constitution discrimination information, which is a prediction value of the constitution discrimination. 12. The method of claim 11, wherein the step (f-8) further comprises fusing with the first Sasang Constitution discrimination information on the Sasang Constitution of the voice main character based on the data created in the form of a questionnaire. Way. The method of claim 11, wherein step (f-1), (f-11) designing a hamming window and modeling a spectrum of the signal using the hamming window; (f-12) modeling the modeled speech through a mathematical autoregression model and calculating a transition function of the speech track; (f-13) calculating frequency response characteristics of the voice track; And (f-14) detecting a peak point for a frequency bandwidth based on the calculated frequency response characteristic, and extracting a feature factor based on the peak point. The method of claim 11, wherein step (f-3), (f-31) designing a low pass filter and setting a loop count counter value; (f-32) checking whether the current loop number count value is larger than the set loop number counter value; (f-33) if it is checked in step f-32 that the current number of loops is less than a predetermined number of loops, reading data and demodulating the read data; (f-34) decimating the demodulated data; (f-35) increasing the current number of loops and removing the DC component of the decimated signal; (f-36) taking an FFT on the decimation signal from which the DC component has been removed and performing spectral prediction, and then feeding back to the step (f-32); (f-37) extracting a feature factor when it is checked in step f-32 that the current loop number is equal to or larger than a preset loop number counter value; And and (f-38) adjusting the display scale to display a color corresponding to the contour line. The method of claim 11, wherein step (f-5), (f-51) setting a parameter to be processed in the frequency analysis; (f-52) dividing the voice signal frequency into a predetermined frequency band, designing filters corresponding to each of the divided frequency bands, and setting the number of loops to '1'; (f-53) checking whether the current loop count count value is larger than a preset loop count counter value; (f-54) if it is checked in step f-53 that the current number of loops is smaller than the preset number of loops, the data is read by overlapping the input data; (f-55) performing frequency shifting on the read data after performing FFT processing; (f-56) performing interpolation for smoothing the frequency data analyzed for each frame; (f-57) increasing the counter value by '1' and feeding back to the step (f-53); (f-58) if it is checked in step f-53 that the current number of loops is equal to or greater than the preset number of loops, selecting respective frequency bandwidths for frequency analysis; (f-59) extracting an energy spectrum for the selected frequency bandwidth and displaying the extracted energy spectrum; (f-60) detecting an average sum of energy, peak frequency, magnitude and bandwidths per frequency bandwidth; (f-61) investigating the mean and standard deviation of each of the tonals, the number of tonals, and the correlations between the tonals; (f-62) normalizing the analyzed frequency for each band; And (f-63) adjusting the display scale and tracing the frequency lines to calculate the average and standard deviation of the tonal frequencies of each frame. The method of claim 11, wherein step (f-6), (f-6a) filtering the time domain analysis information with a specific gain for each specific frequency band; (f-6b) filtering the signal filtered for each of the multiple bands using a Hamming window; (f-6c) estimating a parameter by obtaining a mathematical autoregression model of the filtered speech signal; (f-6d) obtaining a transfer function of the voice track and obtaining a frequency response of the voice track; (f-6e) finding the frequency, amplitude, and bandwidth of the peak formant frequencies, and performing normalization on the remaining formants except the first formant, which is the highest; (f-6f) finding the number of formants for each of the bands and obtaining energy through the integration process for the found formants; And (f-6g) a method for discriminating filamentous constitution, comprising examining the correlation between the formants for each of the bands. An analog-digital converter for converting an analog voice signal provided from a microphone into a digital voice signal; A pure voice signal extracting unit which performs low pass filtering on the digitally converted voice signal and extracts a pure voice signal by extracting a signal having a predetermined threshold or more; A time domain analyzer for analyzing the waveform of the pure voice signal and outputting time domain analysis information; An analysis unit configured to output first analysis information through pitch analysis on the time domain analysis information, output second analysis information through formant analysis, and output third analysis information through frequency domain analysis; A gram display and a feature factor extractor configured to display the first to third analysis information and extract a feature factor; A feature factor fusion unit for fusing and outputting the extracted feature factor; And Sasang Constitution Discrimination System including the Sasang Constitution analysis unit for outputting the Sasang Constitution analysis information corresponding to the corresponding speech signal by checking