KR101463955B1 - Blind source extraction method using direction of arrival information and de-mixing system therefor - Google Patents

Abstract

A method for extracting a blind signal using direction information according to the present invention is a method for extracting a signal (y) from a sound source located in a specific direction in a frequency domain, the method comprising: extracting a signal y from two or more sound sources The signals are mixed and received through the signal receiver; And demixing the received signal using the non-Gaussian function characteristic of the received signal, using the direction information.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a blind signal extracting method using direction information and a demixing system for the blind signal extracting method.

The present invention relates to a blind signal extracting method using direction information and a demixing system therefor. More particularly, the present invention relates to a blind signal extracting method using direction information, and more particularly to a method and apparatus for extracting a signal from a sound source located in a specific direction among mixed signals received by a signal receiving unit in a frequency domain To a blind signal extracting algorithm.

When receiving a signal such as voice, the signal may be a mixed signal of signals originating from two or more different sources. Therefore, it is necessary to separate or extract only the signal from the desired source among the mixed signals from these two or more sources. Blind Source Separation (BSS) and Blind Source Extraction (BSE) algorithms are known as methods for this purpose.

The blind signal separation method separates signals from two or more sources and separately obtains signals from each source. However, the blind signal separation method results in the separation of signals from undesired sources, e.g., noise. This increases the amount of unnecessary calculation and complicates circuit configuration. On the other hand, the blind signal extraction method is a method of extracting signals from a desired signal from a mixed signal. An algorithm for blind signal extraction in the time domain has already been proposed, but it has a drawback that the computation time is very long and the computation time is long.

The blind signal extraction algorithm in the frequency domain is not active. When the blind signal separation is performed in the frequency domain, a permutation problem occurs in order to align the signals separated in different frequency domains. In order to eliminate such permutation problems, pairing is performed using signals separated in different frequency domains. However, in the case of blind signal extraction, only one signal is extracted, so that a matching operation for eliminating a permutation problem can not be performed.

Accordingly, there is a need for a technique capable of effectively extracting a desired signal by performing an implicit signal extraction algorithm without causing a permutation problem in the frequency domain.

Korean Laid-Open Publication No. 10-2008-0019879 (2008.03.05) U.S. Patent Publication No. 7,797,153B2 (September 14, 2010)

SUMMARY OF THE INVENTION The present invention has been conceived to solve the problems of the prior art, and it is an object of the present invention to provide a technique capable of quickly extracting a signal from a sound source in a specific direction in a mixed signal by performing blind signal extraction without generating a permutation problem in a frequency domain And to provide the above objects.

The technical objects to be achieved by the present invention are not limited to the above-mentioned technical problems, and other technical subjects which are not mentioned can be clearly understood by those skilled in the art from the description of the present invention .

A method for extracting a blind signal using direction information according to an aspect of the present invention is a method for extracting a signal y from a sound source located in a specific direction in a frequency domain, Receiving signals from a sound source through a signal receiving unit; And demixing the received signal using information of the direction using the non-Gaussian function characteristic of the received signal.

In another aspect of the present invention, the step of demultiplexing the received sound source signal may calculate the transfer function by initializing a transfer function W of the demixing filter using information in the direction.

According to another aspect of the present invention, there is provided a method of extracting a signal y in a frequency domain, comprising: receiving signals from two or more sound sources through two or more signal receivers; And demixing the received signal using a non-Gaussian function characteristic of the received signal, wherein the demixing step includes a step of calculating a local constraint indicating a flatness of a transfer function of the mixing filter in a frequency domain And the value of the signal arrival delay time to the two or more signal receiving units is adaptively updated as the learning is repeatedly performed .

According to another aspect of the present invention, a demixing system for blind signal extraction using direction information is a demixing system for extracting a signal y from a sound source located in a specific direction in a frequency domain, A signal receiving unit in which signals from two or more sound sources including a plurality of sound sources are mixed and received; And a de-mixing filter using the non-Gaussian function characteristic of the received signal, wherein the de-mixing filter demultiplexes the received signal using information of the direction.

A demixing system for blind signal extraction using direction information according to another aspect of the present invention includes a filter parameter calculation unit for initializing a transfer function (W) of the demixing filter using the information in the direction and calculating the transfer function .

