JP2008060635A - Blind signal extracting device, method thereof, program thereof, and recording medium stored with this program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To extract a target signal from a mixed signal using a beam former even if a user does not have prior knowledge of an impulse response vector of the target signal. <P>SOLUTION: In observing signals emitted from N-sets of signal sources using M-sets of sensors 4m to extract one signal or more (where N, M are each an integer of not less than 2), the observation signal x<SB>m</SB>(t) observed with the sensor 4m is converted into a signal x<SB>n</SB>(f, τ) of a frequency region (5), x<SB>n</SB>(f, τ) is normalized (22), the normalized x<SB>n</SB>(f, τ) is clustered to N-cluster C<SB>n</SB>(24), a correlation matrix R<SP>n</SP><SB>J</SB>(f) of the observation signal including only the unwanted signal is estimated from the C<SB>n</SB>(25), the beam former w<SB>n</SB>(f) is calculated from the cluster information C<SB>n</SB>and the R<SP>n</SP><SB>J</SB>(f) (28), a target signal y<SB>n</SB>(f, τ) is extracted from x<SB>n</SB>(f, τ) using the w<SB>n</SB>(f) (30), and the y<SB>n</SB>(f, τ) is converted into a signal of a time region (32). <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

    In the present invention, it is not possible to directly observe only a necessary source signal (target signal), but the target signal is estimated and extracted in a situation where other noise, interference signal, etc. are observed superimposed on the target signal. The present invention relates to a blind signal extraction device, a method, a program, and a recording medium.

Here, the observation signal is first modeled and the frequency domain of the signal is defined, and then the prior art is briefly described.
[Observation signal]
All signals are sampled at a certain sampling frequency f _s and expressed discretely. It is assumed that N (N is an integer of 2 or more) source signals are mixed and observed by M (M is an integer of 2 or more) sensors. The present invention deals with a situation in which a signal is attenuated / delayed depending on a distance from a signal generation source to a sensor, or a signal is reflected by a wall or the like, thereby causing transmission path distortion. In such a situation, the source signals s _n (t) (n = 1,..., N) from a plurality of signal sources are observed by a plurality of sensors to be observed signals x _m (t) (m = 1,. , M), and the impulse response from each signal source n to sensor m is h _mn (u) (u represents time). The observation signal x _m (t) at the sensor m is convolutionally mixed with the corresponding impulse response h _mn (u) for each source signal s _n (t), and is expressed by the following equation.
x _m (t) = Σ _{n = 1} ^N Σ _{u = 0} ^∞ h _mn (u) s _n (tu) (1)
Here, the source signals s ₁ (t),. . . , S _N (t) and impulse response h ₁₁ (u),. . . , H _1N (u),. . . , H _M1 (u),. . . Consider a situation where information about h _MN (u) cannot be obtained in advance. In this situation, the observation signals x ₁ (t),. . . , X _M (t) only and source signals s ₁ (t),. . . , S _N (t) is a broad object of this invention.
[Frequency domain expression]
In the present invention, each operation is performed in the frequency domain. Therefore, a time series x _m (f, τ) for each frequency by applying a short-time Fourier transform, for example, a known technique to the L point (L is an arbitrary integer) on the observation signal x _m (t) of the sensor. ) = Σ _{u = −L / 2} ^{L / 2-1} x _m (τ + u) g (u) e ^−i2πfu (2)
Ask for. Where f is the frequency, f = 0, f _s / L,. . . , F _s (L−1) / L, and τ is an arbitrary time, and as described above, f _s is a sampling frequency. g (u) is a window function such as a Hanning window.

In the frequency domain, the convolutional mixing in the time domain expressed by Equation (1) is
x _m (f, τ) = Σ _{n = 1} ^N h _mn (f) s _n (f, τ) (3)
And an approximate representation of simple mixing at each frequency. Here, h _mn (f) is a frequency response (impulse response) for the frequency component f from the signal source n to the sensor m, and s _n (f, τ) is a source signal s according to an equation similar to equation (2). _n (t) is subjected to a short-time Fourier transform, and so on. Observation signals x ₁ (f, τ),. . . , X _M (f, τ) as a vector using equation (3),
x (f, τ) = Σ _{n = 1} ^N h _n (f) s _n (f, τ) (4)
It becomes. Here, x (f, τ) is x (f, τ) = [x ₁ (f, τ),. . . , X _m (f, τ),. . . , X _M (f, τ)] is an observed signal vector ^T , and h _n (f) is h _n (f) = [h _1n (f),. . . , H _mn (f),. . . , H _Mn (f)] ^T , a vector summarizing the frequency response from the signal source to each sensor. [A] ^T indicates a transposed vector of the vector A. The same applies to the following description.
[Representative conventional technology]
As a typical signal extraction method for extracting a target signal from a mixed signal, an adaptive beamformer (ABF) is described in Non-Patent Document 1 and widely known.

An example of the functional configuration of a conventional adaptive beamformer (hereinafter referred to as a conventional beamformer) is shown in FIG. Signals x _m (t) observed by the plurality of sensors 4 _m are input to the frequency domain transform unit 5. The signal x _m (t) is converted into the frequency domain signal x _m (f, τ) by the frequency domain converter 5. x _m (f, τ) (m = 1,..., M) are all input to the conventional beamformer 6.

The conventional beamformer 6 emphasizes the target signal s _n (t) in the system using a plurality of sensors, and unnecessary signals s ₁ (t),. . . , S _n-1 (t),. . . , S _{n + 1} (t),. . . , S _N (t) as much as possible filter w _n (f) = [w _1n (f),. . . , W _mn (f),. . . , W _Mn (f)] ^T is realized.

When designing the conventional beamformer 6, "impulse response vector _h n (f) from the target signal source to each sensor or a steering vector _{a n} is the approximation (f) = [exp (-i2πfτ 1n , ..., exp (-i2πfτ _Mn )] ^T (5)
Is known ". Here, τ _mn is the difference between the time when the signal source n reaches the sensor m and the time when it reaches the origin 0. Conventionally, often using a sensor system arranged in a straight line as shown in FIG. 2, _n the direction of the signal source n _theta, when the sensor 41 of the sensor 4m and coordinates d _m on the basis of above tau _mn Is given by τ _mn = d _m cos θ _n / c. Where c is the speed of the signal.

