JPWO2006132249A1 - Signal separation device - Google Patents

Abstract

Provided is a signal separation device that can quickly and reliably separate output signals output from different positions in a two-dimensional plane or a three-dimensional space with a simple processing process and device configuration. A superimposed signal in which output signals output from output sources 201 and 202 at different positions are superimposed is detected as an instantaneous mixed sum of time gradients related to each output signal, and each output signal is based on the instantaneous mixed sum of time gradients. Since the detected signal can be detected not as a scalar quantity (sound pressure) but as a vector quantity (sound pressure gradient), it is easy to separate each output signal along with the direction of each output source. Can be quickly and reliably separated by simple processing steps and apparatus configuration. [Selection] Figure 1

Description

  The present invention relates to a signal separation device that separates output signals before superposition from superposed signals on which output signals such as audio signals or radio wave signals outputted from different positions in a two-dimensional plane or three-dimensional space are superposed.

  While many people are speaking, humans can select and listen to specific voices. This human auditory ability is called the cocktail party effect. However, with many sound sources mixed, even if recording is performed using a normal microphone, only the necessary sound is efficiently collected due to the influence of the distance to the sound source, the direction of arrival of the sound source, the strength of the sound, etc. Is very difficult.

  One technique for restoring the original signal from a signal in which a plurality of signals are superimposed is blind signal separation. Blind signal separation is a technique for estimating a source signal based on the assumption that the source signal is statistically independent and the observed signal is linearly mixed by the source signal.

  Various methods have been proposed for blind signal separation, but convolution-type methods that take into account the observation of speech in a real environment have been studied. In recent years, research has been actively conducted on a separation method in which a signal is divided into narrowband signals and instantaneous mixed blind signal separation is performed for each band.

  Japanese Patent Laid-Open No. 2000-181499 discloses a sound source signal circuit related to these separation methods and a microphone device using the same.

A sound source separation / sound collection microphone device and method thereof are disclosed in Japanese Patent Laid-Open No. 2003-98003. Further, a sound source separation method, apparatus, and recording medium are disclosed in Japanese Patent Laid-Open No. 10-313497.
JP 2000-181499 A JP 2003-98003 A Japanese Patent Laid-Open No. 10-313497

In Patent Document 1 in the prior art, a plurality of mixed signals x ₁ (t) and x ₂ (t) obtained by linearly adding a plurality of linearly independent sounds (sound source signals) are divided into frames and divided for each frame. The sound (sound source signal) is separated from the plurality of sound sources by multiplying the inverse matrix of the mixing matrix that minimizes the correlation with zero lag time between the plurality of signals y ₁ (t) and y ₂ (t). Therefore, it is necessary to divide a plurality of mixed signals for each frame as a premise of the separation processing, and there are problems that the processing steps and the circuit configuration are complicated and the separation processing time is delayed.

  Similarly to Patent Document 1, Patent Document 2 has a similar problem because the mixed signal is divided for each sound source signal after being divided for each frame.

  Further, Patent Document 3 divides each channel signal from a plurality of microphones into a plurality of bands so that the main component is composed of only one sound source signal component. Since each sound source signal is separated by detecting time and determining and separating which sound source signal for each band from these, a plurality of sound source signals are composed of only one sound source signal. The division process cannot be executed unless the pre-processing for dividing into bands is performed, so that the processing steps and the circuit configuration are complicated, and the separation processing time is delayed.

  The present invention has been made to solve the above-described problems, and is a signal separation device capable of quickly and reliably separating output signals output from different positions in a two-dimensional plane or three-dimensional space with simple processing steps and device configurations. The purpose is to provide.

  The signal separation device according to the present invention is a signal separation device that separates superimposed signals on which output signals output from output sources at different positions in a two-dimensional plane or a three-dimensional space are superimposed. A spatial gradient detecting means for detecting an instantaneous mixture sum of time gradients with respect to each of the output signals, and directly capturing the signals detected by the spatial gradient detection means without storing them, and each of the output signals based on the instantaneous mixture sum of the time gradients of the output signals. And a signal separation means for separating the signal.

  As described above, according to the present invention, a superimposed signal in which output signals to be output from output sources at different positions are superimposed is detected as an instantaneous mixed sum of time gradients related to each output signal, and based on the instantaneous mixed sum of time gradients. Since each output signal is separated, the detected signal can be detected not as a scalar quantity (sound pressure) but as a vector (sound pressure gradient), and together with the direction of each output source, each output signal can be detected. There is an effect that the separation can be performed quickly and reliably by a simple processing process and apparatus configuration.

  Further, in the signal separation device according to the present invention, the signal separation unit includes an integrator for time-integrating the spatial gradient signal output from the gradient detection unit, as necessary.

  As described above, according to the present invention, the signal separation unit includes an integrator that temporally integrates the spatial gradient signal output from the gradient detection unit, and therefore noise in matrix operation at the subsequent stage of the spatial gradient detection unit. The effect which can reduce the bad influence by is produced.