A demixing system for extracting blind signals according to another aspect of the present invention is a demixing system for extracting one signal (y) in a frequency domain and includes two or more signals A receiving unit; A demixing filter for demultiplexing the received signal using a non-Gaussian function characteristic of the received signal; And a filter parameter calculator for repeatedly performing learning to calculate a transfer function (W) of the demixing filter using a near-field constraint expressing the flatness of the transfer function of the mixing filter in the frequency domain, The value of the signal arrival delay time τ to the two or more signal receiving units can be adaptively updated.

According to the present invention, it is possible to provide a technique of extracting a signal from a sound source located in a specific direction in a mixed signal by performing blind signal extraction without generating a permutation problem in the frequency domain. Also, according to the present invention, it is possible to improve the convergence speed of the blind signal extraction algorithm and reduce the calculation amount.

1 illustrates an environment of a demixing system in which a blind signal extraction algorithm according to an embodiment of the present invention may be implemented.
FIG. 2 shows a room impulse response according to the location of a sound source from a signal receiver in order to find a local constraint according to an embodiment of the present invention.
3 shows the flatness of the signal in the frequency domain according to the distance between the sound source and the signal receiving unit.
FIG. 4 illustrates a geographic positional relationship between a sound source and a signal receiving unit to which a blind signal extraction algorithm according to an embodiment of the present invention can be applied.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a detailed description of preferred embodiments of the present invention will be given with reference to the accompanying drawings. The shape and the size of the elements in the drawings may be exaggerated for clarity of explanation and the same reference numerals are used for the same elements and the same elements are denoted by the same quote symbols as possible even if they are displayed on different drawings Should be. In the following description, well-known functions or constructions are not described in detail to avoid unnecessarily obscuring the subject matter of the present invention.

1 illustrates an environment of a demixing system in which a blind signal extraction algorithm according to an embodiment of the present invention may be implemented. As shown in FIG. 1, it can be considered that a sound source signal is mixed from two or more sources 10 and 12 and received by one or more signal receiving units 20 and 22. 1 illustrates an indoor environment. Therefore, the signals from the sources 10 and 12 not only reach the signal receiving units 20 and 22 through the direct paths D11, D12, D21 and D22, , R22). The sound source signal thus received may be input to the demixing system 30. [ The de-mixing performed through the demixing system 30 may separate the received sound source signals. Hereinafter, the demixing system 30 may be referred to as a concept including the signal receiving units 20 and 22. [

At this time, a state in which there is no information on the sound source signal or mixed environment is referred to as an "implicit" state. That is, an embodiment of the present invention provides an algorithm for extracting signals received in an implicit state.

Extract random signals from mixed signals

Hereinafter, the blind signal extraction algorithm introduced in U.S. Patent No. 7,797,153 B2, which can be used in the embodiment of the present invention, will be described. The blind signal extraction algorithm introduced in U.S. Patent Publication No. 7,797,153B2 shows that each of the mixed signals is stochastically independent of one another while the other frequency bins within one signal are stochastically dependent on each other . When such a blind signal extraction algorithm is used, the permutation problem (permutation problem) can be solved at the same time as the mixed signal is separated.

More specifically, Independent Component Analysis (ICA) is a Blind Source Separation (BSS) algorithm that utilizes statistical independence between output signals. Frequency Domain Independent Component Analysis (FDICA) has been used for the convolutive BSS algorithm because it allows the convolution of the convolutional signal in the time domain in the frequency domain to an instantaneous mixture Because it can be modeled. This modeling can simplify the separation problem. FDCIA successfully separates the signal components of each frequency channel. However, the random permutation problem of separated frequency components between frequency bins occurs.

The blind signal extraction algorithm introduced in U.S. Patent No. 7,797,153 B2 corresponds to a multivariate extension of the independent component analysis and solves the permutation uncertainty problem by using dependency between frequency components. A more detailed description can be found in U.S. Patent Publication No. 7,797,153B2.

First, the mixed signal received in the demixing system 30 can be expressed in the frequency domain through a short-time Fourier transform (STFT). The signal y extracted by the demixing filter (not shown) in the demixing system 30 should be the same as the sound source signal from the sources 10 and 12. Therefore, by multiplying the sound source signal from the first source 11, 12 by the transfer function A (transfer function for the mixing filter) with respect to the path of the sound source signal and multiplying the result by the transfer function W of the demixing filter, The source signal from the sources 11 and 12 must be restored. This is expressed as a matrix.