で得られる。 When Returning to FIG. 1, a conventional beamformer 6 for suppressing undesired signals, the following output power A represented by the formula _{'(w n} (f)) filter group that minimizes the (vector) w _n ( f) is estimated.
A ′ (w _n (f)) = E {| y _n | ² (f, τ)}
= E {y _n (f, τ) y _n ^* (f, τ)}
= E {w _n ^H (f) x (f, τ) x ^H (f, τ) w _n (f)}
= W _n ^H (f) R _x (f) w _n (f) (6)
Here, E {•} is an average operation with respect to time τ, A ^* is a complex conjugate of A, and R _x (f) = E {x (f, τ) x ^H (f, τ)} is a correlation matrix of an observation signal , [A] ^H is a conjugate transpose matrix (vector) of the matrix (vector) A, and y _n (f, τ) is an output of the conventional beamformer 6 and can be expressed by the following equation (7). .
y _n (f, τ) = w _n ^H (f) x (f, τ) (7)
Here, in order to avoid a meaningless solution (w _n (f) = 0 = [0,..., 0] ^T ), the constraint condition shown in the following equation that the target signal can be obtained without distortion is given. Give.
w _n ^H (f) h _n (f) = 1 (8)
As a result, the problem of obtaining the value of w _n (f) that satisfies equation (8) and minimizes the value of A ′ (w _n (f)) in equation (6) uses Lagrange's undetermined multiplier p. And can be represented by the following formula (9).
A (w _n (f)) = A ′ (w _n (f)) + p (w _n ^H (f) h _n (f) −1) (9)
By solving equation (9), the conventional beamformer 6 is

It is obtained by.

In conventional beamformer 6 (conventional adaptive beamformer), the impulse response vector in equation (10) h _{n (f)} the impulse response vector or may be stored by actually measuring the impulse response storage unit 10 h _{n (f} ) Is ideally read and used. However, instead, the steering vector a _n (f) shown in the above equation (5) is stored in the steering vector storage unit 12, and the read steering vector a _n (f) is converted into the impulse response vector h of the above equation (10). It is widely used instead of _n (f).

However, in a real environment, the impulse response vector h _{n (f)} and the steering vector a _{n (f)} is given correctly are rare, minimization of A (w _{n (f))} represented by the above formula (9) In many cases, however, it is not always necessary to minimize only unnecessary signals. Therefore, instead of the correlation matrix R _x (f) of the mixed signal (observation signal), the correlation matrix R _J (f) = E {ξ (f) of the signal ξ (f, τ) in the time interval of only the unnecessary signal. , Τ) ξ ^H (f, τ)} is widely used in Non-Patent Document 2 and the like. This is known to achieve higher performance than using R _x (f). That is, in the conventional beamformer, it is desirable that the correlation matrix can be accurately estimated in the time section of only unnecessary signals (time section in the absence of the target sound).

The conventional beamformer 6 outputs an output signal vector y _n (f, τ) according to the above equation (7) from w _n (f) in the above equation (10) and the observed signal vector x (f, τ). . The output signal y _n (f, τ) is input to the time domain transforming unit 8 and transformed from the frequency domain to the time domain to generate y _n (t).
Haykin, S .; Adaptive Filter Theory Science and Technology Publication 2001, pages 690-693 Toshiro Oga Yoshio Yamazaki Yutaka Kaneda Acoustic system and digital processing The Institute of Electronics, Information and Communication Engineers Corona Corporation pages 190-191

As described above, the conventional beamformer requires the impulse response vector h _n (f) from the target signal source to each sensor or the steering vector a _n (f) that is an approximation thereof. That is, there is a difficulty that prior knowledge about the target signal is necessary. Furthermore, they are difficult to obtain correctly in the actual environment, and the performance of the conventional beamformer is significantly degraded when the prior knowledge and the impulse response vector h _n (f) in the use environment are deviated.
Further, in order to obtain high performance, it is necessary to estimate the correlation matrix R _J (f) of the signal in the time interval of only the unnecessary signal. However, when the unnecessary signal is a non-stationary signal, it is very Have difficulty.

In a signal extraction apparatus for observing signals emitted from N signal sources with M sensors and extracting one or more of the observed signals, where N and M are integers of 2 or more Yes, the observation signals observed by the M sensors are converted into frequency domain signals, the frequency domain signals are normalized, a normalized observation signal vector is calculated, and the normalized observation signal vector Are clustered into the N clusters,
An unnecessary signal correlation matrix that is a correlation matrix of an observation signal including only unnecessary signals is estimated from the cluster information, a beam former is calculated from the cluster information and the unnecessary signal correlation matrix, and the beam former is calculated. The target signal is extracted from the frequency domain signal, and the extracted target signal is converted into a time domain signal.

    With the above configuration, it is possible to accurately extract and extract a target signal without measuring an impulse response vector or a steering vector in advance and even if an unnecessary signal is a non-stationary signal.

  The best mode for carrying out the invention will be described below.

    An example of the functional configuration of the present invention is shown in FIG. 3, and the main processing flow of the present invention is shown in FIG. The same functional components as those in FIG. 1 are denoted by the same reference numerals, and redundant description is omitted. The same applies to the following.

    In the present invention, signal sparsity is assumed. Sparse means that the signal is zero at most times τ. The sparsity of the signal is confirmed by an audio signal, for example. By assuming the sparseness of the signal, it can be assumed that even if there are a plurality of source signals, at each time frequency point (f, τ), they do not overlap each other and there is at most one. That is, the above formula (4) can be expressed by the following formula.

x (f, τ) = h _n (f) s _n (f, τ) (11)
Here, h _n (f) represents an impulse response vector, and s _n (f, τ) represents a source signal existing at (f, τ).

また、ノルム正規化を以下の式で行う。 Each observation signal x _m (t) (m = 1,... M) collected by the sensor 4 m is input to the frequency domain conversion unit 5. Each observation signal x _m (t) is converted from the time domain to the frequency domain by, for example, the above-described short-time Fourier transform, which is a known technique, and converted to x _m (f, τ) by the frequency domain transform unit 5. Conversion is performed (step S2). Further, x _m (f, τ) is output as an observation signal vector x (f, τ). The observation signal vector x (f, τ) is input to the normalization unit 22, and a normalized observation signal vector is calculated (step S4). Specifically, the observed signal vector x (f, τ) = [x ₁ (f, τ),. . . , X _m (f, τ)] For ^T , normalization of the declination is performed by the following equation.

Norm normalization is performed using the following equation.

x ⁻ (f, τ) ← x ⁻ (f, τ) / ‖x ⁻ (f, τ) ‖ (13)
Where x ⁻ (f, τ) represents the normalized observed signal vector x (f, τ), arg (r) represents the declination of r, i represents the imaginary unit, and | r | represents the absolute value of r, ‖r‖ represents the norm of r, Q represents the reference sensor number (Q∈ {1,..., M}), c represents the speed of the signal, α Represents any positive constant. For α, α = 4d _max is most preferred. However, d _max represents the maximum value of the distance between an arbitrary sensor Q selected as a reference and another sensor. Α may be another numerical value.

ここで、Ａ_ｎ＝（Σ_ｍ＝１ ^Ｍ│ｈ_ｍｎ│^２）^１／２であり、信号ｓ_ｎ（ｆ、τ）に関するインパルス応答情報にのみ依存することが分かる。 From the above equations (11) to (13), the normalized observation signal vector x ⁻ (f, τ) can be expressed by the following equation.

