KR101113770B1 - The same speaker's voice to change the algorithm for speech recognition error rate reduction - Google Patents

Abstract

The present invention relates to a speaker recognition method using a voice recognition error rate reduction algorithm for the voice change of the same speaker, to obtain a voice that has been multiple times from a plurality of speakers to classify the voice according to each speaker and to the corresponding speakers If the speech distribution is large according to the speech distribution of each speaker analyzed by the support vector machine model in the analysis step and the speech distribution diagram based on the support vector machine model, a large weight is assigned. If the speech distribution is small, the weighting step of assigning a small weight to obtain the slope according to the speech distribution and the weighting value of each speaker by calculating the weight value according to the speech distribution of each speaker are all equal. Apply a eigenvalue to each speaker so that it is a value. An individual change amount calculation step of obtaining a change amount of the component eigenvalues, and an application step of applying the main component eigenvalue change amount for each speaker obtained in the individual change amount calculation step to the speaker's training or recognition process.

The present invention classifies speech spoken by each speaker using high-dimensional vectors and grasps the classified speech distribution to give each speaker an individual weight. After that, the weights assigned to the speakers are individually assigned to the same eigenvalues so that each speaker's voice can be found and the error rate of speech recognition can be reduced.

Speech Recognition, Eigen Analysis, Eigen Value, Eigen Value Variation, Principal Component Analysis

Description

Speaker recognition method using speech recognition error reduction algorithm for speech change of same speaker {The same speaker's voice to change the algorithm for speech recognition error rate reduction}

The present invention relates to a speaker recognition method using a speech recognition error rate reduction algorithm for the voice change of the same speaker, and more particularly, when recognizing the input voice to make a basic voice or speech recognition from various speakers, When generating a session change model by analyzing the amount of information of speech according to the speech spoken by each speaker, it is possible to reduce the error rate of speech recognition by obtaining different weights for each speaker and obtain an optimal speech for each speaker. The present invention relates to a speaker recognition method using a speech recognition error rate reduction algorithm for the voice change of the same speaker.

In general, when speech recognition is performed from the speakers, the error rate of the speech recognition is that noise is generated from the speech produced by the speakers according to the external environment such as a device (microphone, telephone, etc.) receiving the spoken speech from the speaker or the surrounding environment. As it is mixed, an error rate of speech recognition occurs.

Therefore, in order to reduce the error rate generated from the external environment, various types of filters are used to remove noise from the speech spoken by the speakers, or the speech distribution diagram is set by analyzing the speech spoken by the speakers and setting a constant speech distribution diagram. All the speakers occupying the above apply the same weight uniformly, and the speakers occupying less than the constant speech distribution apply the same weight to reduce the error rate of speech recognition.

However, although the error rate of speech recognized according to the external environment can be reduced as described above, when classifying speakers, the two categories are classified into only those with a wide speech distribution and those with a low speech distribution based on the speech distribution of the speakers. When all speakers belonging to the same weight are given to each other, there is a problem in that the state of the voice that is converted according to the state of each speaker cannot be obtained correctly.

For example, even if the same speaker inputs the voice to the same device, each speaker cannot be individually weighted according to the voice distribution due to the internal environment such as the health status of the speaker rather than the health condition of the speaker. The error rate of the speech recognition for the speakers occurred because the speech for the speakers was not measured accurately.

The voice recognition error reduction algorithm for the voice change of the same speaker used in the speaker recognition method of the present invention obtains the voices uttered several times from a plurality of speakers, classifies the voices according to the corresponding speakers, and distributes the voices to the corresponding speakers. If the speech distribution is large according to the speech distribution of each speaker analyzed by the support vector machine model based on the support vector machine model based on the support vector machine model, the weight is given a large value and the speech distribution map If is small, the weighting step of calculating the slope according to the speech distribution and giving the weight of the small value, and the weight value assigned to each speaker by calculating the weight value according to the speech distribution of each speaker, Apply the eigenvalues to each speaker as much as possible so that the principal component And individual variation calculation step of obtaining an amount of change, is made the main component eigenvalues change amount for each of the speaker determined by the individual variation amount calculation step to the application step of applying the training or recognition of the speaker.