  In addition, the signal separation device according to the present invention includes a first integrator that temporally integrates a spatial gradient signal output from the gradient detection unit, and a first integrator as necessary. Of the output time integration signals, an average calculator that calculates the average of two signals, and a time differentiator that calculates the gradient of the spatial gradient by spatial differentiation of the time integration signal output from the first integrator. And a second integrator for time-integrating the spatial gradient signal output from the spatial differentiator.

  According to the present invention, the signal separation means includes a first integrator for time-integrating the spatial gradient signal output from the gradient detection means, and a time integration signal output from the first integrator. An average calculator that calculates the average of two signals, a time integrated signal output from the first integrator, a spatial differentiator that calculates the gradient of the spatial gradient by spatial differentiation, and the spatial differentiator. Since the second integrator for time-integrating the spatial gradient signal is provided, there is an effect that the signal separation device can cope with up to four sound sources in the two-dimensional plane.

  The signal separation device according to the present invention is configured as a microphone array in which the spatial gradient detection unit detects each output signal as each audio signal, as necessary, and the microphone array serves as a sound pressure gradient of the audio signal. It is to detect.

  According to the present invention, the detection of the time gradient for each output signal is detected by the microarray as each sound signal, and the microphone array detects the sound pressure gradient of the sound signal. Along with the direction, each audio signal can be separated quickly and reliably by a simple processing process and apparatus configuration.

  Further, the signal separation device according to the present invention obtains the sound pressure gradient detected by the microphone array by time differentiation of the particle velocity, if necessary. According to the present invention, since the sound pressure gradient detected by the microphone array is obtained by time differentiation of the particle velocity, each output signal itself is estimated instead of time differentiation of each output signal. Each of these processes is unnecessary, and it is possible to achieve high-speed and reliable separation with a simpler process and apparatus configuration.

  In addition, the signal separation device according to the present invention is a bi-directional device in which the microphone array measures a sound pressure difference in at least the x-axis direction or the y-axis direction at an observation point with respect to different positions from which each audio signal is output, as necessary. It consists of a sex microphone.

  According to the present invention, since the microphone array is composed of a bidirectional microphone that measures a sound pressure difference at least in the x-axis direction or the y-axis direction at an observation point with respect to different positions from which each audio signal is output, Along with the direction of the sound source, each audio signal can be separated quickly and reliably with a simple processing process and apparatus configuration.

  Furthermore, in the signal separation device according to the present invention, the signal separation means separates the superimposed signal into output signals and outputs the direction and / or standard deviation of each output signal at each output source as necessary. Is.

  According to the present invention, the signal separation means separates the superimposed signal into output signals, and outputs the direction and / or standard deviation of each output signal at each output source. It has the effect of separating signals quickly and reliably with a simple processing process and apparatus configuration.

1 is an overall circuit configuration diagram of a signal separation device according to a first embodiment of the present invention. FIG. 2 is an arrangement configuration diagram of the bidirectional microphone described in FIG. 1. 2 is an operation flowchart of the signal separation device described in FIG. 1. It is a whole circuit block diagram in the signal separation apparatus which concerns on the 2nd Embodiment of this invention. 2 is an operation flowchart of the signal separation device described in FIG. 1. It is a whole circuit block diagram in the signal separation apparatus which concerns on the 3rd Embodiment of this invention. FIG. 7 is an arrangement configuration diagram of particle velocity microphones described in FIG. 6. It is an operation | movement flowchart of the signal separation apparatus of FIG. It is a block diagram of a confusion process / separation process for demonstrating the arithmetic logic of each embodiment of this invention. It is an audio | voice waveform diagram used for the numerical experiment and acoustic experiment in each embodiment of this invention. It is a power spectrum of two source signals of each embodiment of the present invention. It is a time gradient characteristic view of the source signal of each embodiment of the present invention. It is a space gradient figure of the observation signal in each embodiment of the present invention. It is a characteristic view of the separation signal in each embodiment of the present invention. It is a characteristic view of the separation signal in each embodiment of the present invention. It is a characteristic view of the separation signal in each embodiment of the present invention. It is a block diagram of the spatial integration microphone used for the acoustic experiment of this invention. It is an experiment schematic diagram of an acoustic experiment of the present invention. It is a spatial gradient characteristic view of the observation signal in the acoustic experiment of the present invention. It is a spatial gradient characteristic view of the observation signal in the acoustic experiment of the present invention. It is a wave form diagram of a separated signal in an acoustic experiment of the present invention. It is a characteristic view of the arrival direction in the acoustic experiment of this invention. It is a whole circuit block diagram of the signal separation apparatus which concerns on the 4th Embodiment of this invention. It is explanatory drawing for demonstrating the omnidirectional microphone unit in the omnidirectional microphone shown in FIG. It is explanatory drawing for demonstrating the directivity characteristics of the x-axis direction by the 1st space differentiator and 1st integrator among the signal separation apparatuses shown in FIG. 23, and a y-axis direction. FIG. 24 is an arrangement configuration diagram of the omnidirectional microphone described in FIG. 23. It is an operation | movement flowchart of the signal separation apparatus of FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Bidirectional microphone 10 Particle velocity microphone 11 X-axis direction bidirectional microphone 12 Y-axis direction bidirectional microphone 11a, 12a, 111a, 111b Directivity 2 Spatial differentiator 21 x-axis direction spatial differentiator 22 y-axis direction space Differentiator 3, 30, 103 Matrix operation circuit 4, 104 Separation matrix element calculation circuit 5, 105 Direction / standard deviation calculation circuit 6, 7 Integrator 61 X-axis time integrator 62 Y-axis time integrator 100 Observation point 101 Non-directional Directional microphones 101a, 101b, 101c, 101d omnidirectional microphone units 102, 102a first spatial differentiator 107, 107a first integrator 108 average calculator 109 second spatial differentiator 110 second integrator 201, 202, 203 and 204 source _{P 1 (t), P 2} (t), P 3 (t), P 4 (t) source signal x (t), fy (t) sound pressure gradient delivery No.