Here, it can be seen that A _n = (Σ _{m = 1} ^M | h _mn | ² ) ^1/2 and depends only on the impulse response information regarding the signal s _n (f, τ).

Normalized observation signal vectors x ⁻ (f, τ) of all time frequencies are input to the clustering unit 24 and clustered into N clusters (step S6). This clustering can be effectively performed using, for example, the k-means method. Details are described in “RO Duda, PE Hart, and DG Stock, Pattern Classification, Wiley Interscience, 2nd edition, 2000”. The clustering method will be specifically described below.

In the storage unit 26, the normalized observation signal vector x ⁻ (f, τ) represented by the above formula (14) and the initial centroid value c ^j _n (j = 0, n = 1,..., N ) Is stored.
The clustering unit 24 reads the normalized observation signal vector x ⁻ (f, τ) from the storage unit 26 and clusters them to generate _N clusters C ₁ ,..., CN. That is, the normalized observation signal vector x ⁻ (f, τ), which is an M-dimensional complex vector, is directly clustered in the following procedure in the M-dimensional complex space.

1. The initial value c ^j _n of the cluster centroid is read from the storage unit 26. The centroid initial value c ^j _n is a vector (M-dimensional complex vector) having the same dimension as the normalized observation signal vector x ⁻ (f, τ). Note that how to select the initial value c ⁰ _n of the centroid will be described later.
2. Let j + 1 be a new j.
3. The normalized observation signal vector x ⁻ (f, τ) at all time frequencies (f, τ) is assigned to the cluster C _n represented by the nearest centroid c ^j−1 _n . That is, for each normalized vector x ⁻ (f, τ), n is selected such that ‖x ⁻ (f, τ) −c ^j−1 _n ‖ is the smallest.
4). The average value of the normalized observation signal vector x ⁻ (f, τ) assigned to each cluster C _n is calculated, and the centroid is updated by setting the norm to 1. That is, for the normalized observation signal vector x ⁻ (f, τ) assigned to each cluster C _n ,
^{_{c n j = E {x (}} f, τ)} n / ‖E {x - (f, τ)} n ‖ (15)
The centroid is updated by performing the above calculation. Here, E {·} _n represents an average operation for the members of cluster C _n .
5. Repeat steps 2-5 until centroid c ^j _n converges. The finally converged centroid is stored in the storage unit 26 as c _n (n = 1,..., N). The above is the clustering procedure.

Next, an example of how to select the initial value of the centroid will be described.
<< Initial value setting method 1 >>
N vectors are randomly selected from the normalized observation signal vector x ⁻ (f, τ), and set as the initial value c ⁰ _n (n = 1,..., N) of the centroid.
<< Initial value setting method 2 >>
The centroid can be written as the following equation (16) assuming that | h _mn (f) | = 1 for all m and n in the equations (11) to (15), and this is used.
^{_{{C n} q = E [}} x - (f, τ)] n
_{= Exp [i2π (d m -d} Q) T v n / α] / M 1/2
= Exp [i2π‖d _m -d _Q || _{^{T cosΘ n mQ / α] /}} M 1/2
(16)
Here, _{d m} represents the position vector of the sensor _{4 m, v n = cosΘ n} mQ represents the arrival direction vector of the signal _s n (t) with respect to the axis connecting the selected sensor 4Q as a sensor 4m and a reference, FIG. In FIG. 5, this is indicated by a thick vector. Furthermore, _{v n} is the unit vector, a ‖V _n || = 1.

The sensor position d _m (m = 1,..., M) appropriately gives the value held in the storage unit 26 as the azimuth θ _n and the elevation angle φ _n (n = 1,..., N). . Here, since the sensor position d _m , the azimuth θ _n and the elevation angle φ _n are initial values, they may be appropriate values. For example, when θ _n = 2πn / N and φ _n = 0, initial values dispersed spatially are obtained.

Returning to FIG. 3, the cluster C _n obtained by the clustering unit 24 corresponds to each source signal s _n (f, τ). Further, the centroid ^{_{c - n = E {x -}} (f, τ)} x- (f, τ) ∈Cn , as will be appreciated from the above equation (14), the source signal _s n (f, tau) It can be seen that the impulse response information for is represented.

推定された不要信号相関行列Ｒ^ｎ _Ｊ（ｆ）とクラスタリング部２４よりのクラスタのセントロイド情報ｃ^― _ｎは、ビームフォーマ計算部２８に入力される。ビームフォーマ計算部２８ではクラスタの情報Ｃ_ｎと、不要信号相関行列Ｒ^ｎ _Ｊ（ｆ）とからビームフォーマｗ_ｎ（ｆ）を計算する（ステップＳ１０）。ビームフォーマｗ_ｎ（ｆ）の具体的な計算方法については、実施例２で詳細に説明する。 Each cluster C _n is input to the unnecessary signal correlation matrix estimation unit 25. The unnecessary signal correlation matrix estimator 25 calculates the correlation matrix of the unnecessary signal section with respect to the source signal s _n (f, τ) from the information for each cluster C _n , that is, the correlation matrix of the observation signal including only the unnecessary signal. The unnecessary signal correlation matrix R ⁿ _J (f) is estimated by the following equations (17) and (18) (step S8).

The estimated unnecessary signal correlation matrix R ⁿ _J (f) and the cluster centroid information c ⁻ _n from the clustering unit 24 are input to the beamformer calculation unit 28. The beamformer calculation unit 28 calculates the beamformer w _n (f) from the cluster information C _n and the unnecessary signal correlation matrix R ⁿ _J (f) (step S10). A specific calculation method of the beamformer w _n (f) will be described in detail in the second embodiment.

The calculated beamformer w _n (f) is input to the target signal extraction unit 30. The target signal extraction unit 30 calculates the following equation (19) using the beamformer w _n (f), and calculates the target signal y _n (f, τ) from the frequency domain observation signal x (f, τ). Is extracted (step S12).
y _n (f, τ) = w _n (f) ^H x (f, τ) (19)
By performing the equations (17) to (19) for all n (n = 1,..., N), all N signals are extracted. All y _n (f, τ) are input to the time domain conversion unit 32. The target signal y _n (f, τ) extracted by the target signal extraction unit 30 is converted into a time domain signal y _n (t) by, for example, a short-time inverse Fourier transform which is a known technique (step S14). .

    Next, a second embodiment of the present invention will be described. The second embodiment is an example in which the beamformer calculation unit 28 described in the first embodiment is configured in more detail. FIG. 6 shows a functional configuration example of the beamformer calculation unit 28 according to the second embodiment and parts related thereto. The parts not shown in FIG. 6 perform the same processing as that described in the first embodiment, and the same applies to the following embodiments. The beamformer calculation unit 28 includes an impulse response estimation unit 40 and an adaptive beamformer calculation unit 42.

