KR100653173B1 - Multi-channel blind source separation mechanism for solving the permutation ambiguity - Google Patents

Abstract

A method and an apparatus for solving the permutation ambiguity of a separation coefficient of a multi-channel mixed signal are provided to prevent the mixing of different sound source signals, thereby accurately separating an original signal from the mixed signal. An ICA(Independent Component Analysis) input matrix is created in a row vector type of a back-mixing matrix corresponding to a separation coefficient(401). A base vector matrix and a deflection coefficient matrix are extracted using the created ICA method(402). A base vector index value regarding a column vector of a mixing vector is selected by comparing the extracted base vector matrix and deflection coefficient matrix with each other(403). A corresponding clustering including the column vector of the mixing vector is determined based on the selected base vector index value(404).

Description

Method for solving exchange ambiguity of multipath mixed signal separation coefficient and its apparatus {Multi-channel blind source separation mechanism for solving the permutation ambiguity}

1 is a configuration diagram of an embodiment of a multipath mixed signal separation system to which the present invention is applied;

FIG. 2 is a diagram illustrating an embodiment in which an exchange ambiguity of a separation coefficient is generated when an original signal is separated from a mixed signal collected in an echoic real environment; FIG.

Figure 3 is an embodiment configuration diagram for the separation coefficient exchange ambiguity resolution device according to the present invention,

4 is a flow chart of an embodiment of a separation coefficient exchange ambiguity resolution method according to the present invention;

5 is an embodiment graph for showing the direction of the column vector distribution of the frequency response derived using the independent component analysis method in the present invention,

6A and 6B are graphs of one embodiment for showing that the exchange ambiguity of the separation coefficient is solved by applying the present invention.

* Explanation of symbols for the main parts of the drawings

41: matrix generator 42: independent component analysis unit

43: base vector index value selection unit 44: matrix clustering determination unit

The present invention relates to a method and a device for solving the exchange ambiguity of a multipath mixed signal separation coefficient, and more particularly, to a mixed signal collected in a reflective real environment without any prior information on a sound source and a recording environment. In the case of separating the original signal of, the independent component analysis method (ICA) is used to estimate the inherent direction of the separation coefficient in the frequency domain in order to solve the permutation ambiguity. A method and apparatus for solving the exchange ambiguity of a multipath mixed signal separation coefficient for clustering each separation coefficient on the basis of directionality so that each original signal is aligned.

Blind signal separation is a technique of extracting an original signal from a mixed signal, and aims to extract the original signal from the mixed signal without information on the mixed environment and the original signal as blinds mean. Recently, blind signal separation has been spotlighted in various fields such as acoustic / image processing, biosignal analysis, and communication.

This blind signal separation has traditionally been based on the assumption that "mutual signals are independent of each other" and thus the solution to the problem, such as the assumption of signal independence, is subject to significant constraints. A method for finding blind signal separation using a second order statistics of mixed signals is being studied under the assumption that "they are temporally correlated". That is, the solution of blind signal separation is to estimate the mixed circuit to obtain the separated circuit, and to recover the original signal through the separated circuit.

In general, the signals collected in the real environment are modeled by a double mixing circuit. In this double mixing circuit, due to the long impulse response of the system to be modeled, the parallel circuit coefficient (hereinafter referred to as "separation coefficient") is used. Since there is a problem that the convergence of the above) causes a time delay and an unstable convergence, a method of converting and estimating coefficients of the multiplication integration circuit into the frequency domain is mainly used.

As described above, the method of estimating the separation coefficient in the frequency domain is to describe the fundamental ambiguity of the linear mixed model, that is, the order (hereinafter, referred to as “exchange”) in parallel when each sound source separation is performed for each frequency band. Permutation and scale ambiguity have a significant impact on separation performance. Therefore, this technique is very likely to mix with other sound sources in separating one sound source because it cannot guarantee that the order and magnitude correspond to the original signal for the separated sound sources based on the linear mixing model obtained for each frequency band. .

Thus, the following methods have recently been proposed to overcome the above problems.