(First embodiment of the present invention)
Hereinafter, a signal separation device according to a first embodiment of the present invention will be described with reference to FIGS. 1 is an overall circuit configuration diagram of the signal separation device according to the present embodiment, FIG. 2 is an arrangement configuration diagram of the bidirectional microphone described in FIG. 1, and FIG. 3 is an operation flowchart of the signal separation device described in FIG. Show.

In each of the drawings, the signal separation device according to the present embodiment collects both the source signals P ₁ (t) and P ₂ (t) from the sound sources 201 and 202 existing at different positions in a homogeneous three-dimensional space. Directional microphone 1, spatial differentiator 2 for obtaining a sound pressure gradient related to source signals P ₁ (t) and P ₂ (t) collected by bidirectional microphone 1, and the obtained sound pressure gradient signal A matrix operation circuit 3 that multiplies the vector amount (sound pressure gradient) of fx (t) and fy (t) by the inverse matrix of the mixing matrix, and the matrix operation value is estimated by the blind signal separation to estimate the separation matrix element a′ij. The separation matrix element calculation circuit 4 for estimating and calculating the estimated separation signal Pti (t); i = 1, 2 and the arrival directions θ ₁ and θ ₂ (x of the sound sources 201 and 202 based on the separation matrix element a′ij) Direction to calculate the standard deviations σti and σtj of the time gradient of the sound sources 201 and 202) Quasi deviation calculating circuit 5, the estimated separation signal Pti (t) separated signal P ₁ as the time integral (t), it is configured to include an integrator 6 that calculates the P ₂ (t).

The bidirectional microphone 1 includes an x-axis bidirectional microphone 11 that measures a sound pressure difference in the x-axis direction from the vicinity of the observation point 100, and a y-axis that measures a sound pressure difference in the y-axis direction from the vicinity of the observation point 100. The directional bidirectional microphone 12 is provided. The x-axis direction bidirectional microphone 11 has directivity 11a used for spatial differentiation in the x-axis direction at the observation point 100 as shown in FIGS. The y-axis direction bidirectional microphone 12 has a directivity 12 a used for spatial differentiation in the y-axis direction, like the x-axis direction bidirectional microphone 11. The spatial differentiator 2 includes an x-axis direction spatial differentiator 21 for _obtaining a spatial gradient (f _x1 −f _x2 ) / Δx in the x-axis direction and a spatial gradient (f _y1 −f _y2 ) / Δx in the y-axis direction. The y-axis direction spatial differentiator 22 is obtained.

Next, the separation processing operation of the signal separation device according to the present embodiment based on the above configuration will be described. First, the source signals P ₁ (t) and P ₂ (t) from the sound sources 201 and 202 arranged in the homogeneous three-dimensional space are converted into the x-axis direction bidirectional microphone 11 and the y-axis direction near the observation point 100. Sound is picked up by the bidirectional microphone 12 (step 1). The collected source signals P1 (t) and P2 (t) are converted into spatial gradients fx and fy by the spatial differentiation in the x-axis and y-axis directions by the x-axis direction spatial differentiator 21 and the y-axis direction spatial differentiator 22, respectively. Based on the spatial gradients fx and fy, the matrix calculation circuit 3 multiplies the vector amount (sound pressure gradient) of the spatial gradient signals fx (t) and fy (t) by the inverse matrix of the mixing matrix (step 2). .

The matrix calculation circuit 3 determines whether or not there is a signal input from the bidirectional microphone 1, the spatial differentiator 2, and the matrix calculation circuit 3 which are input devices in the previous stage (step 3). If it is determined that there is a separation matrix element a′ij by blind signal separation (step 4), the estimated separation signals P _t1 (t) and P _t2 (t) are calculated (step 5).

The estimated separation signals P _t1 (t) and P _t2 (t) based on the separation matrix element a′ij are time-integrated by the x-axis time integrator 61 and the y-axis time integrator 62, respectively, and the separation signal P ₁ (t) , P ₂ (t) is calculated (step 6), and the calculated separated signals P ₁ (t) and P ₂ (t) are output to a display device (not shown) (step 7).