As described above, the present invention classifies speech spoken from each speaker by using a high-dimensional vector and grasps the classified speech distribution from each speaker. After assigning the individual weights to the individual weights, the weights assigned to the speakers are individually assigned to the same eigenvalue so that all of the characteristics of the speech of each speaker can be found and the error rate of speech recognition can be reduced.

In addition, by giving an individual eigenvalue to each speaker in the same manner as described above, there is an advantage that the error rate of speech recognition can be reduced by checking the difference of the minute voice that changes according to the speaker's physical state.

Looking at the configuration of the speech recognition error rate reduction algorithm for the voice change of the same speaker, used in the speaker recognition method of the present invention for achieving the above object as follows.

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

The voice recognition error reduction algorithm for the voice change of the same speaker used in the present invention obtains the uttered voices from a plurality of speakers, classifies the voices according to the corresponding speakers, and supports the voice distribution for the corresponding speakers. If the speech distribution is large according to the speech distribution of each speaker analyzed by the support vector machine model based in the analysis step and the analysis based on the model-based, the weighting of a large value and the speech distribution is small The weighting step of calculating the slope according to the speech distribution and giving the weight of the small value and calculating the value of the weight according to the speech distribution of each speaker so that the weights assigned to each speaker are equal to each other Apply the eigenvalues to the speakers to determine the amount of change in the principal component eigenvalues for each speaker. And individual variation calculation step of, is made applying step of applying the principal component eigenvalues change amount for each of the speakers calculated by the change amount calculating step to the individual training or recognition of the speaker.

에 의해 구해지도록 하고, 상기 개별변화량연산단계에서 각각의 화자들의 주성분 고유치의 변화량(α_s)은

에 의해 구해지게 된다.Here, the analyzing step is to obtain the direction vector occupied by the samples of each speaker according to the eigenvalue analysis of the speech distribution of each speaker, then to determine the distance between the samples for the direction vector and the slope of the average , The distribution of speech spoken by each speaker in the analysis step is

The change amount (α _s ) of the principal component eigenvalue of each speaker in the individual change amount calculation step is calculated by

It is saved by.

에 의해 구해지도록 한다.In addition, in the applying step, the change in the voice of each speaker through the weight associated with the change in the principal component eigenvalue

To be obtained.

Looking at the operation according to a preferred embodiment of the present invention having the configuration as described above are as follows.

First, voices obtained by obtaining multiple spoken voices from a plurality of speakers are classified according to the respective speakers, and a speech distribution diagram of the spoken voices is analyzed using a support vector machine.

In detail, the support vector machine (SVM) is a linear relational expression which is a mathematical model according to the causal relationship between independent and dependent variables for the continuous variables observed in statistical classification and regression analysis (statistics). To predict a dependent variable given an independent variable, which is an analytical method to measure the goodness-of-fit to determine how well this mathematical model describes it. Kernel tricks apply linear classification techniques to nonlinear classification problems.

In addition, as one of the most recognizable learning models among many known methods, a linear threshold element is used and a two-class pattern identifier is constructed.

This is to grasp the characteristics (aggregation and direction) of each speaker's speech distribution map and extract it as a feature (distributed speech of each speaker: connecting a large number of small vectors into one high-dimensional vectorization). will be.

Here, the speaker's speech distribution is calculated using the following equation.

[Equation 1]

 The above equation is to measure the distance (sound component grapes) between the voices spoken many times from the same speakers.

In the above equation,
s is the speaker index,
k is the session index spoken from the speakers,
m _sk is the GMM spervector (spatial distribution) of each session,
μ is the average vector of each class (sessions belonging to one speaker),
n _s is the number of sessions held by the sth speaker,
μ _s is the average sample value of the s-th speaker,
The parentheses indicate the covariance matrix of the sample,
U _s is the eigen vevtor matrix,
_{S s} is an eigen value matrix.
Λ = (λ ₁ , λ ₂ , ......... λ _n ), and λ _n is an eigenvalue.
Using the above equation, the speech spoken by the speakers is used to find a unique vector value where all values have a certain value other than zero.

