JP4714892B2 - High reverberation blind signal separation apparatus and method - Google Patents

Description

  The present invention relates to a signal separation technique for performing signal separation of a mixed signal obtained by linearly mixing a plurality of statistically independent original signals, and in particular, using original component analysis (hereinafter referred to as “ICA”). The present invention relates to a technique for separating signals.

  A signal processing technique for separating a mixed signal into an original signal is called blind signal separation. In recent years, ICA has been attracting attention as a blind signal separation technique, and research and development are being promoted in the fields of audio signal processing, image processing, communication, and the like. In ICA, for example, each original signal is statistically independent of multiple observed signals obtained by receiving mixed sound, which is a linear mixture of original signals emitted from multiple sound sources, with R receivers. By utilizing the fact that the observation signal is processed, it is possible to separate the same number of R or less than R original signals as the receiver.

  The ICA is divided into a time domain ICA (Time Domain ICA: hereinafter referred to as “TDICA”) that performs signal separation in the time domain and a frequency domain ICA (Frequency Domain ICA: hereinafter referred to as “FDICA”) that performs signal separation in the frequency domain. Broadly divided.

  "Spatial mixture" or "instantaneous mixture" when the sound wave reaching two or more receivers has no time delay or convolution and the elements of the mixing matrix are represented by real constants That's it. In the case of instantaneous mixing, the sound source can be completely separated by either TDICA or FDICA.

  However, in an actual environment, due to the difference in distance between the sound source and the sound receiver, there is a time difference in the signals received by the individual sound receivers. In addition, echoes and reverberations are convoluted and observed. A case where the original signals are mixed spatially and temporally in this manner is called “spatio-temporal mixture” or “convolutive mixture”. The problem of separation of the original signal of convolutional mixing is called “Blind Source Deconvolution (hereinafter referred to as“ BSD ”)”. The problem of blind source deconvolution is formulated as follows:

[1] Blind source deconvolution (BSD)
Let R _{1 be} the number of sound sources and R ₂ the number of sound receivers. The original signal output from the sound source is denoted as s _i (t), and the observation signal at the receiver is denoted as x _j (t). i is an index representing the number of the sound source, and j is an index representing the number of the receiver.

The observation signal x _j (t) in the time domain is modeled by the Lth order filter as follows:

Here, t represents time and τ represents delay time. Also, the impulse response from each sound source to each receiver is denoted as a _ji (τ). The matrix A (τ) = (a _ji (τ)) is called a “mixing matrix”. The shape of a general impulse response a _ji (τ) has a strong pulse-like response after a lapse of an appropriate time and decays with time. Let this impulse response length be T. However, T is unknown.

The purpose of the BSD is to eliminate the length w of the FIR (Finite Impulse Response) filter coefficient w _ij (τ) and the separated signal without the information of the original signal s _i (t) and impulse response a _ji (τ). It is to obtain y _i (t). The coefficient matrix W (τ) = (w _ij (τ)) is referred to as a “separation matrix”.

FIG. 12 shows the BSD configuration when R ₁ = 2 and R ₂ = 2. When Formula (1) and Formula (2) are converted to the frequency domain, the following formula is obtained.

Here, x _j (ω _n , k) and s _i (ω _n , k) are discrete Fourier transforms (Discrete Fourier Transform: hereinafter referred to as “DFT”) of x _j (t) and s _i (t), respectively. a _ji (ω _n ) and w _ij (ω _n ) are DFTs of a _ji (t) and w _ij (t), respectively, ω _n is a normalized angular frequency, and k is a frame number.

In TDICA, signal separation is performed by calculating w _ij (τ) of Equation (2) in the time domain, and in FDICA, signal separation is performed by calculating w _ij (ω _n ) of Equation (4) in the frequency domain. . In BSD, when trying to separate the original original signal from the mixed signal by TDICA, a filter with order (tap length) L of several thousand is required, which causes problems of algorithm stability and calculation load. Therefore, in general, BSD is often solved by FDICA. This is because convolution in the time domain can be expressed as a product term in the frequency domain and can be handled as easily as in the case of instantaneous mixing.

[2] TDICA
Next, an outline of TDICA will be briefly described. In TDICA, the separation matrix is such that each component of the separation signal Y (t) = [y ₁ (t),..., Y _R1 (t)] ^T (where x ^T represents the transpose of x) is independent. Determine W (τ) = (w _ij (τ)). The distance to measure independence is the Kullback-Leibler information amount (hereinafter referred to as “KL information amount”) between the simultaneous distribution of each component of the separated signal Y (t) and the peripheral distribution. I _KL (Y (t)) Can be used.

However, C (Y) represents the entropy of Y. In order to make each component of the separation signal Y (t) completely independent, the separation matrix W (τ) may be determined so that I _KL (Y (t)) = 0. Since the amount of KL information is non-negative, it becomes minimum when the derivative of I _KL (Y (t)) by W (τ) is zero. When this is solved, the update formula of the separation matrix is formulated as follows.

Here, I is R ₁ = matrix of R ₁ × R ₁ when R _2, R ₁ <when R ₂ of _{R 1 × R 2 (I:} O) becomes the matrix, O is R ₁ × (R ₂ -R ₁ ), where η is the search step width and δ (τ) is the Kronecker delta. Also, φ is a function that is ideally defined by the following equation using the peripheral probability density p _yi (y) of y _i . However, since the exact distribution of the sound source is generally unknown, φ uses a nonlinear function such as tanh (•).

  First, consider the off-diagonal component of the TDICA algorithm (Equation (6)). The non-diagonal component is expressed as the following equation (10).

  Therefore, the following equation (11) is established at the convergence point of W (τ).

This indicates that the separated separated signals y _i (t) and y _j (t) are independent not only spatially but also temporally. That is, the off-diagonal component of Equation (6) functions so that the separated signal is independent both spatially and temporally.

  On the other hand, consider the diagonal component of equation (6). The diagonal component is expressed as the following equation (12).

  Therefore, the following equation (13) holds at the convergence point of W (τ).

This indicates that the separated signal y _i (t) is scaled to φ (y _i (t)) y _i (t) = 1 and is also independent in time. That is, the diagonal component of Equation (6) works to scale the separated signal and to make it independent (white) in time. In the case of non-diagonal components, temporal independence is not a problem because it works to independence for different signals. However, in the case of the diagonal component, since one signal is temporally independent, distortion occurs. This distortion is called “whitening distortion”.

FIG. 13 is a diagram showing the mixing process and the separation process as seen from TDICA. Assuming that the distribution of the white signal s _i ′ (t) is a sound source with iid (identically independent distribution), the original signal s _i (t) is an audio signal tuned by the articulator in the oral cavity. Unlike the white signal s _i ′ (t), the original signal s _i (t) has a temporal correlation. Therefore, the TDICA algorithm assumes the iid of the signal to be separated and updates the separation matrix so as to satisfy the process. The separated signal y _i (t) obtained as a result is not only spatially independent as described above, but also temporally independent (whitened). Therefore, when applying TDICA algorithm to the original signal s _i (t) with a temporal correlation as speech, the estimation of the separated signals y _i (t) in s _i (t) is not s _i '(t) It becomes a value and becomes distorted. This distortion is called whitening distortion because it is similar to the distortion of white noise.