First, [L. Parra and C. Spence, Convolutive blind source separation of nonstationary sources, IEEE Trans. on SAP, pp320-327, May 2000]. The method described in the paper is to gradually update the mixed circuit coefficient by limiting the length of the mixed circuit coefficient. However, in this conventional scheme, when the impulse response of the multi-integral mixed circuit system is long, a significant limit must be placed on the length of the separation circuit coefficient, which causes a problem that the separation performance is remarkably degraded. Resulting in unstable convergence).

Second, N. Murata, S. Ikeda and A. Ziehe, "An Approach to Blind Source Separation Based on Temporal Structure of Speech Signals", Neurocomputing, Vol. 41, Issue 1-4, pp.1-24, October 2001] As described in the paper, the separation of sound sources in adjacent frequency bands is based on the fact that "a particular structure of a signal changes gradually as the frequency changes." Refer to the result to rearrange the signals. However, in this conventional method, if the signal arrangement is made at least once, the wrong arrangement is propagated to a continuous frequency domain, resulting in a wrong signal separation result.

Third, [L. Parra and C. Alvino, Geometric source separation: merging convolutive source separation with geometric beamforming, IEEE Trans. on SAP, Vol. 10, No. 6, pp352-361 September 2002] In the method disclosed in the paper, beam arrival is used to estimate the direction of arrival of the signal and to arrange the signal based on the direction. However, this conventional method has a problem in that beamforming cannot be applied to a high quality signal having a high sampling rate.

Fourth, [H. Sawada, S. Araki, R. Mukai, S. Makino, Blind extraction of a dominant source signal from mixtures of many sources using ICA and time-frequency masking, IEEE International Symposium on Circuits and Systems (ISCAS 2005), pp. 5882-5885, May 2005], which derives geometric information of the mixed environment from the mixing coefficients and clusters the separation coefficients based on this information to rearrange the results of exchange ambiguity. That is, in this conventional method, the normalization is performed by specifying the maximum distance information between the sound source and the sensor for the mixed circuit coefficients, and then clustering is performed using only geometric information of the remaining mixed environment as a result of the normalization. . However, this conventional method requires a relaxed prior knowledge in that only the distance information between the sound source and the sensor is used, but there is a problem that the separation performance is not consistently guaranteed according to the reference sensor selected in the normalization process.

The present invention has been proposed to solve the above problems, in the case of separating the high quality original signal without prior information on the sound source and the recording environment for the mixed signal collected in the real environment with echo, separation coefficient In order to solve the permutation ambiguity, Independent Component Analysis method (ICA) is used to estimate the inherent directionality of the separation coefficient in the frequency domain and cluster each separation coefficient based on this direction. It is an object of the present invention to provide a method and apparatus for solving the exchange ambiguity of a multipath mixed signal separation coefficient, in which each original signal is aligned.

Other objects and advantages of the present invention can be understood by the following description, and will be more clearly understood by the embodiments of the present invention. Also, it will be readily appreciated that the objects and advantages of the present invention may be realized by the means and combinations thereof indicated in the claims.

The method of the present invention for achieving the above object, in the method of solving the exchange ambiguity of the multipath mixed signal separation coefficient, the inverse mixing matrix corresponding to this separation coefficient in response to receiving the estimated separation coefficient for the mixed signal A matrix generation step of generating an independent component analysis input matrix with a row vector of; An independent component analysis step of extracting a base vector matrix and a deflection coefficient matrix from the generated matrix using an independent component analysis method; A base vector index value selection step of comparing the extracted base vector matrix and the deflection coefficient matrix and selecting a base vector index value of a column vector of a mixed matrix corresponding to a transpose of the inverse mixed matrix row vector; And a matrix clustering determining step of determining a corresponding cluster to which the column vector of the mixed matrix belongs based on the selected base vector index value.