_수학식(1)

_{Equation (1)}

Here, W _ij (z) will representation of the transfer function for the input j and the output i of the de-mixing filter (not shown) contained in the de-mixing system 30 in the z- domain, A _ij (z) is the source j Domain to the signal receiving unit i in the z-domain.

The signal y extracted in the embodiment of the present invention can be expressed by using a multivariate probability density function proposed by the following equation (2).

_수학식(2)

_{Equation (2)}

Here, y ^f is the output signal of the f-th frequency, and y = [y ¹ , y ^f , ..., y ^K ]. Further, K represents the number of all frequency bins included in the signal y, and? ^F is the standard deviation of the absolute value of the fth frequency signal, and may be set to, for example,

In order to demix the received signal to the demixing system 30, it is necessary to calculate a transfer function W for demixing in the demixing filter. For example, the parameters of the transfer function W can be obtained in a filter parameter calculation unit (not shown) in the following manner.

In order to maximize the non-Gaussianity of a signal in the blind signal extraction algorithm introduced in U.S. Patent No. 7,797,153B2 according to an embodiment of the present invention, a Negentropy function is used as a cost function Lt; / RTI >

수학식(3)

Equation (3)

Where y ^G is a Gaussian function signal with mean and variance equal to y .

The following learning rule can be obtained in order to obtain an optimized extracted signal using the cost function expressed by the above-mentioned equation (3).

_수학식(4)

_{Equation (4)}

Here,? Represents a learning rate.

At this time, the following equation (5) can be obtained by differentiating the equation (3).

수학식(5)

Equation (5)

In the embodiment of the present invention, the learning rule represented by the equation (4) can be determined using the equation (5). Here, x ^f represents a signal of the f-th frequency received at the signal receiving units 20 and 22 and input to the demixing system 30.

In the embodiment of the present invention, the filter parameter calculator (not shown) included in the demixing system 30 obtains a transfer function W for demixing using the cost function expressed by Equation (3) . Thereafter, the signal received by the demixing system 30 is demixed in the demixing filter using the calculated transfer function W.

At this time, the filter parameter calculator receives the output from the demixing filter, and on the basis of the filter parameter calculator, it can repeatedly obtain filter parameters according to the learning rule expressed by Equation (4) and supply it to the demixing filter. Thus, the demixing filter can operate adaptively. In other words, the transfer function W can be adaptively obtained by repeatedly performing an operation in accordance with the learning rule of the equation (4). Thereafter, it is determined whether or not the transfer function W converges. If not, the flow returns to the previous step to calculate the transfer function W again and perform demixing.

Thus, the signal y can be extracted using the transfer function W obtained in an adaptive manner. Which can then be transformed to be represented in the time domain as needed. The signal y extracted according to the algorithm described above may be any of the signals received through the signal receiving units 20 and 22. Generally, the extracted signal y is the largest signal among the signals received through the signal receiving units 20 and 22 in the signal receiving units 20 and 22.

Extracting the signal from the sound source closest to the signal receiver

By extracting the signal y by adding a specific condition to the blind signal extraction algorithm as described above, it is possible to extract the signal generated from the nearest one from the signal receiving units 20 and 22. Hereinafter, a blind signal extraction algorithm capable of extracting a near signal according to an embodiment of the present invention will be described.

First, the characteristics of the signals from the sources 10 and 12 located in the short distance from the signal receiving units 20 and 22 are examined.

FIG. 2 shows a room impulse response according to a location of a sound source from a signal receiver in order to find a local constraint according to an embodiment of the present invention. In Fig. 2, the source of the sound source is indicated by "source" and the signal receiver is indicated by "mic". In FIG. 2, in the graph for each of the sources 1 to 5, the horizontal axis represents the time axis representing the time delay and the vertical axis represents the size of the response.

As can be seen from the room impulse response for each of the sources 1 to 5 in FIG. 2, the closer the distance from the source to the signal receiver (microphone) is, the smaller the magnitude of the first component of the room impulse response is, As shown in FIG. In addition, as the distance from the sound source to the signal receiver (microphone) increases, the difference between the size of the first component of the room impulse response and the size of the remaining components decreases.

