KR100926771B1 - Method of sasang constitution check using speech analysis - Google Patents

Abstract

PURPOSE: A method for classifying sasang constitution using a sound analysis technique is provided to improve accuracy of specific constitution classification by applying a non-parametric statistical method to a sound variable extracted through a sound analysis. CONSTITUTION: In a first step(S10), constitution distribution about a sound feature variable extracted from a plurality of sound data is investigated by a non-parametric statistical method. In a second step(S20), a maximum value and a minimum value about constitution of an arbitrary sound feature variable are searched. In a third step(S30), gRule matrix is generated from the constitution distribution of the arbitrary sound feature variable. In a fourth step(S40), specific constitution is classified by applying the specific sound feature variable of the arbitrary inputted sound data to the gRule matrix corresponding to the specific sound feature variable.

Description

How to classify Sasang Constitution through Speech Analysis {METHOD OF SASANG CONSTITUTION CHECK USING SPEECH ANALYSIS}

The present invention relates to a method for classifying filamentous constitution through voice analysis and to a method for classifying filamentous constitution through voice analysis.

The ideological constitution theory is divided into four constitutions according to the strong and weak differences in the human ledger and named as the sun person, the Taeumin person, the Soyangin person, and the noise person respectively.

The human ledger is divided into lungs, rain, liver, and god, and lungs and liver, rain and god are paired with each other, resulting in large and small differences in constitution.

This thought constitution also affects the voice of a person, and people with the same thought constitution have some commonalities with voice.

Hereinafter, the characteristics of the speech of each general constitution will be described.

First, the sunin is loud and loud and hastily and urgent. Specifically, the Sunin's voice is loud, the voice is far away, and it is hearing. In addition, the main characteristics are treble, clear syllables, and powerful voice.

Second, the Taeumin are silent and silent. Specifically, the voice of the Taeumin is Takum, slow speech, and bass. The main feature is that the voice tone is thick, the voice is heavy, and the voice is not clear.

Third, Soyang is talkative and frivolous. Specifically, the voice of Soyangin is fast speaking, treble, and clear. In addition, the main feature is that the sound is rough, like a beep sound.

Fourth, the noise person is quiet, silent, and calm. Specifically, the voice of the noise person is small in speech, has no power, and speaks slowly. In addition, the main features are clear voice, low voice, and thin line of voice.

In this way, the constitutional constituents of the target voice can be classified according to the characteristics of the speech for each constitutional constituent.

The conventional method of classifying sasang constitutions through speech has a problem in that a simple parametric statistical method is applied to a small number of extracted speech variables, or a specific speech analysis technique and its contents are ambiguous and thus it is not possible to accurately classify specific constitutions.

The present invention has been made to solve the problems described above, to provide a Sasang constitution classification method to increase the accuracy of the specific constitution classification by applying a non-parametric statistical method to extracting the speech variable using the speech analysis technology The purpose is.

It is another object of the present invention to provide a method for classifying filamentous constitution through voice analysis that can enhance the performance of a system for integrating the constitution with various information such as skin, face, and body as well as voice.

In order to achieve the object as described above, the Sasang constitution classification method through voice analysis according to the present invention extracts a plurality of voice feature variables from a plurality of voice data of which the constitution is confirmed, but the distribution of each voice feature variable for each constitution. A first step of searching by a nonparametric statistical method, a second step of searching for a maximum value and a minimum value for each of the constitutions in the distribution of the constitutions for any of the negative feature variables, and for each of the negative feature variables. After comparing the maximum values for each constitution, the first logical rule including the minimum value (uMin) searched among the maximum values for each constitution and the constitution (ui) in which the minimum value (uMin) is generated is obtained. A second log including a maximum value (lMax) searched among the minimum values for each constitution and a constitution (li) in which the maximum value (lMax) is generated after comparing the minimum values for each constitution. A third step of generating a gRule (uMin, ui, lMax, li) matrix consisting of a ical rule and a specific voice feature variable of the randomly input voice data corresponding to the specific voice feature variable gRule (uMin, ui, lMax, li) a fourth step of identifying a specific constitution through the voice data by applying to the matrix.