On the other hand, the apparatus of the present invention, in the apparatus for solving the exchange ambiguity of the multipath mixed signal separation coefficient, the matrix for generating the independent component analysis input matrix as a row vector of the inverse mixed matrix corresponding to the separation coefficient estimated from the mixed signal Generation unit; An independent component analysis unit for extracting a base vector matrix and a deflection coefficient matrix from the matrix generated by the matrix generator using an independent component analysis method; A base vector index value selecting unit for comparing a base vector matrix and a deflection coefficient matrix extracted by the independent component analysis unit and selecting a base vector index value of a column vector of a mixed matrix corresponding to a transpose of the inverse mixed matrix row vector; And a matrix clustering determining unit for determining a corresponding cluster to which the column vector of the mixed matrix belongs based on the base vector index value selected by the base vector index value selecting unit.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, whereby those skilled in the art may easily implement the technical idea of the present invention. There will be. In addition, in describing the present invention, when it is determined that the detailed description of the known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

A characteristic of the present invention is that in the case of separating the high quality original signal through the separation circuit without mixing information of the sound source and the recording environment, the mixed signal collected in the real environment with the echo, especially the sound and the audio signal, Exchange ambiguity (order ambiguity) is solved by adjusting the ambiguity of the separation coefficient estimated for the signal, for example, between the row vectors of the inverse mixing matrix (that is, between the column vectors of the mixing matrix which is the invertible mixture of the inverse mixing matrix). Here, the description through the well-known technology, the signal collected in a real echo environment (e.g. closed places such as recording studios) has a long impulse response for each frequency, high quality signals such as music Refers to data that requires sampling.

To this end, the present invention estimates the intrinsic directionality of separation coefficients in the frequency domain using an independent component analysis method (ICA) to resolve the separation coefficient exchange ambiguity and clusters each separation coefficient based on this direction. (Clustering) so that each original signal is aligned.

In other words, the present invention estimates the inherent direction of the row vectors of the inverse mixing matrix corresponding to the separation coefficient without any prior information by using independent component analysis, and normalizes the row vectors of the inverse mixing matrix as before. It is characterized by performing clustering without. In particular, the present invention can find its direction even from a signal distributed linearly and non-orthogonal crossing using independent component analysis.

The present invention as described above is performed as a post-processing process in a multipath mixed signal separation system, that is, a separation coefficient during the process of performing a convolutive blind source separation problem from the time domain to the frequency domain. It is performed between the estimating process and the process of using this separation coefficient for filtering the mixed signal.

For example, by using the separation coefficient (hereinafter, referred to in parallel with the "separation filter coefficient Ｗ") in which the exchange ambiguity is eliminated according to the present invention as a mixed signal filtering parameter, each original signal is separated from the original signal. By sorting in the order of, avoiding mixing between different original signals improves separation performance.

Hereinafter, to facilitate the understanding of the present invention, the "separation coefficient exchange ambiguity resolution device" of the present invention will be briefly described with reference to the general multipath mixed signal separation system shown in FIG. With reference to FIG. 2, the "separation coefficient exchange ambiguity" will be described.

1 is a diagram illustrating an embodiment of a multipath mixed signal separation system to which the present invention is applied.

The multipath mixed signal separation system to which the present invention is applied includes a signal mixer 100, a discrete Fourier transform unit 200, a separation coefficient estimation unit 300, an inverse discrete Fourier transform unit 500, and a signal separator 600. The separation coefficient exchange ambiguity eliminating apparatus 400 according to the present invention receives the separation coefficient estimated by the separation coefficient estimating unit 300, solves the exchange ambiguity, and outputs it to the inverse discrete Fourier transform unit 500. .

The components of the known multipath mixed signal separation system shown in FIG. 1 will be briefly described as follows.

The signal mixer 100 mixes original signals input in the multipath from the outside through a sensor such as a microphone, and the discrete Fourier transform unit 200 converts the mixed signals into discrete Fourier transforms to convert the time domain into a frequency domain. The separation coefficient estimator 300 estimates the separation coefficient by updating (updating) the separation coefficient using a learning routine of a known sound source separation algorithm for the mixed signal in the frequency domain, and the inverse discrete Fourier transform unit. In operation 500, the discrete coefficients of the frequency domain are inverse discrete Fourier transform to convert the frequency domain to the time domain, and the signal separation unit 600 filters the mixed signal using the separation coefficient of the time domain to filter the original signal. Separate.