On the other hand, based on the separation matrix element a′ij estimated and calculated in step 4, the direction / standard deviation calculation circuit 5 determines the arrival directions θ ₁ and θ _{2 of the} source signals P ₁ (t) and P ₂ (t) and the sound source. The standard stitching σti and σtj in the time gradients 201 and 202 are estimated and calculated (step 8). The arrival directions θ ₁ and θ ₂ of the source signals P ₁ (t) and P ₂ (t) at the observation point 100 estimated by the direction / standard deviation calculating circuit 5 and the standard deviations σ ₁ and σ of the sound sources 201 and 202 are calculated. ₂ is output to the display device or the like (step 9).

As described above, the signal separation device according to the present embodiment is based on the spatiotemporal gradient analysis, and in a homogeneous space, the spatial gradient at an arbitrary observation point 100 is expressed by a linear mixture of temporal gradients at the sound sources 201 and 202. Take advantage of the fact that That is, the features of this embodiment are summarized as follows. By measuring the spatial gradient at only one observation point 100 and in the vicinity thereof, the instantaneous mixed sum of the time gradients of the source signals P ₁ (t) and P ₂ (t) can be acquired. Therefore, the problem can be reduced to the simplest instantaneous mixed blind signal separation problem. Furthermore, since the observed quantity is not a scalar quantity (sound pressure) but a vector quantity (spatial gradient of sound pressure), the source signals P ₁ (t) and P ₂ (t) including the arrival direction of the sound source can be separated. . By adopting this spatiotemporal gradient analysis method, it is not necessary to consider the difference in arrival time of signals generated between the observation points 100, and the spatial gradient of the observation signal can be regarded as an instantaneous linear mixed signal. The separation signals P ₁ (t) and P ₂ (t) can be estimated by processing for linear mixing.

(Second embodiment of the present invention)
Hereinafter, a signal separation device according to a second embodiment of the present invention will be described with reference to FIGS. 4 is an overall circuit configuration diagram of the signal separation device according to the present embodiment, and FIG. 5 is an operation flowchart of the signal separation device described in FIG.

  In each of the drawings, the signal separation apparatus according to the present embodiment is similar to the first embodiment in that the bidirectional microphone 1, the spatial differentiator 2, the matrix operation circuit 3, the separation matrix element calculation circuit 4, and the direction / standard deviation The calculation circuit 5 is provided in common, and the sound pressure gradient signals fx (t) and fy (t) output from the spatial differentiator 2 are time-integrated, and the integrated sound pressure gradient signals fx (t) and fy ( The configuration includes an integrator 7 (corresponding to the integrator 6 in the first embodiment) that outputs t) to the matrix operation circuit 3.

  Here, normally, noise is superimposed on the observation signal at the observation point 100, and when the spatial differentiator 2 simply obtains the spatial gradient by subtraction, the noise is emphasized, and the matrix calculation circuit 3 in the subsequent stage is used. In some cases, a large error may be caused in the calculation. For this reason, in order to remove this noise, a means for applying a filter may be considered. However, depending on the type of filter, the condition of the linear instantaneous mixing necessary for the operation of this system may be broken. .

  Therefore, in the signal separation device according to the second embodiment, by disposing the integrator 7 in the preceding stage of the matrix operation circuit 3, the latter stage of the spatial differentiator 2 can be achieved without causing the above-described inappropriate action. The adverse effect of noise in the matrix operation can be reduced.

  Next, the separation processing operation of the signal separation device according to the present embodiment based on the above configuration is processed in substantially the same manner as in the first embodiment, but the x-axis direction spatial differentiator 21 and the y-axis direction are processed. After calculating the spatial gradients fx and fy by the spatial differentiation in the x-axis and y-axis directions by the spatial differentiator 22 (step 11), the spatial gradients fx and fy are time-integrated by the integrator 7 (step 12).

The spatial gradient signals fx (t) and fy (t) of the time-integrated spatial gradients fx and fy are multiplied by the inverse matrix of the mixing matrix by the matrix operation circuit 3. Thereafter, separation signals P ₁ (t) and P ₂ (t) are output in steps 13 to 18 as in the first embodiment, and source signals P ₁ (t) and P ₂ (t) at the observation point 100 are output. and outputs the arrival direction theta ₁ of the standard deviation sigma ₁ of theta ₂ and the sound source 201 and 202, the sigma ₂ to the display device or the like

(Third embodiment of the present invention)
A signal separation device according to a third embodiment of the present invention will be described with reference to FIGS. 6 is an overall circuit configuration diagram of the signal separation device according to the present embodiment, FIG. 7 is an arrangement configuration diagram of the particle velocity microphone described in FIG. 6, and FIG. 8 is an operation flowchart of the signal separation device described in FIG. .