In the impulse response estimating unit 40, a cluster _{C n} centroid information c ^- _n to estimate the impulse response vector _h n of the target signal (f) from. Specifically, the impulse response vector for the source signal s _n (f, τ) is estimated by denormalizing the centroid information c ⁻ _n from the clustering unit 24.

式（２０）を、ｈ_ｍｎ（ｆ）について解くと、以下の式（２１）を得ることが出来る。 First, the centroid of the cluster C _n is the above equation (15), and x ⁻ (f, τ) can be represented by the equation (14). Here, assuming that | h _mn | = 1 for all m and n, the following equation (20) is established for the m-th component c ⁻ _mn of the centroid c ⁻ _n .

Solving equation (20) with respect to h _mn (f), the following equation (21) can be obtained.

式（２０）の右辺はインパルス応答ｈ_ｍｎ（ｆ）が正規化されたものになっているが、式（２１）は式（２０）から逆にインパルス応答ベクトルｈ_ｎ（ｆ）について求め直しているので、式（２１）は逆正規化と呼ぶ。

The right side of the equation (20) is a normalized impulse response h _mn (f), but the equation (21) is obtained by recalculating the impulse response vector h _n (f) from the equation (20). Therefore, equation (21) is called denormalization.

Returning to the description of FIG. 6, the estimated impulse response vector h _n (f) and R ⁿ _J (f) from the unnecessary signal correlation matrix estimation unit 25 are input to the adaptive beamformer calculation unit 42. .

上記式（２２）は、上記式（１０）のＲ_ｘ（ｆ）をＲ^ｎ _Ｊ（ｆ）に置き換えることで得ることができる。 The adaptive beamformer calculation unit 42 calculates an adaptive beamformer w _n (f) using the impulse response vector h _n (f) and the unnecessary signal correlation matrix R ⁿ _J (f). Specifically, the adaptive beamformer w _n (f) can be calculated by the following equation (22).

The above formula (22) can be obtained by replacing R _x (f) in the above formula (10) with R ⁿ _J (f).

The target signal extraction unit 30 applies the adaptive beamformer w _n (f) in Expression (22) and the signal x (f, τ) in the frequency domain to the above Expression (19), thereby generating the target signal y _n ( f, τ) is extracted.

Next, Example 2 ′ is shown as a modification of Example 2. According to the impulse response estimator 40, instead of outputting the impulse response vector estimation of h _{n (f)} using the above formula (22), to estimate the steering vector a _{n (f),} it is conceivable to output. Signal _s n (f, _τ) the expression steering vectors as well as _{(5) a n (f)} = [exp (-i2πfτ 1n), ···, exp (-i2πfτ Mn)] T (23)
Then, since the steering vector a _n (f) is an estimation of the impulse response vector h _n (f), when the phase terms of the equation (21) and the equation (23) are compared, τ _mn (m = 1, ..., M) can be estimated by the following equation (24).
^{_{τ ^ mn = αc -1 arg [}} c - mn c - Qn] / 2π (24)
The calculation of Expression (24) is performed by the impulse response estimation unit 40 ′.

The steering vector a _n (f) using τ ^ _mn is output as an estimate of the impulse response vector h _n (f). That is, from the above equations (23) and (24), the steering vector a _n (f) can be expressed by the following equation (27), and is output from the impulse response estimation unit 40 ′ as the impulse response h _n (f).
_{a n (f) = [exp} (-i2πfτ ^ 1n), ···, exp (-i2πfτ ^ Mn)] T ≒ h ^ n (f) (25)
An adaptive beamformer is calculated by the above equation (22) from h ^ _n (f) and R ⁿ _{J of} equation (25).

In addition, when the unnecessary signal correlation matrix estimation unit 25 estimates the adaptive beamformer w _n (f) using the above equation (22), the unnecessary signal correlation matrix R ⁿ _J (f) is used to generate an acoustic signal. If the transfer characteristics, that is, the impulse response vector h _n (f) and the steering vector a _n (f) are known, they may be used. Further, instead of the unnecessary signal correlation matrix R ⁿ _J (f) in the equation (22), the correlation matrix R _x (f) of the observation signal may be used. The same applies to the following embodiments.

In the adaptive beamformer shown in the second embodiment, high performance can be obtained when N ≦ M. However, when N> M, the performance is limited. Specifically, the adaptive beamformer can effectively extract the target signal y _n (f, τ) if the number of unnecessary signals is M−1 or less. Is known to be inadequate. Therefore, the third embodiment shows that the target signal can be extracted even when N> M, that is, when there are N−1 (> M−1) unnecessary signals. A functional configuration example of the third embodiment is shown in FIG. In the third embodiment, an unnecessary signal selection unit 49 and an input signal estimation unit 50 are added as compared with the second embodiment, and an unnecessary signal correlation matrix estimation unit 25, an impulse response estimation unit 40, an adaptive beamformer calculation unit 42, The processing of has been changed.

An unnecessary signal selection unit 49 estimates an unnecessary signal correlation matrix R ⁿ _J (f) for K unnecessary signals. Here, K is an integer that satisfies K ≦ M−1. That is, from clusters C _L (L = 1,..., N−1, n + 1,..., N) corresponding to unnecessary signals other than the cluster C _n corresponding to the target signal s _n (f, τ) to K. Choose clusters and let these clusters be C _J.
As the choice of K clusters, in cluster C _L, and a method of selecting the K clusters in order from many cluster member, the power of the following formula (26) represented by xi] _{L (f,} tau) A method of selecting K clusters in order from the largest is conceivable.

Using the K clusters C _J selected by the unnecessary signal selection unit 49, the unnecessary signal correlation matrix R ⁿ _J (f) for the K unnecessary signals is calculated by the following equations (27) and (28).

適応型ビームフォーマ計算部４２では、上記式（２７）（２８）で得られた不要信号相関行列Ｒ^ｎ _Ｊ（ｆ）とインパルス応答推定部４０または４０’よりのインパルス応答ベクトルｈ_ｎ（ｆ）を用いて、上記式（２２）を用いて、適応型ビームフォーマｗ_ｎ（ｆ）を計算する。 Further, the input signal estimation unit 50 estimates a beamformer input signal vector x (f, τ) in which the K unnecessary signals selected by the unnecessary signal selection unit 49 and the target signal C _n are mixed. It uses the undesired signal cluster C _J and the target signal clusters C _n, can be obtained by the following equation (29).

In the adaptive beamformer calculation unit 42, the unnecessary signal correlation matrix R ⁿ _J (f) obtained by the above equations (27) and (28) and the impulse response vector h _n (f) from the impulse response estimation unit 40 or 40 ′. Is used to calculate the adaptive beamformer w _n (f) using the above equation (22).