In addition, the fourth step may include a case in which a classification of a specific constitution is suspended for a randomly input voice data, a case in which a specific constitution is distinguished, and a case in which a specific constitution is not distinguished.

In addition, when distinguishing that the input voice data is not a specific constitution, when a value of a specific voice characteristic variable of the input voice data is arbitrarily larger than uMin of the gRule (uMin, ui, lMax, li) matrix, It can be distinguished that it is not ui by the first logical rule.

<1st Logical Rule>

if (X> uMin), then NOT ui

(X is a value of a specific voice characteristic variable of the voice data input arbitrarily)

In addition, in the case of distinguishing that the input voice data is not a specific constitution, when the value of a specific voice characteristic variable of the input voice data is arbitrarily smaller than the lMax of the gRule (uMin, ui, lMax, li) matrix, It can be distinguished that it is not li by the second logical rule.

delete

<2nd Logical Rule>

if (X <lMax), then NOT li

(X is a value of a specific voice characteristic variable of the voice data input arbitrarily)

Further, in the third step, when the voice feature variable is a voice feature variable correlated with specific body information, the sub matrix for each group of the specific body information is included in the gRule (uMin, ui, lMax, li) matrix. It may further comprise a step.

In addition, the correlation between the voice feature variable and specific body information may be 0.3 or more.

In addition, the specific body information may be a subject's age, height, weight, and body mass index (BMI).

In addition, the first step of initializing the score for each constitution of the randomly input voice data and each voice feature variable of the randomly input voice data is applied to the gRule (uMin, ui, lMax, li) matrix by applying the gRule ( uMin, ui, lMax, li) may include a second step of generating a final constitution score for each voice feature variable by assigning a -1 point as a penalty to the score of the constitution related thereto.

The fourth step may further include a third process of finding a maximum value, a median value, and a minimum value in the final constitution score of the voice feature variable, and calculating a difference value between the maximum value and the minimum value to determine the voice feature variable. A fourth step of comparing with a calculated value of number N_SP × 0.1, and if the difference value is less than the calculated value, the classification of a specific constitution is suspended; if the difference value is larger than the calculated value, the maximum value and the median value The fifth step of calculating the difference between the difference value of the median value and the minimum value, and if the difference value between the maximum value and the median value is greater than the difference value between the median value and the minimum value, distinguish the specific constitution, the maximum value and the median When the difference value of the value is smaller than the difference value between the median value and the minimum value, it may include a sixth step of distinguishing that the specific constitution.

As described above, according to the Sasang constitution classification method through speech analysis according to the present invention, it is effective to extract a speech variable using a speech analysis technique and to apply a nonparametric statistical method to increase the accuracy of a specific constitution classification.

In addition, not only the voice but also the skin, face, body shape and various information to improve the performance of the system that can be divided into constitution.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention; First, it should be noted that the same components or parts in the drawings represent the same reference numerals as much as possible. In describing the present invention, detailed descriptions of related well-known functions or configurations are omitted in order not to obscure the gist of the present invention.

13A and 13B are block diagrams of a method for classifying filamentous constitution through voice analysis according to an embodiment of the present invention.

Sasang constitution method through voice analysis according to an embodiment of the present invention, as shown in Figure 13a, the first step (S10), the second step (S20), the third step (S30) and fourth Step S40 is included.

The first step (S10) is a step of extracting a voice feature variable from a plurality of voice data of which the constitution is confirmed, the method for investigating the distribution for each voice feature variable by the non-parametric statistical method.

The non-parametric statistics refers to an analysis method that calculates a probability directly from a given data and performs a statistical test regardless of the shape of a population without assuming assumptions about parameters in statistics.

In order to extract voice feature variables from a plurality of confirmed voice data, in the present invention, 'a (a)', 'e (e)', '(i)', '(o)', 'right' (u) 'and 5 sentences and' We've been drinking fresh air 'are repeated twice to collect voice data, and 134 voice feature variables are extracted from the collected voice data.