[3] Principle of minimum distortion
As a method for solving whitening distortion in TDICA, a method based on the principle of minimum distortion (Minimum Distortion Principle: hereinafter referred to as “MDP”) described in Non-Patent Document 3 is known. The outline of MDP and its application to TDICA are briefly described below.

First, the mixing process from s _i ′ (t) to x _j (t) and the separation process from x _j (t) to y _i (t) are considered as shown in FIG. The separation filter matrix obtained by TDICA is expressed in the form of z-transform as shown in Equation (14).

Even if the source s _i (t) is separated by TDICA, the separation filter matrix W (z) does not become the inverse matrix A ⁻¹ (z) of the mixing matrix, and is expressed by whitening as shown in Equation (15). A diagonal matrix D (z) is given over A ^-1 (z).

Here, D (z) is a diagonal matrix having the filters d _i (z) (i = 1,..., R ₁ ) as diagonal elements. On the other hand, the sound source s (t) is considered to be a white signal s ′ (t) modulated by a diagonal matrix E (z) whose diagonal components are filters e _i (z) (i = 1,..., R ₁ ). It is done. Therefore, it can be interpreted that TDICA estimates E ⁻¹ (z) as D (z).

  Here, D (z) can be canceled by performing an operation such as the following equation (16) on the separation matrix W (z).

  Therefore, by replacing the left side of Equation (16) with W (z), the arbitraryness of D (z) can be removed from W (z).

Equation (17) is equivalent to D (z) = diag A (z) in Equation (16). At this time, the separated signal y _i (t) becomes a _ii (z) s _i (t). Therefore, the scaling of y _i (t) is specified by the transfer function and makes sense, and the distortion of the separated signal is minimized.

To make D (z) become diag A (z), evaluate the distortion with E [‖y (t) −x (t) ‖ ² ], and minimize this evaluation function W (z) Has been shown to be obtained (see Non-Patent Document 3). This is the principle of minimum distortion (MDP).

Next, the TDICA algorithm based on MDP will be described. To minimize E [‖y (t) -x ( t) ‖ ^2] according to MDP, the natural gradient of E [‖y (t) -x ( t) ‖ ^2]

とすると、

Then,

は式（18）で与えられる。また、そのon-line展開形は式（19）で与えられる。

Is given by equation (18). The on-line expansion is given by equation (19).

On the other hand, the update formula of TDICA separation matrix (Equation (14)) is divided into non-diagonal component ΔW _off (τ) and diagonal component ΔW _diag (τ) as shown in Equations (20) to (22). be able to.

Here, ξ represents the ratio of ΔW _diag (τ) to ΔW _off (τ), and ξ = 1 in equation (6). As described above, off-diagonal components do not contribute to whitening and are used as they are. When ΔW _off (τ) is expressed in an on-line expansion form, Equation (23) is obtained.

On the other hand, the diagonal component ΔW _diag (τ) contributes to scaling adjustment and also causes whitening, so it is derived based on MDP.

と置換する。以上の結果、MDPに基づくTDICA法における分離行列の更新式のon-line展開形は、式（24）により表される。

Replace with As a result, the on-line expansion form of the separation matrix update formula in the TDICA method based on MDP is expressed by formula (24).

  Here, ψ (t) = y (t) −x (t) and γ = 2ξ. γ represents the update ratio of the diagonal term to the off-diagonal term.

[4] FDICA
Next, an outline of FDICA will be briefly described. In FDICA, each of the separated signals Y (ω _n , k) = [y ₁ (ω _n , k),..., Y _R1 (ω _n , k)] ^T (where x ^T represents transposition of x). The separation matrix W (ω _n ) = (w _ij (ω _n )) is determined so that the components are independent. The update formula of the separation matrix W (ω _n ) is obtained in the same manner as in the case of TDICA, and is as follows.

  Here, I is a unit matrix defined in the same manner as in equation (6), η is a search step width, and φ is a nonlinear function defined in the same manner as in equations (8) and (9).

In FDICA, component replacement (permutation) and scaling (scaling) are major problems. That is, as in the case of TDICA, time series Y (ω _n , k) = [y ₁ (ω _n , k),..., Y _R1 (ω _n , k)] ^{T for} each frequency of the separated signal obtained by FDICA Are arbitrary in order between elements and arbitrary in scaling (magnitude and phase). This is because the independence between the elements is maintained even if the order of the elements is changed or the scaling of the elements is changed. From Equation (4), these arbitraryities regarding the separation signal Y (ω _n , k) correspond to the arbitraryities of the separation matrix W (ω _n ) = (w _ij (ω _n )). Therefore, an operation for arranging the row components of the separation matrix W (ω _n ) so as to correspond to the same signal and an operation for appropriately adjusting the scaling of the row components are required. The former is a component replacement problem, and the latter is a scaling problem.

FIG. 14 shows an example where the number of sound sources and the number of sound receivers are both two. In FDICA, the observed signals x ₁ (t) and x ₂ (t) are Fourier transformed for each frame of time width N, and N subbands {x ₁ (ω ₀ , k),..., X ₁ (ω _{N −1} , k)}, {x ₂ (ω ₀ , k),..., X ₂ (ω _N−1 , k)}. Then, for each subband, instantaneous mixing FDICA is performed, and N sets of separated signals {y ₁ (ω _n , k), y ₂ (ω _n , k) | n = 0,..., N−1} Calculated. These separated signals are subjected to inverse Fourier transform and output as separated signals y ₁ (t) and y ₂ (t) in the time domain. Here, in the example of FIG. 14, when the components of the separated signal y ₁ (ω _n , k) and the separated signal y ₂ (ω _n , k) are ω _n = ω ₀ and ω _n = ω ₁ by component replacement. It is replaced at times. When such component replacement occurs, when the separated signals y ₁ (t) and y ₂ (t) in the time domain are reproduced by inverse Fourier transform, the subband components are mixed again and distortion occurs.

Similarly, the gains of the separated signals y ₁ (ω _n , k) and y ₂ (ω _n , k) are different for each subband due to the arbitrary scaling. Therefore, when the separated signals y ₁ (t) and y ₂ (t) in the time domain are reproduced by inverse Fourier transform, unnatural weighting is performed and distortion occurs.

  As a method for solving the scaling problem, for example, a method based on MDP described in Non-Patent Document 3 is known (see Patent Document 4, [0011]).

  As methods for solving the problem of component replacement, for example, Patent Documents 1 to 3 and Non-Patent Documents 1, 2, 5 to 7 are known. Here, as an example, the component replacement elimination method described in Non-Patent Document 1 will be described. For simplicity, the case where the number of sound sources and receivers is both 2 will be described.