It can be seen from FIG. 2 that the room impulse response is similar to the delta function when the distance between the sound source and the signal receiving unit is extremely close. When a room impulse response close to the delta function is expressed in the frequency domain, the signal has a constant value for all the frequency domains. Accordingly, the flatness that is set such that the value increases as the room impulse response has a constant value in the frequency domain can be expressed as Equation (6) below.

수학식(6)

Equation (6)

Where a _ij ^f is the transfer function of the fth frequency domain of the path from source j to signal receiver i.

3 shows the flatness of the signal in the frequency domain according to the distance between the sound source and the signal receiving unit. In FIG. 3, different colors are displayed depending on the degree of echo of the room. 3, the abscissa is a distance axis indicating the distance between the source and the microphone, and the ordinate is a magnitude axis indicating the magnitude of the flatness.

As can be seen from FIG. 3, as the source j is closer to the signal receiving unit i, the size of the transfer function becomes constant in the frequency domain, and therefore, the flatness value defined according to Equation (6) becomes larger. This flatness is calculated using the transfer function a _ij ^f from the source to the signal receiving unit.

In order to extract the near signal in the blind signal extraction algorithm according to the embodiment of the present invention, the above-mentioned characteristic should be applied when calculating the transfer function of the demix filter. In the embodiment of the present invention, it is possible to calculate the transfer function W of the demixing filter by adding the near limit constraint to the learning rule expressed by Equation (4) in the blind signal extraction algorithm.

Hereinafter, the near limit constraint condition will be discussed.

Equation (1) described herein is expressed for a blind signal extraction of 2 x 2 (2 source signals, 2 microphones). This can be expressed by the following equation (7) as an expression for the blind signal extraction for N signals and N microphones.

수학식(7)

Equation (7)

Here, const. Represents a constant value. Therefore, the relation between the demixing filter and the mixing filter can be obtained as shown in equation (8) using equation (7).

수학식(8)

Equation (8)

Equation (8) can be transformed into a simpler form by applying the following two assumptions for obtaining the near limit constraint. The transfer function a _ij ^f corresponding to the mixing filter can be simply modeled using the following two assumptions so as not to violate the object of extracting the signal closest to the signal receiving units 20 and 22.

1. The distance between two neighbors is very close.

2. Since the distance between the nearest signal and the two receivers is very short, the room impulse response, which is a transfer function from the nearest signal to each receiver, is similar to the delta function, ie the first component is very large and the remaining components are very small It will be a form. If we ignore the remaining components except the first component, we can simply express the room impulse responses a _i1 ^f and a _k1 ^f in the two receiving units as shown in Equation (9) with a difference in time delay.

_수학식(9)

_{Equation (9)}

Where τ _l is the closest sound from the receiver l The sound is passed by means of time-consuming and, _{L l} is the distance, v from the closest source to the receiver l denotes the speed of sound. Therefore, by substituting this relationship into the equation (8) for one transfer function a ₁₁ ^f , the following equation (10) can be obtained.

수학식(10)

Equation (10)

Using the above-mentioned assumption 2 once more, the above equation (10) can be expressed in a simpler form. That is, the transfer function a ₁₁ ^f of the mixing filter can also be treated as a constant because of its short distance to the source. Therefore, the above equation (10) can be expressed by the following equation (11).

수학식(11)

Equation (11)

That is, the characteristics of the demixing filter having a signal close to the signal receiving unit through the above two assumptions using the characteristics of the near-field signal can be modeled as in Equation (11). Here, by introducing the flatness expressed by the equation (6), a near constraint expressed by the following equation (12) can be obtained.

수학식(12)

Equation (12)

이고,here,

ego,

이며, 그리고

And

로 표현될 수 있다.

. ≪ / RTI >

이다. 수학식(12)에서 획득되는 Jc가 본 발명의 실시예에 따라 근거리 신호를 추출하기 위한 주파수 영역에서의 근거리 제약 조건이 되는 것이다. 이러한 근거리 제약 조건을 수학식(4)로 표현된 학습 규칙에 적용하여 학습하기 위해서는 Jc의 미분값을 구할 필요가 있다. Jc의 미분값은 다음과 같이 수학식(13)로 표현이 가능하다.Also,

to be. Jc obtained in the equation (12) becomes a local constraint in the frequency domain for extracting the near-field signal according to the embodiment of the present invention. To apply such a local constraint to the learning rule expressed by equation (4), it is necessary to obtain the derivative value of Jc. The differential value of Jc can be expressed by the following equation (13).