The target signal extraction unit 30 receives the beamformer input signal x (f, τ) from the input signal estimation unit 50. The target signal extraction unit 30 calculates the above equation (19) using the adaptive beamformer w _n (f) from the adaptive beamformer calculation unit 42 and the beamformer input signal vector x (f, τ). , The target signal y _n (f, τ) is extracted.

    In the third embodiment, the case where N> M is described as described above. However, it can be implemented even when N ≦ M. In this case, compared with the second embodiment, an extra cost is required for the processing of the input signal estimation unit 50 and the unnecessary signal selection unit 49.

  The fourth embodiment is an embodiment where the position information of the sensor is known. FIG. 8 shows a part of a functional configuration example of the fourth embodiment. The impulse response estimation unit 40 described in the second or third embodiment includes an arrival direction estimation unit 60 and an impulse response calculation unit 62. Moreover, it has the sensor position information storage part 64 which has memorize | stored the sensor position information showing the position of M sensors.

The arrival direction estimation unit 60 estimates the arrival direction of the signal. The arrival direction estimation unit 60, the centroid information c than clustering section 24 ^- three-dimensional vector representing the position of each sensor than _n and the sensor position information storage unit _{64 d m (m = 1,} ..., M) Is entered. _V n dimensional vector of length one representing the direction of arrival of the signal _{s n} (n = 1, ..., N) When the estimated value of the arrival direction of the signal _{s n} is the centroid information c ^- with _n And can be calculated by the following equation (30).

v _n = αD ⁺ arg [c ⁻ _n ] / 2π (30)
Here, D = [d ₁ -d _Q ,. . . , _D m _-d Q,. . . , D _M −d _Q ] ^T , where d _Q is a three-dimensional vector representing the position of the sensor 4Q arbitrarily selected as a reference, and D ⁺ represents a generalized inverse matrix of D.

そして、ステアリングベクトルａ_ｎ＾（ｆ）（インパルス応答ベクトルの推定値ｈ_ｎ（ｆ））がインパルス応答計算部６２から出力され、適応型ビームフォーマ計算部４２に入力される。 Next, in the impulse response calculation unit 62, by using the direction of arrival and the sensor position information of the signals s _n, to calculate the impulse response. The impulse response calculation unit 62 receives the arrival direction estimation value q _n of the signal s _n from the arrival direction estimation unit 60 and the sensor position information d _Q from the sensor position information storage unit 64. The impulse response calculator 62 obtains an estimated value of the steering vector a _n ^ (f) for the signal s _n represented by the following equation (31). This steering vector a _n ^ (f) is calculated as an estimated value h _n (f) of the impulse response vector.

The steering vector a _n ^ (f) (impulse response vector estimation value h _n (f)) is output from the impulse response calculation unit 62 and input to the adaptive beamformer calculation unit 42.

In this embodiment, a configuration using a maximum gain beamformer instead of the adaptive beamformer is shown. The maximum gain beamformer is a method in which a filter w _n (f) that minimizes an unnecessary signal component in the sensor array output while maximizing a target signal in the sensor array output is used as a beamformer. (DH Johnson and DE Dudgeon, “Array Signal Processing Concepts and Technologies”, Prentice Hall, 1993.)
In the maximum gain beamformer, it is one point to estimate the target signal component and unnecessary signal component in the sensor array output. However, when the unnecessary signal is a non-stationary signal, it is very difficult to estimate the unnecessary signal. There was a problem that it was difficult. In the fifth embodiment, this problem is solved by using a sparsity assumption. In other words, (1) the target signal correlation matrix R _T ⁿ (f) that is the correlation matrix of the observation signal only of the target signal and the unnecessary signal correlation matrix R _J ⁿ (f) that is the correlation matrix of the observation signal only of the unnecessary signal (2) The maximum gain beamformer calculation unit estimates the maximum gain beamformer w _n (f) from the target signal correlation matrix R _T ⁿ (f) and the unnecessary signal correlation matrix R _J ⁿ (f). Can be solved.

Further, since the maximum gain beamformer does not have the above-mentioned constraint condition (8) of “minimizing distortion of the target signal”, beamformers w _n (f) having various gain characteristics are formed at each frequency f. Is done. This means that, for example, when the maximum gain beamformer is applied to a wideband signal such as an audio signal, the output is distorted by the frequency characteristic of w _n (f). For this reason, conventionally, it has been difficult to use the maximum gain beamformer for wideband signals. In the fifth embodiment, the maximum gain beamformer w _n (f) is corrected so that the error between the observation signal vector x (f, τ) and the output signal of the maximum gain beamformer w _n (f) is minimized. To solve this.

First, the principle of the maximum gain beamformer will be briefly described. As described above, the evaluation function is expressed by the following expression (32) based on the condition that “the target signal at the sensor array output is maximized while the unnecessary signal component at the sensor array output is minimized”.

Here, the denominator is the output power of the unnecessary signal, the numerator is the output power of the target signal, R _T ⁿ (f) is the correlation matrix of the observation signal only of the target signal, and R _J ⁿ (f) is the observation of only the unnecessary signal. It is a correlation matrix of a signal. Also, (R _J ⁿ (f)) ^1/2 = EF ^1/2 E ^H , where E = [e ₁ ,. . . e _m ], e _i is the eigenvector of R _J ⁿ (f), F = diag (λ ₁ ,..., λ _M ), and λ _i is the eigenvalue of R _J ⁿ corresponding to e _i And w ^~ = (R _J ⁿ (f)) ^1/2 w _n , the above equation (32) can be changed to the following equation (33).

Here, from the Rayleigh quotient theorem described in “Kodama, Suda,“ Matrix Theory for System Control, Corona, 1995 ”, the maximum value of g (w ⁻ ) is (R _J ⁿ (f)) ^{−1. _{/ 2 (R T n (f}} )) is given by ^(R J n ^(f)) the maximum eigenvalue of ^-1/2 lambda, when the corresponding eigenvectors and e, the maximum value ^{maxg (w ~) = λ =} g (E). Maximum gain beamformer _{w n} for obtaining That can be expressed by the following equation (34) (35).
w ^~ = e (34)
w _n = (R _J ⁿ (f)) − ^1/2 e (35)
FIG. 9 shows a functional configuration example of the fifth embodiment. Compared with the first embodiment, an observation signal correlation matrix estimation unit 72 is added, and a beamformer calculation unit 28 includes a target signal correlation matrix estimation unit 70, an eigenvector calculation unit 74, a maximum gain beamformer calculation unit 76, and a correction vector calculation. Part 78 and correction part 80.