First, it is assumed that the separation matrix W (ω _n ) is obtained by FDICA, and the separation signal y _i (ω _n , k) is represented by the observation signal x _j (ω _n , k) by the equation (4). Here, the divided spectrum v _i (ω _n , k) is defined as follows for the separated signal y _i (ω _n , k).

This divided spectrum is expressed by Equation (30) when there is no component replacement, and is expressed by Equation (31) when component replacement occurs, and the transfer function between the sound source s _i (ω _n , k) and the sound source and receiver. It is uniquely represented by the product of a _ji (ω _n ).

Since the first and second components of v _i (ω _n , k) are estimated values for the same sound source s _i (ω _n , k), the two sound sources s ₁ (ω _n , k), The estimated value of s ₂ (ω _n , k) can be represented by a split spectrum of either v ₁₁ (ω _n , k) or v ₂₂ (ω _n , k).

  This component replacement cancellation method utilizes the fact that the decomposition spectrum entropy for speech is higher than the decomposition spectrum entropy for noise.

Therefore, in order to introduce the entropy of the decomposition spectrum, first, each partition when the distribution range of the real part Re {v _ij (ω _n )} of the divided spectrum is equally divided into M is represented by ι _m (m = 0,. Γ−1) and the frequency of entering the block ι _m is κ _ij (ω _n , ι _m ). The probability ρ _ij (ω _n , ι _m ) obtained by normalizing κ _ij (ω _n , ι _m ) is defined by the following equation.

Using this probability ρ _ij (ω _n , ι _m ), entropy for the distribution of the real part Re {v _ij (ω _n )} of the split spectrum is defined by the following equation.

Since the split spectrum v _ij (ω _n , k) is uniquely represented by the product of the sound source s _i (ω _n , k) and the transfer function a _ji (ω _n ) between the sound source and the receiver, C _i1 and C _i2 represents the entropy for the same sound source. Therefore, the difference between the two (formula (34)) is examined, and if the entropy difference ΔC (ω _n ) is negative, it is determined that there is no component replacement, and the final output is z (ω _n , k) = [v ₁₁ (ω _n , k), v ₂₂ (ω _n , k)], conversely, if the difference ΔC (ω _n ) is positive, it is determined that component replacement has occurred, and the final output is z (ω _n , k) = [v ₂₂ (ω _n , k), v ₁₁ (ω _n , k)] is assigned.

In the above description, the case where both the number of sound sources and sound receivers is two has been described. However, it is easy to expand to the case where the number of sound sources is R ₁ and the number of sound receivers is R ₂ . As described above, the problem of component replacement in FDICA can be solved.

[5] Convolution FDICA (ConvFDICA)
In FDICA, there is a problem that the sound source decomposition performance is remarkably lowered when reverberation increases. This is because when the signal is converted to the frequency domain by DFT, if the Fourier transform analysis window width is narrow, it cannot include all of the impulse response, so even if the direct sound part of the sound source can be separated, the subsequent reflection and reverberation part Due to the inseparability. FIG. 15 shows this state. FIG. 15 shows a case where an analysis window having a window width N is used for the impulse response length T. In this case, an impulse response larger than N is not taken into the DFT. Therefore, the reverberation part that is not captured cannot be separated.

  Therefore, it seems that reverberation components can be removed by widening the analysis window width. However, if the analysis window width is widened for the same length of observation signal, the time resolution becomes coarse, and as a result, a sufficient number of data required for ICA cannot be obtained. Therefore, even if the analysis window width is reduced, the ability of ICA to degrade is reduced. In addition, if the observation signal is long, the reverberation component can be removed by widening the analysis window width. However, considering practical use, it is unlikely that the same environment will last for a long time. Moreover, it is necessary to separate the moving sound source with a short-time observation signal.

  In order to solve this problem, Non-Patent Document 4 introduced the concept of split convolution, and divided the impulse response of the mixing process, so that it was approximately modeled as instantaneous mixing in the frequency domain until then. Is modeled strictly as a convolutional mixture. A convolutional FDICA (hereinafter referred to as “ConvFDICA”) method is proposed in which the TDICA algorithm for time-domain convolutional mixing is applied to the frequency domain.

In ConvFDICA, a narrowband separation filter matrix is defined as shown in Equation (35). This W (ω _n , z) is called a “narrowband separation filter matrix”.

  When this is applied to the convolution model in the frequency domain, an update matrix like the following equation (36) is obtained. This is called ConvFDICA.

Here, k is the frame number, M is the tap length of the narrowband separation filter, and L = MN. ΔW (ω _n , p), W (ω _n , p), φ (y (ω _n , p)), and y (ω _n , p) are each expressed by the following equations.

  In ConvFDICA, the analysis window width is N, the tap length of the separation filter at each frequency is M, and the time domain filter of length MN is decomposed into MN and each frequency in the frequency domain. As a result, a separation filter that is effective even under high reverberation and has excellent convergence can be realized with an appropriate analysis window width sufficient to obtain sufficient statistics even from short-time observation signals.

However, since ConvFDICA inherits the properties of TDICA, it involves the problem of whitening distortion inherent to TDICA. Furthermore, the problem of scaling instability and component replacement also arises due to the separation in the frequency domain.
JP-A-2005-91560 JP 2005-91732 A JP 2004-302122 A JP 2005-79781 A Kanada, Satoshi, Furuya Takeshi, Gotanda Hiroshi, “Resolving component replacement based on entropy of split spectrum”, IEICE Transactions, Vol.J87-A, No.7, pp. 1065-1069, July 2004. H. Sawada, R. Mukai, S. Araki, S. Makino, "A Robust and Precise Method for Solving the Permutation Problem of Frequency-Domain Blind Source Separation," IEEE Trans. Speech and Audio Processing, vol.12, no. 5, pp. 530-538, Sep. 2004. K. Matsuoka and S. Nakashima, "Minimal Distortion Principle for Blind Source Separation". Proc. ICA 2001, pp.722-727, 2001. C. Serviere, "Separation of speech signals with segmentation of the impulse responses under reverberant conditions ICA, Int.Workshop on Independent Component Analysis and Signal Separation, Nara, Avril 2003 Hiroshi Saruwatari, Satoshi Kurita, Kazuya Takeda, Fumitada Itakura, Kiyohiro Shikano, “Blind sound source separation using band-split ICA and beamforming”, IEICE Technical Report, EA2000-14, pp.1-8, 2000. Takaaki Ishibashi, Satoshi Kaneda, Takeshi Furuya, Hiroshi Gotanda, "Estimation of sound source direction based on phase of split spectrum", Proceedings of the 56th Kyushu Branch Joint Conference on Electrical Engineering, CD-ROM, 12-2A -10, (2003, H15-9, Sojo University) H. Gotanda, K. Nobu, T. Koya, K. Kaneda, T. Ishibashi, N. Haratani, "Permutation correction and speech extraction based on split spectrum through FastICA," Proc. ICA2003, pp.379-384, 2003. S. Araki, S. Makino, R. Aichner, T. Nishikawa, and H. Saruwatari, "Subband based blind source separation with appropriate processing for each frequency band," Proc. ICA2003, pp.499-504, 2003.