In each of the drawings, the signal separation device according to the present embodiment is similar to the first embodiment in that the matrix operation circuit 30 (corresponding to the matrix operation circuit 3 in the first embodiment), the separation matrix element calculation circuit 4 and the direction A standard deviation calculation circuit 5 is provided in common, and a particle velocity microphone 10 that measures source signals P ₁ (t) and P ₂ (t) from the sound sources 201 and 202 as particle velocities Vx and Vy is provided. In this configuration, the particle velocity signals Vx (t) and Vy (t) measured by the microphone 10 are output to the matrix operation circuit 30.

Next, in the signal separation device separation processing operation according to the present embodiment based on the above configuration, first, the source signals P ₁ (t) and P ₂ (t) from the sound sources 201 and 202 are converted into the particle velocity Vx, It is measured as Vy (step 20). It is determined whether or not particle velocity signals Vx (t) and Vy (t) based on the particle velocities Vx and Vy from the input device including the particle velocity microphone 10 are output (step 21).

If it is determined in step 21 that the particle velocity signals Vx (t) and Vy (t) are output, the subsequent particle velocity signals Vx (t) and Vy (t) are mixed by the matrix operation circuit 30. Multiply by matrix inverse. Thereafter, as in the second embodiment, the separation signals P ₁ (t) and P ₂ (t) are output in steps 22 through 26 and the source signals P ₁ (t) and P ₂ ( The arrival directions θ ₁ and θ _{2 of} t) and the standard deviations σ ₁ and σ ₂ of the sound sources 201 and 202 are output to the display device or the like.

(Operational logic of each embodiment of the present invention)
The arithmetic logic of each embodiment will be described below with reference to FIG. 2 of the first embodiment based on FIG.
2 and 9, in the signal separation device according to each embodiment of the present invention, the spatial gradient at the observation point 100 is expressed by a linear combination of the time gradients of the source signal by the wave equation. In the spatiotemporal gradient method of the wave field, the sound pressure satisfies the following wave equation in a far field that does not include the sound source 201 and the sound source 202.

ここでｃは位相速度ベクトルである。式(1)は逆向きに進行する２つの波面の存在を示している。ここでは片方の波面に着目し次の移流方程式に注目する。

Here, c is a phase velocity vector. Equation (1) shows the existence of two wavefronts traveling in opposite directions. Here, we focus on one wavefront and focus on the following advection equation.

この方程式は、観測点１００における音圧の時間勾配と空間勾配の線形関係を示している。均質な空間を仮定すると、ある観測点での音圧の時間勾配は、各音源２０１、２０２における伝播遅延時間過去の音圧の時間勾配値の総和で表される。本発明は式(2)で示される移流型の方程式を満たす波動場において、時間勾配と空間勾配の線形関係を利用する時空間勾配法をブラインド信号分離問題に適用させる。

This equation shows a linear relationship between the time gradient of sound pressure at the observation point 100 and the spatial gradient. Assuming a homogeneous space, the time gradient of the sound pressure at a certain observation point is represented by the sum of the time gradient values of the sound pressures past the propagation delay times in the sound sources 201 and 202. The present invention applies a spatiotemporal gradient method using a linear relationship between a temporal gradient and a spatial gradient to a blind signal separation problem in a wave field satisfying an advection type equation represented by equation (2).

  Next, for wavefront superposition and instantaneous mixing, for the sake of simplicity, two independent plane waves traveling on the same plane are assumed. The sound pressure at any point is given by:

ここでｃは音速、θ₁、θ₂は各波面の到来方向を示す。関数Ｐ₁(t)、Ｐ₂(t)は帯域幅ωの信号の音圧であり、次式で定義する。

Here, c represents the speed of sound, and θ ₁ and θ ₂ represent the arrival directions of the wavefronts. The functions P ₁ (t) and P ₂ (t) are the sound pressures of the signal of the bandwidth ω and are defined by the following equations.

任意の点で得られる情報は観測信号ｆ（ｘ，ｙ，ｔ）のみであり、元の信号や音源の到来方向は未知である。このような観測信号から源信号Ｐ₁(t)、Ｐ₂(t)を推定するために、時空間勾配法を適用し、観測信号の空間勾配を源信号Ｐ₁(t)、Ｐ₂(t)の時間勾配の線形混合信号として表現する。原点０（観測点１００）における時間勾配は次式で得られる。

Information obtained at an arbitrary point is only the observation signal f (x, y, t), and the arrival direction of the original signal and the sound source is unknown. In order to estimate the source signals P ₁ (t) and P ₂ (t) from such observed signals, the spatiotemporal gradient method is applied, and the spatial gradient of the observed signals is determined as the source signals P ₁ (t) and P ₂ ( Expressed as a linear mixed signal with a time gradient of t). The time gradient at the origin 0 (observation point 100) is obtained by the following equation.

ここでＰ_ti(t)、Ｐ_i(t)の時間勾配である。

Here, it is a time gradient of P _ti (t) and P _i (t).

式（５）を用いると、ｆ（ｘ，ｙ，ｔ）の音圧勾配は次のように導出される。

Using equation (5), the sound pressure gradient of f (x, y, t) is derived as follows.