_{수학식(13)}

_{Equation (13)}

Using the equation (13), a learning rule for extracting a near signal as in the following equation (14) can be obtained by adding a near constraint in an algorithm for extracting a signal represented by equation (4) have.

수학식(14)

Equation (14)

이고,

이고, 그리고

이다. here,

ego,

And

to be.

이고, η는 학습률, λ는 근거리 제약조건의 가중치로써 임의의 값을 정해두고 사용한다. η와 λ의 값이 너무 크면 수학식(14)로 표시된 학습 규칙이 발산하여 학습에 실패할 수 있다. 하지만, 이들 값이 너무 작은 경우에는 학습하는데 장시간이 소요된다. 일반적으로 η와 λ의 값은, 여러 번의 시행착오를 통해 학습 규칙이 수렴하는 경우 중 가장 큰 값으로 설정될 수 있다. 안정성을 위해서 상기 가장 큰 값의 70~80%에 해당하는 값으로 설정될 수도 있다. J_Neg는 네겐트로피 함수를 이용하여 임의의 신호 하나를 추출하는 부분이며 J_C는 근거리 제약조건(Closeness Constraint)을 나타낸다. here

, Η is the learning rate, and λ is an arbitrary value as the weight of the near constraint. If the values of [eta] and [lambda] are too large, the learning rule indicated by equation (14) may diverge and fail to learn. However, if these values are too small, it takes a long time to learn. In general, the values of η and λ can be set to the largest value among the cases where the learning rule converges through several trial and error. And may be set to a value corresponding to 70 to 80% of the largest value for stability. J _Neg is the part that extracts one arbitrary signal by using the nagent trophy function and J _C represents the close constraint (Closeness Constraint).

As described above, by using the transfer function W calculated by applying the learning rule including the local constraint condition, for example, the learning rule expressed by the equation (14), the blind signal extraction algorithm according to the embodiment of the present invention The signals closest to the signal receiving units 20 and 22 can be extracted.

The function Jc for the near limit constraint includes the delay time? As a variable. So far, the value of the delay time (tau) has been fixed to a value of 0 or a specific value to execute the blind signal extraction algorithm. When extracting the closest signal from the signal receiving units 20 and 22 through the blind signal extracting algorithm according to the embodiment of the present invention by adding the near limit constraint, the delay time which is a time difference between signals arriving at the plurality of signal receiving units τ), instead of using a fixed value, the value can be calculated and updated adaptively. That is, the value of the delay time? Can be adaptively updated for each learning of the blind signal extraction algorithm according to the embodiment of the present invention. In this way, the blind signal extraction algorithm according to the embodiment of the present invention can converge more quickly by adaptively updating the delay time (τ) for each learning of the blind signal extraction algorithm. That is, it is possible to predict the direction of a sound source located nearby from the signal receiving units 20 and 22 without inputting a value of the delay time? From outside, and more efficient learning is possible. Also, through this, the amount of calculation can be reduced.

FIG. 4 illustrates a geographic positional relationship between a sound source and a signal receiving unit to which a blind signal extraction algorithm according to an embodiment of the present invention can be applied. In FIG. 4, a first microphone (Microphone 1) and a second microphone (Microphone 2) are shown as a signal receiving unit and a first sound source 1 is shown as a sound source.

The value of the delay time?, Which is a time difference between the arrival of the signal at the plurality of signal receiving units 20 and 22, is closely related to the direction in which the sound source is located. That is, in FIG. 4, the maximum value of the delay time (.tau.), Which is the time difference between the first sound source and the second microphone, is determined by the distance between the two microphones.

In Fig. 4, the abscissa axis is defined as x axis and the ordinate axis is defined as y axis. If the first sound source is located on the x-axis, the value of the delay time (tau), which is the time delay taken for the signal from the first sound source to reach the first microphone and the second microphone, becomes the maximum. The value of the delay time (tau) at this time is referred to as T and represents the maximum delay time. If the first sound source is located on the y-axis, the value of the delay time (tau), which is the time delay taken for the signal from the first sound source to reach the first microphone and the second microphone, becomes zero. Therefore, it can be seen that the value of the delay time? Has a value of -T to T.