the output feature factor. The method of claim 17, wherein the analysis unit, A band pass filter for outputting only a signal included in a predetermined bandwidth among the time domain analysis information; A square detection module for outputting a squared process for the reduced bandwidth; A low pass filter for outputting only a signal up to a predetermined frequency in which a pitch may exist among voice signals; A decimation module for receiving a low-pass filtered signal and decimating at a maximum decimation rate at which no aging occurs; A high pass filter removing a DC component included in the decimated signal and performing high pass filtering; A hanning window module for modeling a spectrum of the frame; A real time fast Fourier transform module for performing a real time Fourier transform process on a spectrum modeled through a hanning window; A linear integration module for integrating the Fourier transformed spectrum; And And a pitch analysis unit comprising a normalization module for providing the first analysis information obtained by standardizing the integrated frequency spectrum to the gram display unit and the feature factor extraction unit. The method of claim 17, wherein the analysis unit, A hanning window module receiving the time domain analysis information to model a spectrum and output a modeled signal; A mathematical autoregression model calculation module for estimating parameters for the signal modeled by the hanning window module; A transfer function extraction module for extracting a transfer function of an output signal with respect to an input signal using the estimated parameters; A frequency response module for extracting a frequency response characteristic corresponding to the extracted transfer function and providing the extracted frequency response characteristic to the gram display and feature factor extractor; A multiple adaptive band pass filter configured to filter the time domain analysis information by receiving the time domain analysis information and varying desired frequency bands and gains; A hanning window module for modeling a signal filtered through different bandwidths and gains through a hanning window; A real time FFT module receiving a modeled signal from the hanning window module and performing a real time Fourier transform; And And a formant analysis unit comprising a normalization module configured to provide a second analysis information to the gram display unit and the feature factor extraction unit after normalizing a Fourier transformed signal. The method of claim 17, wherein the analysis unit, A square detection module which squares the time domain analysis information; A low pass filter for performing low pass filtering on the squared time domain analysis information and outputting only the low band information; A decimation module for receiving a voice signal having a reduced frequency bandwidth from the low pass filter to decimate at a maximum decimation rate at which no aging occurs; A high pass filter for high pass filtering the decimated time domain analysis information and outputting only information of a high band; A data overlap and frequency shifting module for overlapping data in the time domain so as not to cause an aliasing in the frequency domain, and shifting a frequency corresponding to the overlapped data; A band selection module for selecting a desired bandwidth among the shifted frequencies and outputting a signal of the selected band; An interpolation module for interpolating the selected bandwidth; And And a frequency domain analyzer comprising a normalization module configured to provide third analysis information to the gram display and feature factor extractor after normalizing the interpolated signal. (a) checking whether or not the sacrificial constitution discriminating voice is recorded, and if the voice is checked as being recorded, digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; And (c) a computer-readable recording medium incorporating a Sasang Constitution Discrimination Program comprising performing a digital signal processing process for classifying Sasang Constitution on the basis of the pure voice data, and outputting the information classified Sasang Constitution Discrimination. . (a) checking whether or not the sacrificial constitution discriminating voice is recorded, and if the voice is checked as being recorded, digitally converting and storing the recorded voice; (b) extracting pure voice data from the stored voice data; (c) checking whether the prediction is completed by the basic survey for Sasang Constitution Discrimination; (d) if it is checked in step (c) that the prediction has been completed, processing the prediction information to display the first Sasang Constitution discrimination information; (e) storing the first filamentous constitution discrimination information processed in step (d), and checking whether to set the filamentous constitution classification; And (f) If it is checked in the step (e) that the filamentous constitution classification setting is obtained, second filamentous constitution discrimination information is obtained through a digital signal processing process based on the pure voice data, and the first filamentous constitution discrimination information and the A computer-readable recording medium incorporating a Sasang Constitution Discrimination Program comprising outputting third Sasang Constitution Discrimination Information through fusion with second Sasang Constitution Discrimination Information.