すなわちｆ（ｘ，ｙ，ｔ）の空間勾配はＰ_t1(t)、Ｐ_t2(t)を源信号とした場合の瞬時線形混合信号として表される。

That is, the spatial gradient of f (x, y, t) is expressed as an instantaneous linear mixed signal when P _t1 (t) and P _t2 (t) are used as source signals.

ここで行列Ａを混合行列として、次式で定義する。

Here, the matrix A is defined as a mixing matrix by the following equation.

また、信号の分離と音源方向の推定については、瞬時線形混合型ブラインド信号分離問題の概略を図９（Ａ）、（Ｂ）に示す。（Ａ）は混合過程を示し、（Ｂ）は分離過程を示す。（Ｂ）に示されるように、分離過程は観測された２信号をＷで無相関化する白色化過程とＲ（η）で独立した成分に分離する回転過程から構成される。白色化後、分離行列を求めるため、分離信号Ｐ_t1(t)、Ｐ_t2(t)と観測信号の音圧勾配は次の関係式を満たす。

As for signal separation and sound source direction estimation, an outline of the instantaneous linear mixed blind signal separation problem is shown in FIGS. (A) shows the mixing process, and (B) shows the separation process. As shown in (B), the separation process is composed of a whitening process in which two observed signals are decorrelated with W and a rotation process in which R (η) is separated into independent components. In order to obtain a separation matrix after whitening, the separation signals P _t1 (t) and P _t2 (t) and the sound pressure gradient of the observation signal satisfy the following relational expression.

ここで分離行列を

Where the separation matrix

とおく。Ｒ（η）とＷは後述するブラインド信号分離のアルゴリズムによって決定されるが、混合行列の逆行列Ａ^-1を用いて次式でも表わすことができる。

far. R (η) and W are determined by the blind signal separation algorithm described later, but can also be expressed by the following equation using the inverse matrix A ⁻¹ of the mixing matrix.

ここでσ_t1とσ_t2はそれぞれ観測点Ｐ_t1(t)、Ｐ_t2(t)におけるの標準偏差である。式（１３)においてＡ′の要素とＡ^-1の各要素を比較することで、式（１４）、（１５）の推定式を得る。

Here, σ _t1 and σ _t2 are standard deviations at the observation points P _t1 (t) and P _t2 (t), respectively. By comparing the element of A ′ and each element of A ⁻¹ in Expression (13), the estimation expressions of Expressions (14) and (15) are obtained.

このとき分離信号と各パラメータＰ_ti(t)、θ_i、σ_ti、ｉ＝１、２は次の２通りのどちらかで推定される。

At this time, the separated signal and the parameters P _ti (t), θ _i , σ _ti , i = 1, 2 are estimated in one of the following two ways.

すなわち分離信号を推定するのと同時に、個々の音源の到来方向、標準偏差も推定することができる。

That is, the arrival direction and standard deviation of each sound source can be estimated simultaneously with the estimation of the separated signal.

  Furthermore, for blind signal separation for instantaneous mixing, the sound pressure gradient of the observed signal is whitened by the matrix W, and the separated signal is estimated by rotationally transforming the output signals so that they are independent of each other by the matrix R (η). To do. About whitening, it can obtain | require by following Formula.

回転変換については、次式で求めることができる。

About rotation conversion, it can obtain | require by following Formula.

ここでηは分離パラメータであり、単位は度(°)である。最適なηを推定するために次式の評価関数を導入する。

Here, η is a separation parameter, and the unit is degree (°). In order to estimate the optimum η, the following evaluation function is introduced.

ここで関数Ｈ₄(・)は４次のHermite多項式である。

Here, the function H ₄ (•) is a fourth-order Hermite polynomial.

勾配法を適用したηを逐次的に推定する式は次式となる。

An equation for sequentially estimating η to which the gradient method is applied is as follows.

ここでμは適当な定数である。

Here, μ is an appropriate constant.

(Numerical experiment in the present invention)
Through the numerical experiment in the present invention, the direction of the sound source is estimated using blind signal separation. The speech waveform shown in FIG. 10 is a source signal used in numerical experiments and acoustic experiments. The audio signals of two different women for 10 seconds are used as the source signals P ₁ (t) and P ₂ (t). Table 1 shows the standards of the source signal. Looking at FIG. 11 showing the power spectra of these two source signals, it can be seen that each has a component mainly in the same band (around 200 to 350 Hz). In the present invention, blind signal separation is applied to the spatial gradient of the observation signal formed by the time gradient of the source signal created based on the Fourier series expansion. In the blind signal separation used in the present invention, in order to stabilize the moment used in the whitening process, the estimation of the separation matrix starts one second after the moment calculation starts. Further, in order to obtain the source signal by time-integrating the estimated separated signal, the following equation is adopted.

表２に数値実験の設定値を示す。

Table 2 shows the setting values for the numerical experiment.