In general, a major component in a known multipath mixed signal separation system may be referred to as a separation coefficient estimator 300. In the present invention, learning of a known sound source separation algorithm used in such a separation coefficient estimator 300 is performed. Any algorithm [eg; Relative optimization, gradient method, trust-region method, least square estimation method, relative gradient method, relative trust domain method -region method, etc.] may be used.

On the other hand, prior to explaining the separation coefficient exchange ambiguity eliminating apparatus 400 of the present invention, to solve the real-world sound source separation problem modeled as a signal defined in the blind source separation, that is, instantaneous linear mixing model for each frequency domain Introducing the signals used is as follows.

을 Ｍ개의 마이크로폰을 사용하여 녹음한 혼합신호를

이라고 정의한 상태에서, 시간 지연과 다중경로 효과를 고려하 면 ｉ번째 마이크로폰의 출력신호

를 다음의 [수학식 1]과 같이 나타낼 수 있다[모델링].N original signals [source sound source]

Mixed signals recorded using M microphones

Is defined as, the output signal of the i-th microphone, considering the time delay and the multipath effect

Can be expressed as Equation 1 below [modeling].

는 ｊ번째 원신호와 i번째 마이크로폰간의 녹음환경 특성을 나타내는 룸 임펄스 응답(room impulse response)이고, P는 채널의 길이를 나타낸다.here,

Is a room impulse response representing the recording environment characteristic between the i-th original signal and the i-th microphone, and P is the length of the channel.

In the state in which Equation 2 is assumed below, Equation 1 can be summarized as Equation 3 below with reference to Equation 2.

은 시간 이동 연산자이다.here,

Is a time shift operator.

When the above Equation 3 is arranged in a matrix form, it is as shown in Equation 4 below, and it can be seen that it is an output signal of the signal mixer 100 shown in FIG. 1.

As mentioned above, an approach to solve the blind signal separation problem is to convert the mixed signal in the time domain represented by Equation 4 into the mixed signal in the frequency domain, thereby providing an instantaneous mixed signal in each frequency domain. mixture is a solution to the problem of separation.

That is, the discrete Fourier transform unit 200 transforms the time domain into the frequency domain by performing discrete Fourier transform on the mixed signal as shown in [Equation 4]. For example, if the frame size T of the Discrete Fourier Transform (DFT) is much larger than the length P of the channel, the linear convolution is a circular convolution as shown in Equation 5 below. Approximated. Equation 5 below converts Equation 4 to the frequency domain.

(mixing matrix)의 가역인 역혼합 행렬(demixing matrix)의 추정치, 즉 분리계수

를 도출한다. 이를 통해, 혼합 행렬

의 열벡터

는 역혼합 행렬

의 행벡터

와 서로 대응됨을 알 수 있다. 다음의 [수학식 6]은 혼합 행렬과 역혼합 행렬간의 관계를 나타낸다.In addition, the separation coefficient estimator 300 to which a specific blind signal separation algorithm is applied based on the frequency domain mentioned through Equation 5 above may be used for each frequency bin. Per matrix

estimate of the demixing matrix, the reversible of the mixing matrix

To derive This allows the mixing matrix

Column vector

Is the inverse mixing matrix

Row of

And correspond to each other. Equation 6 below shows a relationship between a mixing matrix and an inverse mixing matrix.

에는 크기 모호성(scaling ambiguities) 및 교환 모호성(permutation ambiguities)이 존재한다. 여기서, 역혼합 행렬의 크기 모호성은 주지 기술인 음원분리알고리즘으로도 쉽게 해소되는 바, 이하 본 발명 에서는 분리계수 크기 모호성에 대해서는 언급하지 않기로 한다.On the other hand, as described through the prior art inverse mixing matrix as shown in [Equation 6] estimated by the separation coefficient estimation unit 300

There are scaling ambiguities and permutation ambiguities. Here, the size ambiguity of the inverse mixing matrix is easily solved by a sound source separation algorithm, which is a well-known technique. Hereinafter, the separation coefficient size ambiguity will not be described.