ここで、Ｃ_ｎは目的信号に対応するクラスタである。不要信号相関行列推定部２５よりの不要信号相関行列Ｒ_Ｊ ^ｎ（ｆ）と、目的信号相関行列Ｒ_Ｔ ^ｎ（ｆ）とが、固有ベクトル計算部７４に入力される。固有ベクトル計算部７４で（Ｒ_Ｊ ^ｎ（ｆ））^−１／２（Ｒ_Ｔ ^ｎ（ｆ））（Ｒ_Ｊ ^ｎ（ｆ））^−１／２の最大固有ベクトルｅ_ｎ（ｆ）を上記で説明したレイリー商の定理より、計算する。 The target signal correlation matrix estimation unit 70 estimates the correlation matrix of the time interval of only the target signal s _n (f, τ) from the cluster information using the following equations (36) and (37).

Here, C _n is a cluster corresponding to the target signal. The unnecessary signal correlation matrix R _J ⁿ (f) and the target signal correlation matrix R _T ⁿ (f) from the unnecessary signal correlation matrix estimation unit 25 are input to the eigenvector calculation unit 74. In the eigenvector calculation unit 74, the maximum eigenvector e _n (f) of (R _J ⁿ (f)) − ^1/2 (R _T ⁿ (f)) (R _J ⁿ (f)) − ^1/2 has been described above. Calculate from the Rayleigh quotient theorem.

R _J ⁿ (f) from the unnecessary signal correlation matrix estimation unit 25 and e _n (f) from the eigenvector calculation unit 74 are input to the maximum gain beamformer calculation unit 76. The maximum gain beamformer calculation unit 76 calculates the maximum gain beamformer w _n (f) from the following equation (38).
_{w n (f) = (R} J n (f)) -1/2 e n (f) (38)
This equation (38) is based on the above equation (35).

On the other hand, the observation signal correlation matrix estimation unit 72 estimates an observation signal correlation matrix R _x (f) that is a correlation matrix of the observation signal vector x (f, τ) using the following equation (39).
R _x (f) = E {x (f, τ) x ^H (f, τ)} (39)
The maximum gain beamformer w _n (f) from the maximum gain beamformer calculation unit 76 and the observation signal correlation matrix R _x (f) from the observation signal correlation matrix estimation unit 72 are input to the correction vector calculation unit 78. The correction vector calculator 78 generates a correction vector α _n (f) for correcting the maximum gain beamformer w _n (f). This correction converts the maximum gain beamformer w _n (f) so that the distortion that the maximum gain beamformer w _n (f) gives to the output is minimized. For example, a correction vector α _n (f) that minimizes an error A between the observed signal vector x (f, τ) and the output signal vector y _n (f, τ) expressed by the following equation (40) is calculated.
_{A (α n (f))} = E {‖x (f, τ) -α n (f) y n (f, τ) ‖ ^2} (40)
Here, y _n (f, τ) is the output y _n (f, τ) = w _n (f) x (f, τ) of the maximum gain beamformer w _n (f).
When the right side of the above equation (40) is expanded,
A (α _n (f)) = {E [‖x (f, τ) ‖]} ² −α _n (f) E [x ^H (f, τ) y _n (f, τ)] − α _n ^H (f) E [y _n (f, τ) ^* x (f, τ)]
^{_{+ Α n α n H E [}} │y n (f, τ) │ 2] (41)
In equation (41), when both sides are partially differentiated by α _n ^H (f), the following equation (42) is obtained.
∂A (α _n (f)) / ∂ α _n ^H (f) =
−E [y _n (f, τ) ^* x (f, τ)] + α _n E [| y _n (f, τ) | ² ] (42)
When the left side of the equation (42) is set to 0 and α _n is obtained, the following equation (43) is obtained.
α _n (f) = E [y _n (f, τ) ^* x (f, τ)] / E [| y _n (f, τ) | ² ]
(43)
Here, from the above equation (19) and the above equation (39), the above equation (43) becomes the following equation (44).

ここで、上述したように、Ｒ_ｘ（ｆ）は観測信号ベクトルｘ（ｆ、τ）の相関行列である。上記式（４４）から理解されるように、最大利得ビームフォーマｗ_ｎ（ｆ）と観測信号ベクトルｘ（ｆ、τ）を用いて、補正ベクトル計算部７８では、補正ベクトルα_ｎ(ｆ)が計算される。

Here, as described above, R _x (f) is a correlation matrix of the observed signal vector x (f, τ). As can be understood from the above equation (44), the correction vector α _n (f) is calculated by the correction vector calculator 78 using the maximum gain beamformer w _n (f) and the observed signal vector x (f, τ). Calculated.

The correction unit 80 corrects the frequency distortion for the maximum gain beamformer w _n (f) using the correction vector α _n (f), and calculates a corrected beamformer. Specifically, the corrected beamformer w _n ′ (f) can be obtained by correction using the following equation (45).
w _n ′ (f) = [α _n (f)] _B w _n (f) (45)
Here, B is the number of an arbitrary sensor, and Bε {1,. . . , M}, and [q] _B indicates the Bth element of the vector q.

The target signal extraction unit 30 extracts the target signal y _n (f, τ) by the following equation (46) using the corrected beam former w _n ′ (f).
y _n (f, τ) = w _n ′ ^H (f) x (f, τ) (46)
FIG. 10 shows a functional configuration example of a modification of the fifth embodiment. The beamformer calculation unit 28 includes an objective signal correlation matrix estimation unit 70, an eigenvector calculation unit 74, and a maximum gain beamformer calculation unit 76, and the target signal extraction unit 30 includes a signal extraction unit 81 and a distortion correction unit 82. Has been.

The maximum gain beamformer w _n (f) from the maximum gain beamformer calculation unit 76 and the observation signal vector x (f, τ) from the frequency domain conversion unit 5 are input to the signal extraction unit 81. The signal extraction unit 81 calculates the following equation (47) to extract the target signal y _n (f, τ) including distortion.
y _n (f, τ) = w _n ^H (f) x (f, τ) (47)
The target signal y _n (f, τ) including distortion is input to the distortion correction unit 82.
The correction vector α _n (f) from the correction vector calculation unit 78 is also input to the distortion correction unit 82. The distortion correction unit 82 converts the output signal according to the following equation (48) to correct the distortion and output a corrected output signal y _n ′ (f, τ).
_{y n '(f, τ)} = [α n (f)] B y n (f, τ) (48)
In the first to fifth embodiments described above, signals are extracted for all n. However, a beam former may be configured only for a single signal (one n). For purposes signal selection, for example, by comparing the impulse response vector h _n, which is estimated for all the sound sources n by the impulse response vector h _d and invention method object signal on the database, the h _n closest to the h _d The sound source n can be selected by selecting it. For example, an algorithm such as min _n (h ₁ · h _n ) is conceivable. An adaptive beamformer for a target signal can be obtained by configuring a beamformer using the above formula (24) by the beamformer calculation unit 28 described in the second to fifth embodiments for only selected n.