1 is a diagram showing an extracted speech feature variable according to the present invention, FIG. 2 is a diagram showing an algorithm for obtaining a pitch according to the present invention, and FIG. 3 is a diagram showing an equation for obtaining an MDVP variable according to the present invention. 4 is a diagram showing an algorithm for obtaining a formant according to the present invention, FIG. 5 is a flowchart for obtaining an MFCC according to the present invention, and FIG. 6 is a diagram showing a distribution of arbitrary speech feature variables according to the present invention. .

As shown in FIG. 1, first, 12 voice feature variables related to pitch and formant and 13 Mel-Frequency Cepstral Coefficients are extracted for each vowel. A total of 5 vowels × 25 = 125 speech feature variables are extracted.

Secondly, a series of pitch and intensity for each sentence is obtained, and a total of 134 voice feature variables are extracted by extracting a total of 9 voice feature variables from the values of 10, 50, 90, percentile and the ratio between them. Extract

Specifically, in speech signal analysis, a pitch representing the height of sound is first obtained.

As an algorithm for obtaining the pitch, as shown in FIG. 2, an autocorrelation method may be used. In this case, two extracted variables are xF0 and xT0. In this case, x means any one of 'a (a)', 'e (e)', 'yi (i)', 'o (o)', and 'u (u)'.

In addition, as shown in FIG. 3, MDVP variables are obtained using the pitch value and the voice signal waveform. The MDVP variable is widely used in the clinic. At this time, the extracted variable is eight, xSTD, xJITA, xJITT, xRAP, xPPQ, xSHDB, xSHIM, and xAPQ.

In this case, STD, Jita, RAP, and PPQ are calculated values of pitch change in a window of a constant size, and ShDB, ShIm, APQ are calculated values of change of amplitude of a voice signal in a window of a predetermined size. Value.

In addition, as shown in FIG. 4, the first formant and the second formant are obtained by using a linear prediction coding (LPC) model and a root extraction method. At this time, two extracted variables are xF1 and xF2.

In addition, as shown in FIG. 5, Mel-Frequency Cepstral Coefficients (MFCC), which are widely used in speech analysis and recognition, are obtained. At this time, there are 13 extracted variables, that is, the twelfth order MFCC coefficient and one energy value (C0). That is, xMFCC1 to xMFCC12 and xC0.

As described above, after extracting a total of 134 voice feature variables, as shown in FIG. 6, the distribution of each voice feature variable is examined for each constitution.

7 is a diagram showing the generation of the gRule (uMin, ui, lMax, li) matrix from the distribution of any speech feature variable according to the present invention, Figure 8 is a gRule (uMin, of 4 × 134 Matrix according to the invention ui, lMax, li) matrix.

The second step (S20) is a step of searching for the maximum value and the minimum value for each constitution for the speech feature variable.

The third step (S30) is a first logical rule consisting of a minimum value (uMin) searched from the maximum value for each constitution and a constitution (ui) in which the minimum value (uMin) occurred, and the search (at the minimum value for each constitution) A step of generating a gRule (uMin, ui, lMax, li) matrix consisting of a second Logical Rule consisting of the searched maximum value lMax and the maximum value lMax generated.

Specifically, as shown in FIG. 7, the minimum value is searched at three maximum values and stored in uMin, and the constitution when the minimum value is generated is stored in ui.

In addition, the maximum value is searched for in three minimum values and stored in lMax, and the constitution when the maximum value occurs is stored in li.

Therefore, using the above four values, a logical expression of the first logical rule and the second logical rule can be obtained, and as shown in FIG. 8 using the above logical expression, gRule (4) including four condition variables ( uMin, ui, lMax, li) matrix can be created.

As described above, the gRule (uMin, ui, lMax, li) matrix can be generated by two logical expressions, so when four variables are expanded to a total of 134 voice feature variables, a total of 268 logical expressions can be made.

It can be stored and managed as a 4 × 134 matrix.

Meanwhile, as illustrated in FIG. 13B, when the voice feature variable is a voice feature variable correlated with specific body information, the third step S30 may include a sub-matrix for each group of the specific body information as gRule (uMin). , ui, lMax, li) further includes a correlation step (S31) to include in the matrix.

A total of 134 voice characteristic variables include variables that have some correlation with specific physical information such as sex, age, height, and weight of the subject. Accordingly, the correlation addition step S31 of including the sub matrix in the gRule (uMin, ui, lMax, li) matrix is a step for reflecting such correlation.