는 이산 푸리에 변환(또는 단시간 푸리에 변환; short time fourier transform 등)을 거쳐 주파수 빈별 혼합신호

를 생성한다. 이에, 분리계수 추정부(300)에서는 이 주파수 빈별 혼합신호를 입력받아 각 주파수 빈에 대응되는 역혼합 행렬의 추정치

를 도출한다.That is, mixed signals collected through multipath

Is a mixed signal by frequency bin through a discrete Fourier transform (or short time fourier transform, etc.)

Create Therefore, the separation coefficient estimator 300 receives the mixed signal for each frequency bin and estimates an inverse mixing matrix corresponding to each frequency bin.

To derive

의 행벡터

는 다음의 [수학식 7]과 같이 교환 행렬(permutation matrix)에 의해 모호성이 발생될 수 있다. 예컨대, "동일한 색인값 i를 갖는 두 주파수 영역의 행벡터

가 실제적으로 서로 다른 음원에 해당되는 역혼합 벡터일 수도 있다"라는 것이 분리계수 교환 모호성에 대한 정의이다.By the way, inverse mixing matrix

Row of

Ambiguity may be generated by a permutation matrix as shown in Equation 7 below. For example, "row vectors of two frequency domains with the same index value i.

May actually be an inverse mixing vector corresponding to different sound sources ".

Here, P represents an exchange matrix.

는 교환 모호성이 해소되지 않은 분리계수를 의미하며[공지의 음원분리알고리즘으로 추정한 분리계수],

는 본 발명에 의해 교환 모호성이 해소된 분리계수를 의미하는 것으로 설명한다. 이와 같은 분리계수 교환 모호성에 대해서 도 2를 참조하여 좀 더 상세히 살펴보기로 한다.1 and as described above

Means the separation coefficient whose exchange ambiguity has not been resolved [separation coefficient estimated by the known sound source separation algorithm],

Explains that the separation coefficient solved the exchange ambiguity by the present invention. This separation coefficient exchange ambiguity will be described in more detail with reference to FIG. 2.

FIG. 2 is a diagram illustrating an embodiment in which an exchange ambiguity of a separation coefficient is generated when an original signal is separated from a mixed signal collected in an echoic real environment.

가 2개의 센서를 통해 수집된 경우에, 2개의 혼합신호

는 각각 이산 푸리에 변환을 거쳐 L개의 주파수 빈별 혼합신호

와

로 변환된다.As shown in FIG. 2, two sound source signals

Is mixed through two sensors, two mixed signals

Is a mixed signal for each of L frequency bins through a discrete Fourier transform.

Wow

Is converted to.

Therefore, through the L source repetition algorithm through the L source repetition algorithm, the instantaneous sound source separation problem is solved for each frequency bin. As a result, a separation coefficient which is an estimate of the sound source signal in two frequency domains is generated.

However, since it is impossible to know in what order the sound source signals in the frequency domain, which are the final products, will be generated, if one sound source signal is estimated by simply integrating these results, signals in specific frequency domains of different sound source signals may be mixed with each other. Will be generated.

의 교환 모호성을 해소, 즉 역혼합 행렬을 이루는 행벡터[다시 말하면 역혼합 행렬의 가역인 혼합 행렬을 이루는 열벡터]간의 위치를 서로 조정하여 분리계수 교환 모호성을 해소한다. 이러한 본 발명에서 제시하는 분리계수 교환 모호성 해소 매커니즘에 대해서는 도 3, 도 4, 도 5, 도 6a 및 도 6b를 참조하여 보다 상세히 후술하기로 한다.Therefore, in the present invention, the inverse mixing matrix estimated by the known separation coefficient estimator 300 is known.

The separation ambiguity is solved by adjusting the position between the row vectors constituting the inverse mixing matrix (that is, the column vectors constituting the mixing matrix that is the invertible mixture of the inverse mixing matrix). The separation coefficient exchange ambiguity-resolving mechanism proposed in the present invention will be described later in more detail with reference to FIGS. 3, 4, 5, 6A, and 6B.

를 공지의 역이산 푸리에 변환부(500) 및 공지의 신호 분리부(600)에서 사용하여 다음의 [수학식 8]과 같이 혼합신호로부터 원신호

를 분리할 수 있다.Estimation of Inverse Mixing Matrix with Exchange Ambiguity Eliminated by the Present Invention

Is used in the known inverse discrete Fourier transform unit 500 and the known signal separation unit 600, the original signal from the mixed signal as shown in Equation 8 below.