[Experimental result]
Experiments were conducted to show the effects of the above examples. The impulse response measured in the room shown in FIG. 11 was convolved with a plurality of sounds and mixed to simulate the mixed signal. The experimental conditions are as shown in FIG. Three sensors 41, 42, and 43 were arranged at positions 200 cm from the long side of the bottom surface and 282 cm from the short side in a room having a long side of 880 cm, a short side of 375 cm, a height of 240 cm, and a reverberation of 120 ms. The long side and the parallel axis are x, the short side and the parallel axis are y, and as shown in FIG. 12, three sensors 41, 42 and 43 are provided on the y axis, one on the x axis, and the length of the side. An experiment is conducted in the case where the vertices are arranged in two dimensions at the apex of a regular triangle of 4 cm in length. A microphone was used as the sensor.
The signal-to-unnecessary signal ratio (SIR) and the signal-to-distortion ratio (SDR) were evaluated for four voice combinations. The unit is dB.

  Each of the four sound sources is centered on the intersection of the x-axis and y-axis at the sensor position, the + direction of the x-axis is 0 degree, 30 degrees counterclockwise, 315 degrees, and the sensor position is centered on a circle with a radius of 50 cm. Each sound source is arranged on the intersection, and the sound source is arranged on the intersection between the directions of 225 degrees and 315 degrees and a circle having a radius of 80 cm. In the experiment for confirming the effect of the second embodiment, sound sources with directions of 120 degrees, 225 degrees, and 315 degrees were used, and N (number of source signals) = M (number of sensors) = 3. In an experiment for confirming the effect of Example 3, N = 4 and M = 3.

FIG. 13 shows the results of this experiment. The third embodiment shows a case where the input signal estimation unit 50 shown in FIG. 7 is provided in the conventional method, the second embodiment, the second embodiment 2 ′, the fourth embodiment, and the fifth embodiment. In the conventional method, the equation (10) representing the adaptive beamformer 6 shown in FIG. 1 is used in which a known steering vector a _n (f) is given to h _n (f). In this case, high performance was not obtained in both cases of N = M and N> M. This is because an experiment in an environment with reverberation, given steering vector a _{n (f)} until the influence of the reverberation, considered as a main cause that could not be taken into account. Also, the fact that sufficient SIR cannot be obtained when N> M indicates that the limit of the adaptive beamformer, that is, only M−1 unnecessary signals can be effectively suppressed.

    Compared to the conventional method, even when N = M and N> M, it is understood that the above embodiment has higher performance than the conventional method when comparing the values of SIR and SDR.

    In addition to the above embodiments, the blind signal extraction device according to the present invention is not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. Further, the processing described in the blind signal extraction device is not only executed in time series according to the order of description, but may also be executed in parallel or individually as required by the processing capability of the device that executes the processing. Good.

  Further, when the processing in the blind signal extraction device of the present invention is realized by a computer, the processing contents of the functions that the blind signal extraction device should have are described by a program. Then, by executing this program on a computer, the processing function of the blind signal extraction device is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape, and the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable-Programmable-Ready), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. A configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by an execution instruction and result acquisition without transferring a program from the server computer to the computer. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

    In this embodiment, the blind signal extraction apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

    As an application in the audio field, the present invention separates target voices even in situations where the microphone picks up sounds other than the target speaker voice because the input microphone of the voice recognizer and the speaker are separated from each other. By extracting, it is possible to construct a speech recognition system with a high recognition rate.

The block diagram which shows the function structural example of the system of a prior art. The figure for demonstrating the time difference (tau) between the time when the sound source n reaches the arbitrary sensors j, and the time when it reaches the origin 0 in the case of using the sensor system arranged in a straight line. 1 is a block diagram showing a functional configuration example of a system according to Embodiment 1 of the present invention. The flowchart which shows the flow of the main processes of Example 1 of this invention. The figure for ^{demonstrating} cos ₍ ^theta ) nmQ used by the said Formula (16) in the sensor m and the sensor Q which are arbitrary two sensors. The block diagram which shows a part of example of a function structure of the system of Example 2 of this invention. The block diagram which shows a part of example of a function structure of the system of Example 3 of this invention. The block diagram which shows a part of example of a function structure of the system of Example 4 of this invention. The block diagram which shows a part of example of a function structure of the system of Example 5 of this invention. The block diagram which shows a part of example of a function structure of the system of the modification of Example 5 of this invention. The figure which looked at the comparison experiment of the prior art and the technique of this invention from right above. The figure which shows the detail of the positional relationship of the three sensors 41, 42, and 43 of FIG. The figure which shows the experimental result which compared the effect of the prior art and the technique of this invention.

Claims (14)