In this case, the correlation between the voice feature variable and specific body information is 0.3 or more, and the specific body information is a subject's age, height, weight, and BMI.

Specifically, the subjects are divided by gender, and the correlation between 134 voice feature variables, age, height, weight, and BMI is examined to extract a relationship having a correlation of 0.3 or more. At this time, these correlations are only for men's age and height, and women's age and BMI.

9 is a diagram illustrating a correlation between a voice feature variable and specific body information according to the present invention.

As shown in FIG. 9, in the case of men, there are nine variables having a correlation with age of 0.3 or more, eSHDB, eSHIM, iF2, P10, P50, eMFCC8, iMFCC8, oMFCC8, uMFCC8, and correlation with height of 0.3 or more. There are two variables, uF0 and P50.

In addition, in women, there are 22 variables that have a correlation with age of 0.3 or more, including aF0, aT0, eFO, iFO, oFO, oTO, uFO, p10, p50, p90, aMFCC9, aMFCC12, eMFCC6, eMFCC11, eMFCC12, iMFCC6, iMFCC9, oMFCC9, uMFCC7, uMFCC9, uMFCC12, and five variables that correlate with BMI of 0.3 or more are eFO, P10, P50, P90, and aMFCC12.

10 is a diagram illustrating a gRule (uMin, ui, lMax, li) matrix including a sub matrix for each group according to the present invention.

Variables showing a correlation of 0.3 or more are divided into groups, as shown in Table 1 below, and subgroups of gRules are generated for each group.

Then, as shown in FIG. 10, any input voice data is examined for its body information and the corresponding sub matrix is replaced with gRule.

The fourth step (S40) is a step of distinguishing a specific constitution through the voice data by applying a specific voice feature variable of the voice data randomly input to the gRule (uMin, ui, lMax, li) matrix of the specific voice feature variable. .

At this time, the fourth step (S40) includes a case of holding a classification of a specific constitution with respect to the randomly input voice data, a case of distinguishing a specific constitution and a case of distinguishing that it is not a specific constitution.

In the case of distinguishing that the input voice data is not a specific constitution, first, when a specific voice characteristic variable value of the input voice data is larger than uMin of the gRule (uMin, ui, lMax, li) matrix, It can be distinguished that it is not ui by the first logical rule.

<1st Logical Rule>

if (X> uMin), then NOT ui

(X is the value of a certain voice characteristic variable of the voice data input randomly)

In other words, if the value of a particular voice characteristic variable of the input voice is greater than uMin, the likelihood is not the minimum ui.

Secondly, when the value of a specific voice feature variable of the randomly input voice data is smaller than lMax of the gRule (uMin, ui, lMax, li) matrix, it can be distinguished that it is not li by the following second Logical Rule. .

<2nd Logical Rule>

if (X <lMax), then NOT li

(X is a value of a specific voice characteristic variable of the voice data input arbitrarily)

In other words, if the value of a particular voice characteristic variable of the input voice is less than lMax, the probability of not being at least li is greater than vice versa.

14A and 14B are block diagrams of a fourth step of FIG. 13A or 13B.

As shown in FIG. 14A, the fourth step S40 includes a first step S41 and a second step S42.

11 is a view showing the final constitution score of the randomly input voice data according to the present invention.

The first step (S41) is a step of initializing the score for each constitution of the voice data input arbitrarily.

The second step (S42) applies each voice feature variable of the voice data to the gRule (uMin, ui, lMax, li) matrix to satisfy the gRule (uMin, ui, lMax, li) matrix, A process of generating a final constitution score for each voice feature variable by assigning a -1 point to the score of the constitution related thereto.

Specifically, as shown in FIG. 11, the randomly input voice data is initialized with zero noise / quantity / tone sound scores for the voice data, and then passes the logical formula of the gRule matrix, and the condition of the logical formula is satisfied. In the event of a penalty, -1 points will be awarded in the form of penalty for the score of the related constitution.

For example, if the logical expressions of the gRule matrix are if (X> uMin), then NOT ui, and if the X value exceeds uMin in a specific voice characteristic variable, check what the ui is at this time and correspond to this ui. -1 point is given to the constitution score.