Can be separated.

Figure 3 is a block diagram of an embodiment of the separation coefficient exchange ambiguity solving apparatus according to the present invention, Figure 4 is a flow chart of an embodiment of the separation coefficient exchange ambiguity solving method according to the present invention. Hereinafter, the present invention will be described with reference to FIGS. 3 and 4 for the convenience of description.

The separation coefficient exchange ambiguity elimination apparatus according to the present invention includes a matrix generator 41, an independent component analyzer 42, a base vector index value selector 43, and a matrix clustering determiner 44.

로 독립성분분석법의 입력 파라미터로 사용될 행렬 X를 생성한다(401). 여기서, 행렬 X는 역혼합 행렬의 행벡터

에 대한 전치 행렬(transpose matrix)로 구성된다.The matrix generator 41 is a row vector of an inverse mixing matrix in all frequency domains in which the separation coefficient estimated by the separation coefficient estimator 300, for example, the exchange ambiguity is generated.

In operation 401, a matrix X to be used as an input parameter of the independent component analysis method is generated. Where matrix X is the row vector of the inverse mixing matrix

It consists of a transpose matrix for.

The independent component analyzing unit 42 extracts the matrix A of the basis vector and the matrix S of the deflection coefficient from the matrix X received from the matrix generator 41 using an independent component analysis algorithm. (402). Here, the results of performing the independent component analysis algorithm used in the present invention will be described later with reference to FIG. 5.

Processes "401" and "402" through the equations for the matrix X generated by the aforementioned matrix generator 41 and the base vector matrix A and the deflection coefficient matrix S extracted by the independent component analyzer 42. When described in more detail as follows.

Equation 9 below shows the relationship between the matrix X, the base vector matrix A, and the deflection coefficient matrix S.

이며, 독립성분분석법이 사용되어져 추출된 기초벡터

는 선형적으로 분포한 열벡터의 방향성을 나타내고, 편향계수 S의 요소들은 주어진 열벡터의 포인트가 해당되는 기초벡터에 어느 정도로 편향되어 있는지를 나타내는 계수이다.Where matrix X is the transpose of all the row vectors of the inverse mixture matrix

Basic vector extracted by independent component analysis

Denotes the direction of linearly distributed column vectors, and the elements of the deflection coefficient S are coefficients indicating how much a point of a given column vector is deflected to the corresponding base vector.

에 대응되는, 예컨대 역혼합 행렬의 행벡터의 전치에 해당되는 혼합 행렬의 열벡터

에 대해 편향계수값이 가장 큰 기초벡터 색인값 j를 선정한다(403).The base vector index value selecting unit 43 uses the base vector matrix A and the deflection coefficient matrix S extracted by the independent component analyzing unit 42 as parameters, and the row vectors of the inverse mixing matrix in all frequency domains.

Column vectors of the mixing matrix that correspond to, for example, transpose of the row vectors of the inverse mixing matrix.

The base vector index value j having the largest deflection coefficient value is selected with respect to (403).

가 실제로 어떠한 색인값을 가지는 열로 교환되어야 되는지 알기 위해서

의 해당 계수값들을 조사한다. 이는, 해당 편향계수 행렬 Ｓ의 열벡터

에 대해

의 각 요소들

중 가장 큰 절대값을 가지는 계수값에 대해서 해당 색인값을 추출하는 것으로서, 계수값이 더 큰 기초벡터에 편향되어 있다고 판정하여 이를 기반으로 군집화를 수행한다. 다시 말하면, 여기서 선정된 기초벡터 색인값 j는 이후의 행렬 군집화 결정 과정에서 열벡터

간의 위치 조정 파라미터로 사용된다.That is, the column vector of the mixing matrix

To know which index value should actually be exchanged for

Examine the corresponding coefficients for. This is the column vector of the deflection coefficient matrix S

About

Elements of

The index value is extracted with respect to the coefficient value having the largest absolute value, and it is determined that the coefficient value is biased to the larger base vector, and clustering is performed based on the index value. In other words, the base vector index value j selected here is a column vector in a subsequent matrix clustering decision process.