At this time, the value of the function Jc which is a local constraint can be obtained according to the value of the delay time?. For example, the value of the delay time? May be obtained at an appropriate resolution, for example, 0.1 second intervals. This delay time? And the function Jc value may be plotted. For example, the horizontal axis represents the delay time?, The vertical axis represents the value of the function Jc, which is the near limit constraint condition, and the relationship between the two values can be plotted graphically.

The learning objective of the blind signal extraction algorithm using the near limit constraint is to find the transfer function (W) of the demixing filter that maximizes the function Jc value. Again, we can find the value of the delay time (τ) that maximizes the value of the function Jc. In this way, by adding the local constraint condition and adaptively updating the delay time? For each learning of the blind signal extraction algorithm according to the embodiment of the present invention, the blind signal extraction algorithm according to the embodiment of the present invention is faster Can be converged. Also, through this, the amount of calculation can be reduced.

Extraction of signals from a sound source located in a specific direction

In the case where the near limit constraint is not added, an arbitrary signal such as a signal having the largest size can be extracted by using the blind signal extraction algorithm according to the embodiment of the present invention.

According to the blind signal extraction algorithm according to the embodiment of the present invention, it is possible to extract a signal from a sound source located in a specific direction in a mixed signal.

That is, signals from the plurality of sound sources 10 and 12 are mixed and received in the signal receiving units 20 and 22. At this time, it is possible to extract a signal from a sound source located in a specific direction among the signals received by the signal receiving units 20 and 22.

Information in a desired direction can be used to extract signals from the sound sources 10 and 12 located in a specific direction in relation to the arrangement of the signal receiving units 20 and 22. [ Here, the direction information may refer to a relative direction of the sound source to the signal receiving units 20 and 22. [ For example, when the sound source is located on the y-axis of FIG. 4, the direction of the sound source may be referred to as being located in the 0 degree direction. In addition, when the sound source is located on the x-axis of FIG. 4 and is located on the right or left side of the first and second microphones, the direction of the sound source may be referred to as being located in the +90 degrees or -90 degrees direction.

Hereinafter, a blind signal extraction algorithm capable of extracting a signal from a sound source located in a specific direction will be described according to an embodiment of the present invention.

The transfer function from the sound sources 10 and 12 to the signal receiving units 20 and 22, that is, the transfer function of the mixing filter, according to the positions of the sound sources 10 and 12 with respect to the signal receiving units 20 and 22, 15). ≪ / RTI >

_{수학식(15)}

_{Equation (15)}

Here, τ _kl is the time required from the sound signal to the receiver k l there is a voice transmission, l _kl is the distance between the sound source and the signal receiver l k, v is the speed of sound. The information about the direction of the sound source appears as a phase difference in the transfer function (A) of each micro from one sound source when two microphones are used, for example. That is, the relation between the two transfer functions through the transfer function (A) of the mixing filter for the m-th microphone and the n-th microphone is expressed as Equation (16) below.

_{수학식(16)}

_{Equation (16)}

에 대하여 정리하면 아래 수학식(17)을 구할 수 있다.Using the equation (15) from the equation (8) representing the relationship between the demixing filter and the mixing filter,

The following equation (17) can be obtained.

_{수학식(17)}

_{Equation (17)}

As can be seen from the above equation (17), the coefficients of the transfer function W of the demixing filter and the coefficients of the transfer function A of the mixing filter are respectively bundled. Here, the coefficients of the mixing filter can be expressed as a phase difference. When the coefficients of the demixing filter are initialized, the phase difference generated in the mixing filter can be compensated based on the direction information. Thus, by initializing the transfer function W of the demixing filter, a signal from a sound source located in a specific direction according to the direction information can be extracted.

Initializing the transfer function W of the demixing filter can be expressed by the following equation (18).