  The problems of the conventional signal separation technology are summarized as follows.

(1) In TDICA, the problem of component replacement does not pose a serious problem in application, but the problem of scaling indefiniteness and whitening distortion pose a serious problem. These problems can be solved by TDICA using MDP.

  However, if TDICA is applied to a BSD model that takes reflection and reverberation into account, the filter order becomes enormous, which causes a large mounting problem. In addition, the problem of algorithm and the problem of calculation load are also great. Therefore, application to the BSD model is inappropriate, and TDICA is not effective in an environment with a lot of reverberations.

(2) Since FDICA separates each subband by instantaneous mixing, it can be realized with a small arithmetic circuit even when applied to the BSD model, and it converges quickly and has excellent solution stability. On the other hand, in FDICA, there are a problem of component replacement, a scaling problem, and a problem that reverberation cannot be completely separated when an impulse response is long.

  The problem of component replacement is solved by, for example, the component replacement elimination method based on the entropy of the decomposition spectrum described above. Further, the scaling problem is solved by the above-described TDICA method based on MDP.

  However, the problem that reverberation cannot be completely separated when the impulse response of the observation signal is long cannot be solved.

(3) According to ConvFDICA described in Non-Patent Document 4, it has an appropriate analysis window width that can provide sufficient statistics even from short-time observation signals. An excellent separation filter can be realized. However, since ConvFDICA inherits the properties of TDICA, it involves the problem of whitening distortion inherent to TDICA. Furthermore, there remain problems of scaling indefiniteness and component replacement that inevitably occur in the frequency domain.

  Therefore, an object of the present invention is to provide a signal separation technique that can effectively separate observation signals even under high reverberation and can solve all the problems of whitening distortion, scaling, and component replacement. There is to do.

  The first configuration of the high reverberation blind signal separation device according to the present invention calculates time series data for each frequency by Fourier transforming a plurality of observation signals, and performs independent component analysis (hereinafter “ICA”) from the time series data. A high-reverb reverberation blind signal separation device that calculates a separation matrix for each frequency and calculates a separation signal using the separation matrix, wherein the separation matrix by ConvFDICA based on MDP is used for the time-series data for each frequency. And a separation matrix updating means for updating the separation matrix using the update formula.

  According to this configuration, the problem of whitening distortion can be solved by using MDP. Using MDP also solves the scaling problem. Furthermore, since the separation matrix in the frequency domain is updated, the convergence is faster than TDICA. Moreover, since the convolution model is used, it is possible to remove reverberation.

  Here, the “separation matrix update formula by ConvFDICA based on MDP” refers to an update formula given by formula (85) described later.

  A second configuration of the high reverberation blind signal separation device according to the present invention is the approximate value of the separation matrix without component replacement of matrix elements between the frequencies in the first configuration based on the observation signal ( (Hereinafter, referred to as “initial separation matrix”). The separation matrix update means includes a separation matrix initialization means, and the separation matrix update means updates the separation matrix using the initial separation matrix as an initial value.

  According to this configuration, first, an initial separation matrix without component replacement of matrix elements between each frequency is calculated, and this initial separation matrix is set as an initial value, so that the vicinity of the final convergence point of the separation matrix An initial value can be set for. Then, by updating the separation matrix using this initial value, the separation matrix can be converged to the final convergence point without causing component replacement.

  That is, the solution space of the separation coefficient is considered to be a complicated curved surface as shown in FIG. In this case, the correct convergence point has the smallest measure of independence, but there are other solutions with component replacement as local minimum values. On the other hand, since ICA searches for a solution in the separation coefficient space by the natural gradient method, if an initial value is set in the vicinity of the correct convergence point, it moves in a direction in which the curvature decreases and becomes the convergence point of component replacement. It converges to the correct convergence point without jumping. The present invention solves the problem of component substitution by utilizing this property.

  Here, the method of “calculating an initial separation matrix without component replacement of matrix elements between frequencies” is not particularly limited, and various known methods described above can be used.

  According to a third configuration of the high reverberation blind signal separation device according to the present invention, in the second configuration, the separation matrix initialization unit has no component replacement in the frequency domain ICA of instantaneous mixing based on the observation signal. The initial separation matrix is calculated.

  In the first configuration of the anti-reverberation blind signal separation method according to the present invention, time series data for each frequency is calculated by Fourier transforming a plurality of observation signals, and a separation matrix for each frequency is calculated from the time series data by ICA. A high-resistance reverberant blind signal separation method for calculating a separation signal using the separation matrix, wherein the time-series data for each frequency is obtained by using an update equation of the separation matrix by ConvFDICA based on MDP. It has the separation matrix update procedure which updates.

  A second configuration of the high reverberation blind signal separation method according to the present invention is an approximate value of the separation matrix without component replacement of matrix elements between frequencies based on the observation signal in the first configuration. A separation matrix initialization procedure for calculating a certain initial separation matrix is provided, and the separation matrix is updated in the separation matrix update procedure using the initial separation matrix as an initial value.

  According to the third configuration of the high reverberation blind signal separation method according to the present invention, in the second configuration, in the separation matrix initialization procedure, no component replacement is performed in the frequency domain ICA of instantaneous mixing based on the observation signal. An initial separation matrix that is an approximate optimum value of the separation matrix is calculated.

  The program according to the present invention, when executed by a computer, causes the computer to function as any one of the first to third anti-reverberation blind signal separation devices.

  As described above, according to the present invention, it is possible to effectively separate observation signals even under high reverberation, and signal separation that can solve all the problems of whitening distortion, scaling, and component replacement. Technology can be provided.

  The best mode for carrying out the present invention will be described below with reference to the drawings. First, after explaining the basic concept, embodiments of the present invention will be described.

[1] Strict modeling of mixing process The high reverberation blind signal separation apparatus of the present invention performs ICA in the time-frequency domain. Here, in the ICA in the time-frequency domain, the observation signal represented by Equation (1) is subbanded by DFT, and the ICA is performed using the subband observation signals time-series in frame units. However, when Equation (3) obtained by simply transforming Equation (1) by DFT is again transformed back to the time domain, the time domain representation becomes cyclic convolution. At this time, the cyclic convolution is equal to the linear convolution of Equation (1) when L >> T, but is different when L >> T is not satisfied. This causes an error.

  Therefore, first, a method for eliminating an error caused by cyclic convolution will be described, and then a model that accurately reflects the mixing process of Equation (1) in the frequency domain will be described.