Used as a parameter for adjusting the position of the liver.

의 해당 계수값이

와 같은 부등식을 만족하는 경우에, 상기

는 첫번째[1st] 기초벡터에 편향되어 있다고 판정할 수 있다.For example, the column vector of a mixing matrix

The corresponding count of

If the following inequality is satisfied,

Can be determined to be biased against the first [1st] basis vector.

가 속하는 해당 군집을 결정한다(404).The matrix clustering determining unit 44 uses the base vector index value j selected by the base vector index value selecting unit 43 as a parameter to obtain a column vector of the mixed matrix.

Determine the corresponding cluster to which it belongs (404).

가 행렬 내에서 실제로 위치해야 되는 자리를 나타내는 색인값이며, 이에 행렬 군집화 결정 과정에서 열벡터

가 행렬 내에서 다른 열벡터간에 서로 교환되어 있다면 기초벡터 색인값 j를 토대로 원래의 자리로 서로 위치 조정을 해 주어 교환 모호성을 해소시킨다.As explained earlier through the process of selecting the base vector index value, the base vector index value j is a column vector.

Is an index value that represents the position of the actual position in the matrix, and thus the column vector during the matrix clustering decision.

If is exchanged between different column vectors in a matrix, the positional ambiguity is eliminated by adjusting the position of each other based on the base vector index value j.

[i>L*N]에 대해서 수행하여 교환 모호성이 해소된

를 출력한다.All column vectors in matrix X for "403" and "404"

exchange ambiguity is resolved by performing [i> L * N].

Outputs

FIG. 5 is an exemplary graph for showing the direction of a column vector distribution of a frequency response derived using the independent component analysis method of the present invention.

를 나타낸다[N=2인 경우].5 and 2 are independent component vectors which are the result of performing independent component analysis on a sound source collected by a reverberant recording studio, such as a microphone. Column vector of frequency response

[In the case of N = 2].

As can be seen from FIG. 5, the column vectors of the frequency response form an intersecting linear distribution, and the direction of the linear distribution of the column vectors of each frequency response is based on geometric information between the sensor and the sound source. In addition, the column vector of the frequency response is linearly distributed not only by the geometric distribution information between the sensor and the sound source but also by the amount of change by frequency. As such, the present invention finds the intrinsic directionality of the data, ie, the column vector of the frequency response, in particular the direction of the column vector distributed linearly and non-orthogonally, using independent component analysis without prior information on the recording environment. The linear distribution directionality of these column vectors is used as the column vector clustering parameter.

6A and 6B are graphs of one embodiment for showing that the exchange ambiguity of the separation coefficient is solved by applying the present invention.

FIG. 6A is a graph showing a column vector distribution of a frequency response in which a separation coefficient exchange ambiguity is generated, in which column vectors of frequency responses having the same index value intersect with each other without forming a linear distribution are exchanged between frequency responses. It belongs to all distributions.

FIG. 6B is a graph showing a column vector distribution of a frequency response in which the separation coefficient exchange ambiguity is solved according to the present invention, and the column vectors of each frequency response belong to one linear distribution by adjusting positions of the column vectors of the frequency response. have.

As described above, the method of the present invention may be implemented as a program and stored in a recording medium (CD-ROM, ROM, RAM, floppy disk, hard disk, magneto-optical disk, etc.) in a computer-readable form. Since this process can be easily implemented by those skilled in the art will not be described in more detail.

The present invention described above is capable of various substitutions, modifications, and changes without departing from the technical spirit of the present invention for those skilled in the art to which the present invention pertains. It is not limited by the drawings.

The present invention as described above has the effect of accurately separating the original signal from the mixed signal by eliminating the separation coefficient exchange ambiguity, which is a bottleneck of separation performance in blind signal separation, to prevent mixing between different sound source signals.

In addition, the present invention is a sound source and recording for the mixed signal collected in the real environment when a signal that requires a high quality sampling rate, such as music, when a long impulse response due to the reverberation characteristics of the closed recording environment is used In the case where there is no prior information on the environment, there is an effect that the signal separation can be performed stably without degrading performance.