In this process, the voice data passing through the gRule matrix has a final constitution score.

On the other hand, the fourth step (S40) is, as shown in Figure 14b, after the first step (S41), the second step (S42), the third step (S43), the fourth step (S44), The fifth step (S45) and the sixth step (S46) are included.

The third step (S43) is a step of finding (finding) the maximum value, the median value and the minimum value in the final constitution score of the speech feature variable.

The fourth step S44 is a step of calculating a difference value between the maximum value and the minimum value and comparing the calculated value with the number N_SP × 0.1 of voice feature variables.

The fifth step (S45) suspends the classification of a specific constitution if the difference value is less than the calculated value, and if the difference value is greater than the calculated value, the difference value between the maximum value and the median value, the median value and the minimum value. The difference is calculated.

The sixth step (S46) distinguishes a specific constitution when the difference between the maximum value and the median value is greater than the difference value between the median value and the minimum value, and the difference value between the maximum value and the median value is the difference between the median value and the minimum value. If it is smaller than the value, it is a process to distinguish that it is not a specific constitution.

12 is a flow chart for classifying a specific constitution through voice analysis according to the present invention.

As shown in FIG. 12 in order to make a final decision on the constitution classification result, in the third step S43, the maximum value, the median value, and the minimum value are found in the constitution score given while passing through the gRule.

Subsequently, in the fourth step S44 and the fifth step S45, the difference between the maximum value and the minimum value is calculated and this value is a calculated value of the number N_SP × 0.1 of the voice characteristic variable, that is, 134 × 0.1 ≒. If it is less than 13, the classification is suspended, and the difference between the maximum and the median and the difference between the median and the minimum are calculated so that the classification is made only when there is more than that.

Then, in the sixth step (S46), if the difference between the maximum value and the median value is greater than the difference value between the median value and the minimum value, the input voice data is 'a specific constitution (having a maximum value).' On the contrary, if the difference between the maximum value and the median value is smaller than the difference value between the median value and the minimum value, the input voice data is 'not a specific constitution (having a minimum value).' Make the same distinction as

As described above with reference to the drawings illustrating a method for distinguishing filamentous constitution through voice analysis according to the present invention, the present invention is not limited by the embodiments and drawings disclosed herein, the scope of the technical spirit of the present invention Of course, various modifications can be made by those skilled in the art.

1 is a diagram illustrating an extracted speech feature variable according to the present invention.

2 illustrates an algorithm for obtaining a pitch in accordance with the present invention.

3 is a diagram showing an equation for obtaining an MDVP variable according to the present invention.

4 is a diagram illustrating an algorithm for obtaining formants according to the present invention.

5 is a flowchart for obtaining an MFCC according to the present invention.

6 illustrates the distribution of any speech feature variable in accordance with the present invention.

7 shows the generation of a gRule (uMin, ui, lMax, li) matrix from the distribution of any speech feature variable according to the present invention.

8 shows a gRule (uMin, ui, lMax, li) matrix of 4 x 134 Matrix in accordance with the present invention.

9 is a diagram showing a correlation between a voice feature variable and specific body information according to the present invention;

10 is a view showing a gRule (uMin, ui, lMax, li) matrix including a sub matrix for each group according to the present invention.

Fig. 11 is a diagram showing the final constitution score of randomly input voice data according to the present invention.

12 is a flow chart for classifying specific constitutions through voice analysis according to the present invention.

13A is a block diagram of a method for classifying filamentous constitution through voice analysis according to an embodiment of the present invention.

Figure 13b is another block diagram of a method for classifying filamentous constitution through voice analysis according to an embodiment of the present invention.

14A is a block diagram of the fourth stage of FIG. 13A or 13B.

14B is another block diagram of the fourth stage of FIG. 13A or 13B.