  First, when the linear convolution of Expression (1) is expressed as a matrix, it is expressed as the following Expression (41).

一方、式（3）は逆DFTにより時間領域に戻すと、次式（45）のように表される。

On the other hand, when the expression (3) is returned to the time domain by inverse DFT, it is expressed as the following expression (45).

Here, it can be seen from the comparison between Equation (41) and Equation (45) that the matrix representation of linear convolution and the matrix representation of cyclic convolution are different. Therefore, when A _ji ^(L) is followed by L 0s added to A _ji ^(2L) , the length is 2L, and the linear convolution expression in equation (3) is rewritten, the following equation (49) become.

  In addition, when the cyclic convolution expression is similarly rewritten, the following equation (53) is obtained.

  Here, from the comparison between Equation (49) and Equation (53),

の後半のL個の

L in the second half

から

From

と、X_j ^（2L）（t）の後半のL個とは同じであるが、前半のL個は異なることが分かる。

And ^L _{j in} the second half of X _j ^(2L) (t) is the same, but L in the first half is different.

  By cyclic convolution

と線形畳み込みによるX_j ^（2L）（t）の間には上述のような相違があり、これに起因して誤差が生じる。しかしながら、両者は以下のような手順を踏むことによって等しくすることができる（非特許文献4参照）。

And the above-mentioned difference between X _j ^(2L) (t) by linear convolution and an error occurs due to this. However, both can be made equal by taking the following procedure (see Non-Patent Document 4).

  First, the first L windows are defined by Expression (57), and a diagonal matrix H is defined with each element of h as a diagonal component.

  By multiplying the diagonal matrix H on both sides of the equation (53), the following equation (58) is established, and the equation (59) is established, and the linear convolution and the cyclic convolution are equal.

  On the other hand, a matrix F that is an operator of Fourier transform is defined by Expression (60).

  When this is applied to both sides of equation (58), equation (62) is obtained.

  Further, by transforming this, the formula (63) is finally obtained. This is a frequency domain representation of the time domain representation of Equation (58) by DFT. Therefore, this equation (63) is called a strict model in the frequency domain of the time domain convolution mixing process.

[2] Mixing model by dividing impulse response Since the time T of the impulse response is long under high reverberation, the filter length L of the convolution of equation (1) is made larger than T to sufficiently eliminate the reverberation. There is a need. However, as described above, if L is increased, a sufficient number of data cannot be obtained for estimating sound sources in each subband, resulting in a problem in convergence. On the other hand, if the filter length L is L <T, the DFT does not accurately reflect the impulse response, and an error occurs because the window cannot include the entire impulse response.

  Therefore, first, assuming an appropriate window width N, the filter length L is equally divided into N so that L = MN. L is set to be sufficiently larger than T. At this time, the linear convolution of Equation (1) is as shown in Equation (69) below.

  Furthermore, when this is expressed as a matrix, it can be expressed as the following equation (70).

  Even if the equation (70) is converted into the frequency domain as it is, it is not reflected accurately, and therefore, the relationship of the equation (58) is used to transform into the following equation (74).

  Further, by multiplying both sides of the equation (74) by the DFT transformation matrix F to convert to the frequency domain, a strict model is finally obtained by the equation (78).

  here,

は式（79）を対角成分とする対角行列であり、また、それぞれの記号は、式（79）〜（83）により表される。

Is a diagonal matrix having the equation (79) as a diagonal component, and each symbol is represented by equations (79) to (83).

If the matrix representation of Equation (78) is described as a mixing process divided into subbands ω _n (n = 0, 1,..., 2N−1), a convolution representation in the frequency domain is obtained as shown in Equation (84). It becomes. This is called a discrete windowed discrete Fourier transform.

This can be expressed in a simple manner as shown in Fig. 1. In FIG. 1, the observation signal x _j (t) is DFT with a window width N in each time frame k, and each subband is convolved between frame 0 and frame K. As a result, all impulse responses can be captured, and a sufficient number of data can be obtained in a short time to estimate the sound source in each subband in order to calculate an appropriate window width N. Become.

[3] ICA algorithm of the present invention In the present invention, by using the discrete windowed discrete Fourier transform as described above, the observation signal of Equation (1) is subband decomposed, and each subband is based on MDP. Perform the ConvFDICA algorithm. In this case, the update formula of the separation matrix is expressed by the following formula (85).

Here, η is a search step width, and γ is an internal ratio. In addition, ΔW (ω _n , p), W (ω _n , p), φ (y (ω _n , k)), ψ (ω _n , k), and y (ω _n , k) are respectively expressed by equations ( 86) to (90). The number of sound sources and the number of sound receivers are both R.

Φi (y (ω _n , k)) and ψi (ω _n , k) are expressed by equations (91) and (92), respectively.

At this time, y (ω _n , k) is calculated as shown in Equation (93).

  The algorithm of the present invention is an application of the concept of MDP to the aforementioned ConvFDICA algorithm by Serviere (see Non-Patent Document 4). Hereinafter, the algorithm of the present invention is referred to as “MDP-based ConvFDICA (MDP-ConvFDICA)”. By executing ICA using this algorithm, it is possible to eliminate distortion caused by whitening, which has been a problem in ConvFDICA, and to solve the problem of scaling indefiniteness.

[4] Method for eliminating component substitution Finally, a method for eliminating component substitution will be described. In ICA, the true separation filter is obtained by updating the coefficients of the separation filter so that the output signals are independent of each other. When the coefficient of the separation filter is a parameter and the true separation filter is a convergence value, the parameter space is searched so as to minimize the measure of independence (evaluation function), and the convergence value is obtained. Therefore, it can be considered as a solution search problem in the parameter space.

  In the case of instantaneous mixing with a tap length of 1, the parameter surface is simple as shown in Fig. 2 (a), and it has excellent convergence and stability. However, in the case of convolutional mixing in consideration of reverberation and reflection, the parameter surface becomes a complex surface as shown in FIG. 2 (b), the convergence stability is lowered, and a minimal solution other than the true separation filter (local minimum) ).

However, as shown in FIG. 3, the results obtained by beam forming (see Non-Patent Document 5) and methods using a priori information (Non-Patent Documents 1, 6, and 7) are used as appropriate initial values W ⁰ (z). By setting, it is possible to search for a solution stably, and at the same time, to solve component replacement (see Non-Patent Document 8).

  Therefore, in the present invention, first, the problem of instability of component replacement and scaling in FDICA of instantaneous mixing is solved, and a separation matrix obtained by that is adopted as an initial value of the above-described ICA algorithm of the present invention. Then, if the ICA algorithm of the present invention is executed based on this initial value, a true separation filter without component replacement can be obtained.

  FIG. 3 is a diagram schematically showing the convergence of the ICA algorithm by setting appropriate initial values. As shown in FIG. 3, the parameter surface of convolutional mixing is complex, and solutions with component replacement exist as local minimums (local minimum). Since the ICA algorithm is a search based on the natural gradient method, the search point moves in the direction in which the curvature of the parameter curved surface decreases. Therefore, by setting an initial value appropriately, it is possible to search for a true separation filter without component replacement without falling into a local minimum.