Then, various voice distributions can be obtained from the plurality of speakers as shown in FIG.

Here, the more the voice distribution obtained from the speakers is clustered at a certain point, the more the speaker utters the voice.

In addition, generally, when obtaining a voice distribution diagram of a speaker, when the voice distribution degree is large, the voice distribution diagram is mixed with noise by an external environment such as the external environment (type of a device to which the voice is input from the speaker and the place where the speaker speaks). Large), and a small voice distribution is slightly different depending on the internal environment of the patient such as the health of the speakers.

In other words, the larger the voice distribution, the more it is assumed to be the cause of the external environment, and if the voice distribution is smaller, it can be assumed to be the cause of the internal environment such as the physical state of the speaker rather than the factor of the external environment.

In this way, the speech distribution for each speaker is obtained, the vector values of the speech distribution for each speaker are obtained, and then the slope is obtained for the speech distribution of each speaker.

In detail, each eigenvalue is sorted in descending or ascending order in the eigenvector matrix ∧ _s of any s-th speaker.

Then, the average value of the total sum of speech groups spoken by the speaker is calculated using the following equation.

[Equation 2]

수학식 1에서 구한 s번째 화자의 고유벡터 행렬을 ∧_s라고 했을 때, 예를 들어 각 고유치를 내림차순으로 전개하여 Λ=(λ₁,λ₂ , ,......λ_n), λ_i >λ_i+1 라 하고, 각 고유치를 속한 클래스 전체의 고유치 합으로 평균한 것을

라 표현한 것이다.
즉

,

이고,

이다.

When the eigenvector matrix of the s-th speaker obtained from Equation 1 is ∧ _s , for example, each eigenvalue is expanded in descending order and Λ = (λ ₁ , λ ₂ ,, ...... λ _n ), λ _{Let i} > λ _{i + 1} and average each sum of the eigenvalues of the classes

It is expressed.
In other words

,

ego,

to be.

Then, the speech uttered from each individual speaker obtains the amount of change or the slope of the principal component eigenvalue (where the principal component represents some characteristic of the variance of the original variable) from the covariance matrix (S10).

Then, as shown in FIG. 2, weights of the speech distributions are assigned to respective speakers according to the values obtained as described above.

For example, if the speaker's voice distribution degree obtained using the above equation is 3, the weighting factor 3 is assigned to the speaker, and if the voice distribution degree is 1, the weighting factor is 1.

As described above, it is a general matter to obtain the inclination of speech of each speaker using a covariance matrix, and detailed description thereof will be omitted.

Then, the weight of each speaker is multiplied by a separate eigenvalue so that the weights obtained from each speaker have the same value.

&Quot; (3) "

For example, if one speaker has a weight of 4, two speakers have a weight of 2, and three speakers have a weight of 1, one speaker has a 0.25 eigenvalue and two speakers have a eigenvalue of 3 Gives 1 speaker an eigenvalue.

Here, the weights assigned to the respective speakers are given a large weight when the speech distribution is large, but the larger the weight is, the more the voice of the speaker is not uniformly distributed and therefore the voice characteristics of the speaker are not properly recognized. This is to refer to the speaker with the lowest voice distribution.

In addition, the low voice distribution means that a certain voice can be obtained from the speaker. (S30)

In addition, after obtaining the change amount of the principal component eigenvalues according to the speech distribution of each speaker as described above, the speech recognition of each speaker is performed according to the following equation.

&Quot; (4) "

위 식에서, 위쪽 조건은 화자 i와 화자 j가 동일한 화자일 경우이고, 아래 쪽 조건은 그렇지 않은 경우이다, 식 중 W는 행렬. W_ij는 행렬 W의 i번째 열, j번째의 칼럼이다. 즉, 화자가 n명 있을 때 i=1~n, j=1~n이 될 수 있고, i=j이면 같은 화자로부터 발화된 문장이며 이땐 2-4와 같은 가중치를 준다. 만약 i≠j이면 (화자가 다르면) 0 값을 준다.