<Description of Symbols for Main Parts of Drawings>

S10: first step of examining the distribution of the constitution for the speech feature variable

S20: a second step of searching for the maximum and minimum values for each constitution of the voice feature variable

S30: third step of generating a gRule (uMin, ui, lMax, li) matrix

S31: Additional Correlation Step to Include the Sub Matrix in gRule (uMin, ui, lMax, li)

S40: fourth step of identifying a specific constitution through voice data

S41: First step of initializing scores for each constitution of voice data

S42: second process of generating final constitution score for speech feature variable

S43: Third process of finding the maximum, middle, and minimum values of the final constitution score

S44: Fourth step of comparing the difference between the maximum value and the minimum value with a calculated value of the number of negative feature variables x 0.1

S45: fifth process withholding the distinction of specific constitution

S46: sixth process of identifying that a specific constitution or not a specific constitution

Claims (9)

Extracting a plurality of voice feature variables from the plurality of voice data in which the constitution is confirmed, and examining a distribution of each voice feature variable for each voice feature variable by a nonparametric statistical method; A second step of searching for a maximum value and a minimum value for each constitution in the distribution for each constitution for any of the voice feature variables; A first logical that consists of a minimum value (uMin) searched among the maximum values for each constitution and a constitution (ui) in which the minimum value (uMin) is generated after comparing the maximum values of each constitution with respect to any of the voice feature variables A second logical comprising a rule and a maximum value (lMax) searched among the minimum values for each constitution and a constitution (li) in which the maximum value (lMax) is generated after comparing the minimum values for each constitution of the arbitrary voice feature variable with each other; Generating a gRule (uMin, ui, lMax, li) matrix composed of a rule; And And applying a specific voice feature variable of the randomly input voice data to a gRule (uMin, ui, lMax, li) matrix corresponding to the specific voice feature variable to distinguish a specific constitution through the voice data. Sasang constitution classification method through voice analysis. The method of claim 1, The fourth step, The method for distinguishing trait constitution through voice analysis, comprising: suspending the classification of a specific constitution with respect to randomly input voice data; and distinguishing between a specific constitution and a case in which it is not a specific constitution. The method of claim 2, When distinguishing that it is not a specific constitution of the input voice data, Speech analysis, characterized in that if the value of a particular voice feature variable of the randomly input voice data is larger than the uMin of the gRule (uMin, ui, lMax, li) matrix, it is not ui by the first Logical Rule below Sasang constitution classification method through. <1st Logical Rule> if (X> uMin), then NOT ui (X is a value of a specific voice characteristic variable of the voice data input arbitrarily) The method of claim 2, When distinguishing that it is not a specific constitution of the input voice data, Speech analysis characterized in that it is not li according to the second Logical Rule below, when the value of a certain voice feature variable of the randomly input voice data is less than lMax of the gRule (uMin, ui, lMax, li) matrix Sasang constitution classification method through. <2nd Logical Rule> if (X <lMax), then NOT li (X is a value of a specific voice characteristic variable of the voice data input arbitrarily) The method of claim 1, The third step, If the voice feature variable is a voice feature variable that correlates with specific body information, further comprising including a sub matrix for each group of the specific body information in the gRule (uMin, ui, lMax, li) matrix Sasang constitution classification method through voice analysis. delete The method of claim 5, The specific body information is Sasang constitution classification method through the voice analysis, characterized in that the subject's age, height, weight, BMI. The method according to claim 3 or 4, The fourth step, A first step of initializing scores for each constitution of the randomly input voice data; And When the speech feature variable of the randomly input speech data is applied to the gRule (uMin, ui, lMax, li) matrix, when the gRule (uMin, ui, lMax, li) matrix is satisfied, the score of the constitution related thereto And a second step of generating a final constitution score for each voice feature variable by assigning -1 as a penalty to the filamentous constitution. The method of claim 8, The fourth step, A third step of finding a maximum value, a median value, and a minimum value in a final constitution score of the speech feature variable; A fourth step of calculating a difference value between the maximum value and the minimum value and comparing the calculated value with a number N_SP × 0.1 of a voice feature variable; If the difference is less than the calculated value, the classification of a specific constitution is suspended; and if the difference is greater than the calculated value, a fifth difference value between the maximum value and the median value and a difference value between the median value and the minimum value are calculated. fair; And If the difference between the maximum value and the median value is greater than the difference between the median value and the minimum value, it is classified as a specific constitution. Sasang constitution classification method through the voice analysis, characterized in that it comprises a sixth step of classifying.

Images