FIG. 4 is a diagram illustrating an overall configuration of the high reverberation blind signal separation device 1 according to the first embodiment of the present invention. The high reverberation blind signal separation device 1 according to the present embodiment uses observation signals x ₁ (t), x ₂ (t) obtained by receiving sound signals emitted from two sound sources by two sound receivers. The separated signals y ₁ (t) and y ₂ (t) are output.

  The high reverberation blind signal separation device 1 includes FFT means 2, subband separation means 3_0 to 3_ (N-1), and IFFT means 4.

The FFT means 2 performs subband decomposition by performing FFT on a frame basis for each of the observation signals x ₁ (t) and x ₂ (t), and N-band observation signals x ₁ (ω _n , k), x ₂ (ω _n , k) (n = 0, ..., N-1) are output. Here, ω _n is an angular frequency, k is a frame number, and n is a subband number.

The subband separation means 3_n (n = 0,..., N−1) separates the observation signals x ₁ (ω _n , k) and x ₂ (ω _n , k) by ICA, and separates the separation signal y ₁ (Ω _n , k), y ₂ (ω _n , k) is output.

The IFFT means 4 performs inverse discrete cosine transform (IFFT) on the frequency domain separation signals y ₁ (ω _n , k) and y ₂ (ω _n , k) (n = 0,..., N−1), respectively. The time domain separation signals y ₁ (t) and y ₂ (t) are output.

  FIG. 5 is a block diagram showing the configuration of the subband separation means 3_n (n = 0,..., N−1) in FIG. The subband separation means 3_n includes a separation matrix storage means 11, a separation calculation means 12, a separation matrix initialization means 13, a separation matrix update means 14, a separation matrix synthesis means 15, a component replacement and scaling means 16, and a convergence determination means 17. I have.

The separation matrix storage unit 11 stores a separation matrix W (ω _n , p). The separation calculation means 12 uses the separation matrix W (ω _n , p) stored in the separation matrix storage means 11 to separate the separation signal from the observation signals x ₁ (ω _n , k) and x ₂ (ω _n , k). Calculate y ₁ (ω _n , k), y ₂ (ω _n , k).

The separation matrix initializing means 13 is based on the observation signals x ₁ (ω _n , k) and x ₂ (ω _n , k), and the updated value ΔW (ω _n , p) of the separation matrix W (ω _n , p) is obtained by FDICA. ). Further, by cooperating with the separation calculating means 12 and the separation matrix synthesizing means 15, the approximate value (initial separation matrix) of the separation matrix without the component replacement of the matrix elements between the frequencies is calculated.

The separation matrix updating means 14 uses the equation (85) based on the observation signals x ₁ (ω _n , k) and x ₂ (ω _n , k) to convert the separation matrix W (ω _n , p) by MDP-ConvFDICA. The update value ΔW (ω _n , p) is calculated. In addition, the separation matrix is updated by cooperating with the separation calculation means 12 and the separation matrix synthesis means 15.

The separation matrix synthesizing unit 15 stores the update value ΔW (ω _n , p) of the separation matrix W (ω _n , p) output from the separation matrix initialization unit 13 or the separation matrix update unit 14 and the separation matrix storage unit 11. It was added and the separation matrix W (ω _n, p) to synthesize a new separation matrix W (ω _n, p), thereby updating the separation matrix W of the separation matrix storage unit 11 (ω _n, p).

The component replacement and scaling means 16 detects the component replacement of the separation signals y ₁ (ω _n , k) and y ₂ (ω _n , k) obtained by FDICA, and the separation matrix W stored in the separation matrix storage means 11 Perform component replacement of (ω _n , p). It also eliminates scaling indefiniteness.

  Convergence determination means 17 determines whether or not the update of the separation matrix by the separation matrix initialization means 13 and the separation matrix update means 14 has converged, and controls the switching of the operations of the separation matrix initialization means 13 and the separation matrix update means 14 I do.

  The operation of the high reverberation-resistant blind signal separation device 1 according to this embodiment configured as described above will be described below.

  FIG. 6 is a flowchart showing the flow of the high reverberation-resistant blind signal separation method according to the first embodiment of the present invention. In FIG. 6, steps S1 to S6 are separation matrix initialization processing, and steps S7 to S12 are signal separation processing by MDP-ConvFDICA.

First, in step S 1, time domain observation signals x ₁ (t), x ₂ (t) are input to the FFT means 2.

In step S2, the FFT means 2 performs FFT on a frame basis for each of the observation signals x ₁ (t) and x ₂ (t), and performs subband decomposition. Thus, the observed signal x ₁ in the time domain _{(t), x 2 (t} ) is the observed signal x ₁ of the N subbands _{(ω n, k), x} 2 (ω n, k) (n = 0 , ..., N-1). Each observation signal x ₁ (ω _n , k), x ₂ (ω _n , k) is input to the separation calculation means 12 and the separation matrix initialization means 13.

In step S3, the separation operation means 12 for each subband reads the separation matrix W (ω _n , p) stored in the separation matrix storage means 11, and the separation signal Y (ω _n , k) according to the equation (93). = [Y ₁ (ω _n , k), y ₂ (ω _n , k)] ^T is calculated. In the initial state, it is assumed that the separation matrix W (ω _n , p) is initialized to a unit matrix.

In step S4, the separation matrix initialization unit 13 of each subband calculates an update value ΔW (ω _n , p) of the separation matrix by FDICA (Equation (25)). The separation matrix synthesizing means 15 reads the separation matrix W (ω _n , p) stored in the separation matrix storage means 11 and adds it to the new separation matrix W by adding it to the updated value ΔW (ω _n , p) of the separation matrix. Generate _new (ω _n , p). Then, the separation matrix W (ω _n , p) stored in the separation matrix storage means 11 is updated to a _new separation matrix W _new (ω _n , p).

In step S5, the convergence determination means 17 determines whether or not the magnitude | ΔW (ω _n , p) | of the update value ΔW (ω _n , p) of the separation matrix is smaller than a predetermined threshold value W _th . If | ΔW (ω _n , p) | ≧ W _th , the process returns to step S1. If | ΔW (ω _n , p) | <W _th , the process _proceeds to the next step S6.

In step S6, the component permutation and scaling means 16 performs entropy on the distribution of the real part of the divided vector of the separation signals y ₁ (ω _n , k) and y ₂ (ω _n , k) of each subband regarding the component permutation. H ₁₁ (ω _n ) and H ₂₂ (ω _n ) are calculated by equation (33), and the difference ΔH (ω _n ) is calculated by equation (34). If ΔH (ω _n )> 0, the order of the separation signals y ₁ (ω _n , k) and y ₂ (ω _n , k) is changed, and the separation matrix W stored in the separation matrix storage means 11 is changed. Replace the line of (ω _n , p). Next, scaling is solved for the separation matrix in accordance with Equation (16). This eliminates component replacement in FDICA. Thus, the separation matrix initialization process by the separation matrix initialization unit 13 is completed, and then the process proceeds to the signal separation process (step S7) by ConvFDICA based on MDP.