_{수학식(18)}

_{Equation (18)}

으로 초기화될 수 있다. Where τ _ml is the time it takes for the signal to travel from the sound source 1 to the signal receiver m to be extracted. This value (τ _ml ) is a value inputted according to the direction information of the sound source l. τ _dmn is the delay time value that occurs because the distances to which signals are transmitted from the sound source 1 to be extracted to the signal receiving unit m and the signal receiving unit n are different. In the case of applying the blind signal extraction algorithm according to the embodiment of the present invention by initializing the transfer function W of the demixing filter as in Equation (17), it is possible to extract a signal from a sound source located in a desired direction have. For reference, in the blind signal extraction algorithm when direction information is not provided, for example, the transfer function W

Lt; / RTI >

The blind signal extraction algorithm using the direction information can also be used together with a technique of adding a near constraint and / or a technique of adaptively updating the delay time (tau). For example, when there is an error in the direction information of the sound source, the blind signal extraction algorithm can be adapted while compensating the direction information of the erroneous sound source by updating the delay time (tau) for each learning of the blind signal extraction algorithm.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. will be. Therefore, it should be understood that the above-described embodiments are to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than the foregoing description, It is intended that all changes and modifications derived from the equivalent concept be included within the scope of the present invention.

10, 12: source
20, 22: Signal receiving section
30: De-mixing system

Claims (12)

A method for extracting a signal (y) from a sound source located in a specific direction in a frequency domain,
Receiving signals from two or more sound sources including a sound source located in the specific direction through a signal receiving unit; And
And demixing the received signal using information of the direction using the non-Gaussian function of the received signal,
Wherein the demixing comprises:
Cost function

(W) of the demixing filter using the learning rule

Repeatedly calculating the transfer function according to the transfer function; And
And demixing the received signal using the transfer function,
here,

And? Is a learning rate. The method according to claim 1,
Wherein the demixing comprises:
Wherein the transfer function (W) of the demixing filter is initialized by using the direction information to calculate the transfer function. delete A method for extracting a signal (y) from a sound source located in a specific direction in a frequency domain,
The signals from two or more sound sources are mixed and received through two or more signal receivers; And
And demixing the received signal using a non-Gaussian function characteristic of the received signal,
The demixing step repeatedly performs learning to calculate a transfer function (W) of a demixing filter using a near-field constraint representing a flatness of a transfer function of a mixing filter in a frequency domain,
And the value of the signal arrival delay time to the two or more signal receiving units is adaptively updated as the learning is repeatedly performed. 5. The method of claim 4,
Wherein the demixing comprises:
Cost function

(W) of the demixing filter using the learning rule

Calculating the transfer function of the demixing filter repeatedly according to the transfer function of the demixing filter; And
And demixing the received signal using a transfer function of the demixing filter,
here,

, Jc is the near limit constraint, η is the learning rate, and λ is the weight of the near limit constraint. 6. The method of claim 5,
The near constraint

Wherein the first and second signals are represented by the following equation: A demixing system for extracting a signal (y) from a sound source located in a specific direction in a frequency domain,
A signal receiving unit in which signals from two or more sound sources including a sound source located in the specific direction are mixed and received; And
And a demixing filter for demultiplexing the received signal using the non-Gaussian characteristic of the received signal,
Wherein the filter parameter calculator comprises:
Cost function

(W) of the demixing filter using the learning rule

The transfer function is repeatedly calculated in accordance with the transfer function,
here,

And? Is a learning rate. The demixing system for blind signal extraction using direction information. 8. The method of claim 7,
Further comprising a filter parameter calculator for calculating the transfer function by initializing a transfer function (W) of the demixing filter using information in the direction. delete A demixing system for extracting a signal (y) from a sound source located in a specific direction in a frequency domain,
Two or more signal receiving units in which signals from two or more sound sources are mixed and received;
A demixing filter for demultiplexing the received signal using a non-Gaussian function characteristic of the received signal; And
And a filter parameter calculator that repeatedly performs learning to calculate a transfer function (W) of the demixing filter using a near-field constraint expressing the flatness of the transfer function of the mixing filter in the frequency domain,
And the value of the signal arrival delay time (τ) to the two or more signal receiving units is adaptively updated as the learning is repeatedly performed. 11. The method of claim 10,
Wherein the filter parameter calculator comprises:
Cost function

(W) of the demixing filter using the learning rule

The transfer function of the demixing filter is repeatedly calculated according to Equation
here,

, Jc is the near-field constraint,? Is the learning rate, and? Is the weight of the near-field constraint. 12. The method of claim 11,
The near constraint

Wherein the demixing signal is expressed by the following equation: " (1) "

Images