In the above equation, the upper condition is when speaker i and speaker j are the same speaker, and the lower condition is not, where W is a matrix. W _ij is the i-th column and the j-th column of the matrix W. That is, when there are n speakers, i = 1 ~ n, j = 1 ~ n can be, if i = j is a sentence uttered from the same speaker, this time gives a weight equal to 2-4. If i ≠ j (if the speakers are different) it gives a value of zero.

This is used to extract a basic voice through a number of training sessions from each individual speaker, and the same applies to any speaker storing the basic voice and performing voice recognition (S40).

In this way, the individual weights of each speaker and the amount of change in principal component eigenvalues are obtained, and the eigenvalue changes obtained above are matched with the extracted speech information to each speaker, thereby more precisely expressing the speech distribution from each speaker. By distinguishing, it is possible to optimize the speech of the speaker by recognizing the change of the voice according to the change of the internal environment of the speaker.

In addition, it can be used to train the speaker's voice to basic data or to perform voice recognition on the basic data.

1 is a graph illustrating a speech distribution diagram of extracting speech of speakers in a general manner;

2 is a graph weighted to the speaker's speech distribution.

Figure 3 is a flow diagram illustrating a speech recognition error rate reduction algorithm for the voice change of the same speaker according to the present invention.

Claims (5)

An analysis step (S10) of acquiring a plurality of speeches from a plurality of speakers, classifying the speeches according to the respective speakers, and analyzing a speech distribution diagram of the speakers using a support vector machine model; If the speech distribution is large according to the speech distribution of each speaker analyzed by the support vector machine model in the analysis step (S10), the weight is given a large value, and if the speech distribution is small, the weight is assigned a small value. A weighting step (S20) of obtaining a slope according to the distribution; Calculate the variation of the principal component eigenvalues for each speaker by calculating the weighted values according to the voice distribution of each speaker and applying the eigenvalues to each speaker so that the weights assigned to each speaker are all the same. Change amount calculation step (S30); Reduction of the speech recognition error rate for the speech change of the same speaker characterized in that the application step (S40) of applying the change in the principal component eigenvalue for each speaker obtained in the individual variation calculation step (S30) Speaker recognition method using algorithm. The method of claim 1, wherein the analyzing step obtains a direction vector occupied by the samples of each speaker according to the eigenvalue analysis of the speech distribution of each speaker, and then grasps the distance between the samples with respect to the direction vector and slopes of the mean. Speaker recognition method using a speech recognition error rate reduction algorithm for the voice change of the same speaker characterized by obtaining. The speech distribution diagram of the speech spoken by the respective speakers in the analyzing step is

Where s is the speaker index, k is the speech index spoken from the speakers, m _sk is the GMM spervector (spatial distribution) of each utterance, μ is the average vector value of each class (sessions belonging to one speaker), n _s is the number of sessions held by the s-th speaker, μ _s is the average sample value of the s-th speaker, parentheses are the covariance matrix of the samples, and U _s is the eigen vevtor matrix. Where _s is the eigen value matrix and λ is the eigenvalue.) A speaker recognition method using a voice recognition error rate reduction algorithm for the voice change of the same speaker characterized by being obtained by. The method according to claim 1, wherein the change amount (α _s ) of the principal component eigenvalue of each speaker in the individual change amount calculation step is

Where s is the speaker index, k is the speech index spoken from the speakers, λ is the eigenvalue,

Is the average of each eigenvalue (λ), which is the sum of the eigenvalues of the entire class.

being.) A speaker recognition method using a voice recognition error rate reduction algorithm for the voice change of the same speaker characterized by being obtained by. The method according to claim 1, wherein the applying step is the change of the voice of each speaker through the weight associated with the change amount of the principal component eigenvalue

(In the above equation, the upper condition is when speaker i and speaker j are the same speaker, and the lower condition is not, where s is the speaker index, k is the speech index spoken from the speakers,

Is the sum of the eigenvalues of the entire class to which each eigenvalue (λ) is derived, and W is the matrix. W _ij is the i th column and j th column of the matrix W.) Speaker recognition method using a speech recognition error rate reduction algorithm for the speech change of the same speaker characterized by being applied to.

Images