Claims (12)

In the method of solving the exchange ambiguity of the multipath mixed signal separation coefficient, A matrix generation step of generating an independent component analysis input matrix based on a row vector of an inverse mixing matrix corresponding to the separation coefficient according to the input of the separation coefficient estimated for the mixed signal; An independent component analysis step of extracting a base vector matrix and a deflection coefficient matrix from the generated matrix using an independent component analysis method; A base vector index value selection step of comparing the extracted base vector matrix and the deflection coefficient matrix and selecting a base vector index value of a column vector of a mixed matrix corresponding to a transpose of the inverse mixed matrix row vector; And A matrix clustering determination step of determining a corresponding cluster to which the column vector of the mixed matrix belongs based on the selected base vector index value. Method of solving the exchange ambiguity of the multipath mixed signal separation coefficient comprising a. The method of claim 1, The independent component analysis input matrix, 12. A method for solving the exchange ambiguity of a multipath mixed signal separation coefficient, characterized in that it consists of a transpose matrix for the row vectors of the inverse mixing matrix corresponding to the estimated separation coefficient. The method of claim 1, The independent component analysis input matrix, the base vector matrix, and the deflection coefficient matrix are A method for solving the exchange ambiguity of a multipath mixed signal separation coefficient, characterized by having a relationship as in the following equation. [Equation]

Where X is the independent component analysis input matrix, A is the base vector matrix, and S is the deflection coefficient matrix. The method of claim 3, wherein The base vector matrix A is A method for solving the exchange ambiguity of a multipath mixed signal separation coefficient, characterized by the directionality of the column vector distribution with respect to the linearly distributed frequency response. The method of claim 3, wherein The deflection coefficient matrix S is And how the elements indicate how much the point of the column vector for the frequency response is biased in the corresponding base vector. The method of claim 1, The row vector of the inverse mixing matrix is A method for solving the exchange ambiguity of a multipath mixed signal separation coefficient, characterized by being distributed linearly and non-orthogonally in the frequency domain. The method of claim 1, The row vector of the inverse mixing matrix is A method for solving the exchange ambiguity of a multipath mixed signal separation coefficient, characterized in that it has a corresponding linear distribution direction in the frequency domain according to the geometric arrangement between the signal collection sensor and the original signal. The method according to any one of claims 1 to 7, The basic vector index value selection step, The exchange ambiguity of multipath mixed signal separation coefficients is selected by selecting a base vector index value corresponding to a coefficient value having the largest absolute value among the elements of the deflection coefficient matrix column vector with respect to the column vector of the mixed matrix. How to solve. The method according to any one of claims 1 to 7, The matrix clustering determination step, And resolving the interchange ambiguity of the multipath mixed signal separation coefficient according to the selected base vector index value. The method of claim 9, Separating the original signal by filtering the mixed signal using a separation coefficient corresponding to the column vector of the mixing matrix determined in the matrix clustering determination step Method of eliminating the exchange ambiguity of the multipath mixed signal separation coefficient further comprising. A device for solving the exchange ambiguity of a multipath mixed signal separation coefficient, A matrix generator for generating an independent component analysis input matrix as a row vector of an inverse mixing matrix corresponding to the separation coefficient estimated from the mixed signal; An independent component analysis unit for extracting a base vector matrix and a deflection coefficient matrix from the matrix generated by the matrix generator using an independent component analysis method; A base vector index value selection unit for comparing a base vector matrix and a deflection coefficient matrix extracted by the independent component analysis unit and selecting a base vector index value of a column vector of a mixed matrix corresponding to a transpose of the inverse mixed matrix row vector; And A matrix clustering determining unit for determining a corresponding cluster to which the column vector of the mixed matrix belongs based on the base vector index value selected by the base vector index value selecting unit. Apparatus for solving the exchange ambiguity of the multipath mixed signal separation coefficient comprising a. The method of claim 11, Signal separation unit for separating the original signal by filtering the mixed signal using the separation coefficient corresponding to the column vector of the mixing matrix determined by the matrix clustering determination unit Apparatus for solving the exchange ambiguity of the multipath mixed signal separation coefficient further comprising.

Images