In step S7, time domain observation signals x ₁ (t), x ₂ (t) are input to the FFT means 2.

In step S8, the FFT means 2 performs FFT on a frame basis for each of the observation signals x ₁ (t) and x ₂ (t) to perform subband decomposition. Thus, the observed signal x ₁ in the time domain _{(t), x 2 (t} ) is the observed signal x ₁ of the N subbands _{(ω n, k), x} 2 (ω n, k) (n = 0 , ..., N-1). Each observation signal x ₁ (ω _n , k), x ₂ (ω _n , k) is input to the separation calculation means 12 and the separation matrix update means 14.

In step S9, the separation calculation means 12 for each subband reads the separation matrix W (ω _n , p) stored in the separation matrix storage means 11, and the separation signal Y (ω _n , k) is obtained by Expression (93). = [Y ₁ (ω _n , k), y ₂ (ω _n , k)] ^T is calculated.

In step S10, the separation matrix update means 14 for each subband calculates the update value ΔW (ω _n , p) of the separation matrix by MDP-ConvFDICA (Equation (85)). The separation matrix synthesizing means 15 reads the separation matrix W (ω _n , p) stored in the separation matrix storage means 11 and adds it to the new separation matrix W by adding it to the updated value ΔW (ω _n , p) of the separation matrix. Generate _new (ω _n , p). Then, the separation matrix W (ω _n , p) stored in the separation matrix storage means 11 is updated to a _new separation matrix W _new (ω _n , p).

In step S11, the IFFT means 4 determines the separation signals {y ₁ (ω _n , k) | n = 0,..., N−1}, {y ₂ (ω _i , k) | n = 0, .., N-1} are subjected to IFFT and converted into time domain separated signals y ₁ (t) and y ₂ (t).

Finally, in step S12, IFFT means 4 outputs time domain separation signals y ₁ (t) and y ₂ (t), returns to step S7 again, and thereafter repeats the same operation.

  In addition, this invention is not limited to the above-mentioned Example, In the range which does not deviate from the meaning of this invention, it can change suitably. The configuration of the present invention can also be realized by a computer. In this case, the processing contents of the functions that each means should have are described by a program, and the above processing functions can be realized on the computer by executing the program on the computer.

〔Experimental result〕
Finally, experimental results of the high reverberation blind signal separation method (MDP-ConvFDICA) according to the present invention will be described. Here, the result of the experiment which separates the sound mixed with one sound source and two sound receivers under reverberation will be described.

  The mixed signal is the RWCP real-world speech / acoustic database (8-second speech data selected from the newspaper reading speech corpus (The Japan Society of Acoustics. Newspaper article reading speech corpus. JNAS Vols. 1-16, 1997.)). Real World Computing Partnership, “RWCP sound scene database in real acoustic environments,” http://tosa.mri.co.jp/sounddb/index.htm) . Here, the arrangement of the impulse response sound source and the sound receiver is as shown in FIG. 7, the distance between the sound receivers is 2.83 cm, and the distance from the center of the sound receiver to the sound source is 2 m. As voice data, two mixed sentences by two men and women were used, and 12 mixed signals were created. The sampling frequency was 8 kHz.

  Here, in addition to the MDP-ConvFDICA of the present invention, as a comparative example, a natural gradient method (S. Amari, “Natural Gradient Works Efficiently in Learning”, Neural Computation, Vol. 10, pp.251-276 , 1998.), FastICA (A. Hyvarinen, J. Karhunen and E. Oja, "Independent Component Analysis", Jhon Wiley & Sons, pp.165-202, 2001, ISBN 0-471-40540), and ConvICA (non- A separation experiment according to Patent Document 4) was also performed. For the natural gradient method and FastICA, the frame length was changed to 128, 256, 512, 1024, and 2048 [point], the frame period was set to 1/4 of the frame length, and the Fourier transform was performed using the window function as the Hamming window. The length of the signal used for learning was 3 seconds and 8 seconds. For the natural gradient method, the initial load was a random number of [-1, 1], and the maximum number of updates was 100. For FastICA, the initial load was a random number of [-1, 1], the maximum number of updates was 1000, and the convergence criterion was that the absolute value of the inner product of the separation load before and after the update was 0.999999 or more.

  For ConvFDICA and the MDP-ConvFDICA of the present invention, Fourier transform was performed using a frame length of 128, a frame period of 1/4, and a window function as a Hamming window. In addition, the tap length of the separation matrix was changed to 4, 8, and 16 for each frame length. The length of the signal used for learning was 3 seconds. The update times for ConvFDICA tap lengths 4, 8, and 16 were 1000, 2000, and 4000, respectively, and the search step width was 0.007, 0.002, and 0.002.

  The maximum number of updates for the tap lengths 4, 8, and 16 of the MDP-ConvFDICA of the present invention is 1000 times, 2000 times, and 4000 times, respectively, the search step width η is 0.02, 0.006, and 0.002, and the internal ratio γ is 0.1. .

  In addition, ConvFDICA and MDP-ConvFDICA were speeded up by making blocks and calculating with FFT. The block size was the same as the tap length. The final separation matrix is a separation matrix with the number of updates that gives the best separation performance.

[1] Comparison of separation performance In this experiment, Noise Reduction Rate (NRR; ≡ output SNR [dB]-input SNR [dB]) is used as an evaluation scale of separation performance (see Non-Patent Document 5).
NRR is defined as the following equation (91), where the ratio of the target voice to the interference voice (noise) is SNR.

Here, OSNR _i and ISNR _i are i ≠ j at the output SNR and the input SNR, respectively. Further, g _ij is an i-row / j-column element of the product G (z) = W (z) A (z) of the separation matrix W (z) and the mixing matrix A (z). Figures 8 and 9 show the separation performance of each method. In FIG. 8, the horizontal axis indicates the tap length, and the vertical axis indicates NNR. In FIG. 9, the horizontal axis indicates the window width, and the vertical axis indicates NNR. From these figures, it can be seen that the NNR value increases in the order of FastICA, natural gradient method (NG), MDP-ConvFDICA, and ConvFDICA. In the case of FastICA and the natural gradient method (NG), whitening distortion does not occur because the instantaneous mixing model corresponds to the tap length of the narrowband separation filter being M = 1. In the case of MDP-ConvFDICA, it has been devised to prevent whitening distortion as described above. However, in the case of ConvFDICA, whitening distortion occurs as pointed out repeatedly above. Conventionally, NNR has been used as an evaluation quantity for measuring separation ability, but there is a drawback that NNR increases as whitening distortion increases. Therefore, SDR (Signal Distortion Ratio) for evaluating the separation ability including the influence of whitening distortion has been proposed as follows.