In a signal extraction apparatus for observing signals emitted from N signal sources with M sensors and extracting one or more of the observed signals, where N and M are integers of 2 or more Yes,
A frequency domain conversion unit that converts observation signals observed by the M sensors into a frequency domain signal;
A normalization unit that normalizes the frequency domain signal and calculates a normalized observation signal vector;
A clustering unit for clustering the normalized observation signal vector into N clusters;
An unnecessary signal correlation matrix estimation unit that estimates an unnecessary signal correlation matrix that is a correlation matrix of an observation signal including only unnecessary signals from the information of the cluster,
A beamformer calculation unit for calculating a beamformer from the information on the cluster and the unnecessary signal correlation matrix;
A target signal extraction unit that extracts a target signal from the frequency domain signal using the beamformer;
A blind signal extraction apparatus comprising: a time domain conversion unit that converts the extracted target signal into a time domain signal. The blind signal extraction device according to claim 1,
The beamformer calculation unit
An impulse response estimator for estimating an impulse response of the target signal from the centroid information of the cluster;
An adaptive beamformer calculation unit that calculates an adaptive beamformer using the impulse response and the unnecessary signal correlation matrix,
The blind signal extraction apparatus, wherein the target signal extraction unit extracts the target signal from the frequency domain signal using the adaptive beamformer. The blind signal extraction device according to claim 1,
The unnecessary signal correlation matrix estimation unit estimates an unnecessary signal correlation matrix that is a correlation matrix of K unnecessary signals selected from the cluster information, where K is an integer that satisfies K ≦ M−1. And
And an input signal estimation unit for estimating a beamformer input signal including only the target signal and the selected K unnecessary signals from the cluster information,
The beamformer calculation unit
An impulse response estimator for estimating an impulse response of the target signal from the centroid information of the cluster;
An adaptive beamformer calculation unit that calculates an adaptive beamformer using the impulse response and the unnecessary signal correlation matrix, and
The blind signal extraction apparatus, wherein the target signal extraction unit extracts the target signal from the beamformer input signal using the adaptive beamformer. The blind signal extraction device according to claim 2 or 3,
Furthermore, a sensor position information storage unit storing sensor position information representing the positions of the M sensors is provided,
The impulse response estimator is
A direction-of-arrival estimation unit that estimates a direction of arrival of a signal using the sensor position information and the centroid information of the cluster;
A blind signal extraction device comprising: an estimated response arrival direction; and an impulse response calculation unit that calculates an impulse response from the sensor position information. The blind signal extraction device according to claim 1,
Furthermore, an observation signal correlation matrix estimation unit for estimating an observation signal correlation matrix that is a correlation matrix of the observation signal from the observation signal is provided,
The beamformer calculation unit
A target signal correlation matrix estimator for estimating a target signal correlation matrix that is a correlation matrix of an observation signal including the target signal from the cluster information;
A maximum gain beamformer calculation unit for calculating a maximum gain beamformer from the unnecessary signal correlation matrix and the target signal correlation matrix;
A correction unit that corrects frequency distortion using the observed signal correlation matrix and calculates a corrected beamformer for the maximum gain beamformer, and
The target signal extraction unit extracts the target signal from the frequency domain signal using the correction beamformer.
A blind signal extraction device characterized by that. The blind signal extraction device according to claim 1,
Furthermore, an observation signal correlation matrix estimation unit for estimating an observation signal correlation matrix that is a correlation matrix of the observation signal from the observation signal is provided,
The beamformer calculation unit
A target signal correlation matrix estimator for estimating a target signal correlation matrix that is a correlation matrix of an observation signal including the target signal from the cluster information;
A maximum gain beamformer calculation unit for calculating a maximum gain beamformer from the unnecessary signal correlation matrix and the target signal correlation matrix;
A correction vector calculation unit that calculates a correction vector using the observed signal correlation matrix and the maximum gain beamformer,
The target signal extraction unit
A signal extraction unit that extracts a target signal including distortion from the frequency domain signal using the maximum beamformer;
A blind signal extraction apparatus comprising: a distortion correction unit that performs distortion correction on the target signal including the extracted distortion using the correction vector and outputs the target signal. In a signal extraction method of observing signals emitted from N signal sources with M sensors and extracting one or more signals from the observed signals, N and M are integers of 2 or more. Yes,
A frequency domain transforming process in which the frequency domain transforming means transforms the observation signals observed by the M sensors into signals in the frequency domain;
A normalization process in which the normalization means normalizes the frequency domain signal and calculates a normalized observation signal vector;
A clustering process in which the clustering means clusters the normalized observation signal vector into N clusters;
An unnecessary signal correlation matrix estimation process is an unnecessary signal correlation matrix estimation process for estimating an unnecessary signal correlation matrix that is a correlation matrix of an observation signal including only an unnecessary signal from the information of the cluster,
A beamformer calculating means for calculating a beamformer from the cluster information and the unnecessary signal correlation matrix;
A target signal extraction means for extracting a target signal from the frequency domain signal using the beam former;
A blind signal extraction method comprising: a time domain conversion process in which time domain conversion means converts the extracted target signal into a time domain signal. The blind signal extraction method according to claim 7,
The beamformer calculation process is as follows:
An impulse response estimation means for estimating an impulse response of the target signal from the centroid information of the cluster;
An adaptive beamformer calculation means comprising: an adaptive beamformer calculation process for calculating an adaptive beamformer using the impulse response and the unnecessary signal correlation matrix;
The blind signal extraction method, wherein the target signal extraction step is a step of extracting the target signal from the frequency domain signal using the adaptive beamformer. The blind signal extraction method according to claim 7,
The unnecessary signal correlation matrix estimation process is a process of estimating an unnecessary signal correlation matrix that is a correlation matrix of K unnecessary signals selected from the cluster information, where K is an integer satisfying K ≦ M−1. And
Further, the input signal estimation means has an input signal estimation process for estimating a beamformer input signal including only the target signal and the selected K unnecessary signals from the cluster information,
The beamformer calculation process is as follows:
An impulse response means for estimating an impulse response of the target signal from the centroid information of the cluster;
An adaptive beamformer calculating means includes an adaptive beamformer calculating step of calculating an adaptive beamformer using the impulse response and the unnecessary signal correlation matrix;
The blind signal extraction method, wherein the target signal extraction step is a step of extracting a target signal from the beamformer input signal using the adaptive beamformer. A blind signal extraction method according to claim 8 or 9, wherein
The impulse response estimation process is as follows:
An arrival direction estimation unit in which an arrival direction estimation unit estimates an arrival direction of a signal using the sensor position information in the sensor position information storage unit and the centroid information of the cluster, and here, a sensor position information storage unit Is stored sensor position information indicating the positions of the M sensors.
A blind signal extraction method, wherein the impulse response calculation means includes an impulse response calculation step of calculating an impulse response from the estimated arrival direction of the signal and the sensor position information. The blind signal extraction method according to claim 7,
Further, the observation signal correlation matrix estimation means includes an observation signal correlation matrix estimation process for estimating an observation signal correlation matrix that is a correlation matrix of the observation signal from the observation signal,
The beamformer calculation process is as follows:
A target signal correlation matrix estimation means for estimating a target signal correlation matrix, which is a correlation matrix of an observation signal including the target signal, from the cluster information;
A maximum gain beamformer calculating means for calculating a maximum gain beamformer from the unnecessary signal correlation matrix and the target signal correlation matrix;
A correcting means for correcting the frequency distortion using the observed signal correlation matrix for the maximum gain beamformer and calculating a corrected beamformer, and
The target signal extraction process is a process of extracting a target signal from the frequency domain signal using the correction beamformer.
And a blind signal extraction method. The blind signal extraction method according to claim 7,
Further, the observation signal correlation matrix estimation means includes an observation signal correlation matrix estimation process for estimating an observation signal correlation matrix that is a correlation matrix of the observation signal from the observation signal,
The beamformer calculation process is as follows:
A target signal correlation matrix estimation means for estimating a target signal correlation matrix, which is a correlation matrix of an observation signal including the target signal, from the cluster information;
A maximum gain beamformer calculating means includes a maximum gain beamformer calculating step of calculating a maximum gain beamformer from the unnecessary signal correlation matrix and the target signal correlation matrix;
Further, the correction vector calculation means has a correction vector calculation step of calculating a correction vector using the observed signal correlation matrix and the maximum gain beamformer,
The target signal extraction process is as follows:
A signal extraction process in which a signal extraction means extracts a target signal including distortion from the frequency domain signal using the maximum beamformer;
A blind signal extraction comprising: a distortion correction process, wherein distortion correction means performs distortion correction on the target signal including the extracted distortion by using the correction vector, and outputs the target signal. Method.   The blind signal extraction program for making a computer perform each process of the blind signal extraction method in any one of Claims 7-12.   A computer-readable recording medium on which the blind signal extraction program according to claim 13 is recorded.

Images