[2] Comparison of sound quality of separated signals In this experiment, SDR (Signal Distortion Ratio) is used as an evaluation scale for whitening distortion of separated signals (Tomoya Takatani, Takeki Nishikawa, Hiroshi Saruwatari, Kiyohiro Shikano, “SIMO model High-fidelity blind sound source separation used ", IEICE Technical Report, EA2002-108, pp.19-24, 2003). SDR is defined as in the following formula (92).

  This evaluates the sound quality of the separated signal, and a higher value means that the signal can be separated without distortion. 10 and 11 show the SDR of the separated signal. In FIG. 10, the horizontal axis indicates the tap length, and the vertical axis indicates SDR. In FIG. 11, the horizontal axis indicates the window width, and the vertical axis indicates the SDR. From these figures, FastICA and natural gradient method (Natural Gradient: NG) maintain high values, and MDP-ConvFDICA of the present invention does not show much deterioration, but ConvFDICA shows low values, and sound quality It can be seen that is deteriorated. From this, it can be seen that MDP-ConvFDICA has a higher SDR than ConvFDICA by about 6 dB and can faithfully reproduce the original sound.

  As described above, according to the high reverberation blind signal separation device and high reverberation blind signal separation method according to the present invention, the following effects can be obtained.

(1) By using a separation model in which the mixing process is strictly formulated in a convolution model in the frequency domain, signals can be well separated even under high reverberation.
(2) The distortion caused by whitening, which was a problem with TDICA and ConvFDICA, is eliminated.
(3) By using the initial value setting method in the present invention, an optimal solution without component replacement can be obtained. This solves the problem of component replacement, which has been a problem until now.
(4) The problem of component replacement and scaling that were unsolved in ConvICA is solved.

  In addition, the following points were confirmed by experiments that the anti-high reverberation blind signal separation method according to the present invention functions effectively under high reverberation.

(1) Since the impulse response is divided into frequency bins, the filter length of each frequency bin can be short even if the reverberation is long. As a result, TDICA based on TDICA and MDP has a long reverberation time, that is, the disadvantage of not converging if the filter length is long.
(2) Since the impulse response is divided into frequency bins, all reverberations can be included even with a short window width. Increasing the window width of FDICA alleviates the disadvantage that sufficient statistics cannot be obtained.
(3) By applying MDP to each frequency bin, the influence of distortion due to whitening can be avoided. Thereby, the influence of distortion due to whitening of ConvFDICA is reduced.
(4) It was confirmed that the separated signal by the MDP-ConvFDICA of the present invention is not deteriorated by whitening distortion, can reproduce the original sound almost faithfully, and the sound quality is improved by 6 dB compared to the ConvFDICA.

It is a figure showing time windowing discrete Fourier transform. It is a schematic diagram showing the state of the parameter curved surface in the case of instantaneous mixing and the case of convolutional mixing. It is a figure explaining the convergence of the solution by the setting of a suitable initial value. 1 is a diagram showing an overall configuration of a high reverberation blind signal separation device 1 according to Embodiment 1 of the present invention. FIG. FIG. 5 is a block diagram showing a configuration of subband separation means 3_n (n = 0,..., N−1) in FIG. 5 is a flowchart showing a flow of a high reverberation blind signal separation method according to Embodiment 1 of the present invention. It is a figure which shows arrangement | positioning of the sound receiver and sound source in experiment. It is a figure which shows the measurement result of NNR of FDICA. It is a figure which shows the measurement result of NNR of ConvFDICA and MDP-ConvFDICA. It is a figure which shows the measurement result of SDR of FDICA. It is a figure which shows the measurement result of SDR of ConvFDICA and MDP-ConvFDICA. It is a diagram illustrating the structure of a BSD in the case of _{R 1 = 2, R 2 =} 2. It is a figure showing the mixing process and separation process seen from TDICA. It is a figure explaining the problem of a component substitution. It is a figure showing the relationship between the analysis window width of a Fourier transform, and an impulse response.

Explanation of symbols

1 High reverberation blind signal separator
2 FFT means
3_0, 3_2, ..., 3_ (N-1) Subband separation means
4 IFFT measures
11 Separation matrix storage means
12 Separation calculation means
13 Separation matrix initialization means
14 Separation matrix update means
15 Separation matrix synthesis means
16 Component replacement and scaling means
17 Convergence judgment means

Claims (5)

Time-series data {x _i (ω _n , k)} (ω _n (n = 0,..., N−1)) is a normalized frequency, N is a window width, k (= 0, 1,...) Is a frame number, i (= 1,..., R) is a receiver number, and R is the number of receivers. A high-reverberation blind signal separation device that calculates a separation matrix for each frequency according to "ICA") and calculates a separation signal using the separation matrix,
Based on the observed signal, the separation matrix W (ω _n , p) without the component replacement of the matrix elements between the frequencies (ω _n is the normalized frequency, p is the filter tap number) (hereinafter referred to as “initial” A separation matrix initialization means for calculating a separation matrix ”),
Update formula of the separation matrix W (ω _n , p) by the convolution frequency domain ICA based on the principle of minimum distortion with the initial separation matrix as an initial value for the time series data for each frequency

Anti-high reverberation blind signal separation apparatus characterized by having a separation matrix updating means for updating the separation matrix W (ω _n, p) with. The separating matrix initialization means, based on said observed signal, resistance to high reverberation blind signal separation device according to claim 1, wherein the calculating the initial separation matrix without component replacement by frequency domain ICA instantaneous mixing. Time-series data {x _i (ω _n , k)} (ω _n (n = 0,..., N−1)) is a normalized frequency, N is a window width, k (= 0, 1,...) Is a frame number, i (= 1,..., R) is a receiver number, and R is the number of receivers. A high-reverberation blind signal separation method that calculates a separation matrix for each frequency according to "ICA") and calculates a separation signal using the separation matrix,
Based on the observed signal, the separation matrix W (ω _n , p) without the component replacement of the matrix elements between the frequencies (ω _n is the normalized frequency, p is the filter tap number) (hereinafter referred to as “initial” A separation matrix initialization procedure for calculating a separation matrix ”),
Update formula of the separation matrix W (ω _n , p) by the convolution frequency domain ICA based on the principle of minimum distortion with the initial separation matrix as an initial value for the time series data for each frequency

Anti-high reverberation blind signal separation method characterized by having a procedure separation matrix update for updating the separation matrix W (ω _n, p) with. 4. The anti-reverberation blind signal separation method according to claim 3, wherein, in the separation matrix initialization procedure, the initial separation matrix without component replacement is calculated based on the observed signal by a frequency domain ICA of instantaneous mixing. A program that, when executed by a computer, causes the computer to function as the high reverberation blind signal separation device according to claim 1 or 2 .

Images