これは以降に示す音響実験と同様の設定にしている。源信号の時間勾配を図１２に、観測信号の空間勾配ｆｘ(t)、ｆｙ(t)を図１３に、分離信号Ｐ_t1(t)、Ｐ_t2(t)とそれらの時間積分、逐次的に推定したθ₁、θ₂を図１４、図１５、図１６にそれぞれ示す。数値実験の初期値として分離パラメータの初期値数η（t₀）＝０°、ｔ₀＝１sec、収束係数μ＝0.001を与えている。図１４より、推定開始直後は波形が乱れているが、ｔ＝２secには元の波形はほぼ復元されている。このとき推定した分離行列Ｒ（η）Ｗを次式に示す。

This is the same setting as the acoustic experiment described below. FIG. 12 shows the time gradient of the source signal, FIG. 13 shows the spatial gradients fx (t) and fy (t) of the observed signal, separation signals P _t1 (t) and P _t2 (t) and their time integration, The estimated θ ₁ and θ ₂ are shown in FIGS. 14, 15, and 16, respectively. As initial values of the numerical experiment, an initial number of separation parameters η (t ₀ ) = 0 °, t ₀ = 1 sec, and a convergence coefficient μ = 0.001 are given. From FIG. 14, the waveform is disturbed immediately after the start of estimation, but the original waveform is almost restored at t = 2 sec. The separation matrix R (η) W estimated at this time is shown in the following equation.

音源の到来方向を推定した図１６（Ａ）では、ｔ＝２secにはθ₁、θ₂は徐々に安定し始め、ｔ＝３sec以降は収束している。その後源信号Ｐ_t1、Ｐ_t2の振幅が同時に急激な変化をしているため、ｔ＝７sec付近に多少乱れが生じているが、再び真値付近に収束している。また最終的に推定された音源の到来方向はθ₁＝６０．８°、θ₂＝１２０．１°であった。標準偏差はσ₁＝１１５３５．９、σ₂＝１１８３５．２であり、表1の真値とほぼ一致する。次に分離パラメータの初期値η（t₀）＝２７０°を与えた場合、分離行列Ｒ（η）Ｗは次式となった。

In FIG. 16A in which the arrival direction of the sound source is estimated, θ ₁ and θ ₂ begin to stabilize gradually at t = 2 sec, and converge after t = 3 sec. After that, since the amplitudes of the source signals P _t1 and P _t2 change rapidly at the same time, some disturbance is generated in the vicinity of t = 7 sec, but it converges again in the vicinity of the true value. The finally estimated direction of arrival of the sound source was θ ₁ = 60.8 ° and θ ₂ = 120.1 °. The standard deviations are σ ₁ = 11535.9 and σ ₂ = 11835.2, which are almost the same as the true values in Table 1. Next, when the initial value η (t ₀ ) = 270 ° of the separation parameter is given, the separation matrix R (η) W is expressed by the following equation.

図１６（Ｂ）には推定した音源の方向を示す。このとき最終的に推定された音源の方向はθ₁＝１２０．１°、θ₂＝６０．８°標準偏差はσ₁＝１１８３５．２、θ₂＝１１５３５．９であった。初期値η（t₀）＝０°の場合、分離信号は式（１６）のように推定されるが、初期値η（t₀）＝２７０°をとると、ブラインド信号分離で得られる分離行列の成分が変化し、全ての推定値が式(１７)のように逆に推定された。このように本手法ではブラインド信号分離によって分離行列がどのように推定されたとしても、出力される分離信号と推定されたそれらの音源２０１、２０２の到来方向、標準偏差を適宜に推定できることが確認された。

FIG. 16B shows the estimated direction of the sound source. At this time, the direction of the sound source finally estimated was θ ₁ = 120.1 °, θ ₂ = 60.8 °, and the standard deviation was σ ₁ = 11835.2 and θ ₂ = 11535.9. When the initial value η (t ₀ ) = 0 °, the separation signal is estimated as shown in Equation (16), but when the initial value η (t ₀ ) = 270 °, the separation matrix obtained by blind signal separation is obtained. The components of were changed, and all estimated values were estimated in reverse as shown in equation (17). As described above, in this method, it is confirmed that no matter how the separation matrix is estimated by the blind signal separation, it is possible to appropriately estimate the arrival directions and standard deviations of the sound sources 201 and 202 estimated as the separation signals to be output. It was done.

  For mutual information, in order to confirm the statistical independence between signals before and after mixing or separation processing, mutual information

を算出する。相互情報量は常に正であり，Ｚ₁とＺ₂が互いに独立のとき０になる。ここでｐ（Ｚ₁）、ｐ（Ｚ₂）は周辺確率密度関数、ｐ（Ｚ₁，Ｚ₂）は結合確率密度関数を表わす。各信号の振幅値を＋側と−側それぞれ６４諧調ずつと０を含む１２９諧調に量子化したものから確率密度関数を求め、相互情報量を算出した。相互情報量は分離信号が安定して得られているｔ＝２〜１０secの区間のものをそれぞれ求めている。各信号間の相互情報量を示した表３をみると、源信号間の相互情報量Ｉ（Ｐ₁(t)、Ｐ₂(t)）は０．０４８ｂｉｔであり、分離によってこの値を下回ることはない限界の値を意味する。表３より、相互情報量における観測信号の空間勾配からのそれぞれの距離の比:

Is calculated. The mutual information amount is always positive and becomes 0 when Z ₁ and Z ₂ are independent of each other. Here, p (Z ₁ ) and p (Z ₂ ) represent a marginal probability density function, and p (Z ₁ , Z ₂ ) represents a joint probability density function. A probability density function was obtained from the quantized amplitude value of each signal in 129 gradations including 0 and 64 gradations in each of the + side and the − side, and a mutual information amount was calculated. The mutual information is obtained for each interval of t = 2 to 10 sec where the separated signal is stably obtained. Looking at Table 3 showing the mutual information amount between the signals, the mutual information amount I (P ₁ (t), P ₂ (t)) between the source signals is 0.048 bits, which is lower than this value due to separation. It means no limit value. From Table 3, the ratio of each distance from the spatial gradient of the observed signal in mutual information:

を比較するとｄ〜−９９．０％となり、得られた分離信号は分離処理によってほぼ統計的独立になったと考えられる。

Are d to -99.0%, and it is considered that the obtained separated signal is almost statistically independent by the separation process.

（本発明における音響実験）
本発明における音響実験において、本発明における空間微分マイクロホンを用いて観測信号の空間微分を実現するために、観測点１００を中心としてｘ軸、ｙ軸各方向に所定間隔離反させて２対を配設させて形成されるマイクロホンを導入する。空間微分マイクロホン本体中央の点（０，０）で得られる観測信号をｆ（０，０，ｔ）とすると、その前後左右に位置するマイクロホンで得られるｆ₁（ｘ、ｙ、ｔ）〜ｆ₄（ｘ、ｙ、ｔ）はそれぞれ次のように表わされる。

(Acoustic experiment in the present invention)
In the acoustic experiment according to the present invention, in order to realize the spatial differentiation of the observation signal using the spatial differential microphone according to the present invention, two pairs are arranged with a predetermined separation in the x-axis and y-axis directions around the observation point 100. Introducing a microphone formed. If the observation signal obtained at the center point (0, 0) of the spatial differential microphone main body is f (0, 0, t), f ₁ (x, y, t) to f obtained by microphones positioned on the front, back, left and right of the observation signal. ₄ (x, y, t) is expressed as follows.

これらの関係式より差分法を用いた観測信号のｘ，ｙ方向の空間微分は、

From these relational expressions, the spatial differentiation in the x and y directions of the observed signal using the difference method is

で表される。ここで△はマイクロホン間の距離であり、対象の信号の波長はこの距離よりも十分に長いものとする。製作した図１７のマイクロホンにおいて、△は１８ｍｍである。

It is represented by Here, Δ is the distance between the microphones, and the wavelength of the signal of interest is sufficiently longer than this distance. In the manufactured microphone of FIG. 17, Δ is 18 mm.

  Regarding noise reduction by time integration, signals observed in a real environment cannot be subjected to accurate separation processing, sound source arrival direction, and standard deviation estimation due to the influence of reflection. Therefore, the time integration performed for the separated signal is performed for the spatial gradient of the observed signal. That is, a separated signal is obtained by the following processing.

ここで忘却係数はα＝ｅ^-△^t/γ〜−０．９９９１であり、時定数γは２５ｍｓに設定している。このときの分離過程を図１７に示す。

Here, the forgetting factor is α = e ⁻ Δt ^/ γ˜−0.99991, and the time constant γ is set to 25 ms. The separation process at this time is shown in FIG.

For signal separation and estimation of sound source direction in a real environment, we conducted signal separation experiments using two speakers and a spatial differential microphone. A schematic diagram of the experiment is shown in FIG. The audio signal P ₁ (t) was simultaneously reproduced from the right speaker, and P ₂ (t) was simultaneously reproduced from the left speaker, and recording was performed using a spatial differential microphone. However, since four voices cannot be recorded simultaneously in this system, recording is performed at the positions of the microphones f _{1 to} f ₄ , and separation processing is performed on the voice signals recorded at the same time. The sound pressure gradients fx (t) and fy (t) of the observation signal and their time integration values are estimated in FIGS. 19 and 20, and the separated signals P ₁ (t) and P ₂ (t) are estimated in θ ₁ in FIG. , Θ ₂ are shown in FIG. As initial values of the experiment, initial values of separation parameters η (t ₀ ) = 270 °, t ₀ = 1 sec, and a convergence coefficient μ = 0.0001 are given. The separation matrix estimated by blind signal separation at this time is shown in the following equation.

推定した音源の到来方向θ₁、θ₂はともにｔ＝１．５secには安定し、ｔ＝３sec以降にはほぼ真値付近に収束しており、精度良く推定されている。このとき最終的に推定された音源の到来方向はθ₁＝６３．４°、θ₂＝１２６．４°であった。標準偏差はσ_t1＝３４５９．２、σ_t2＝３４８３．３であり、σ_t1：σ_t2はほぼ１：１でとなるので、表1の真値の比と一致する。また分離信号を音声に変換して聞いてみると、多少相手側の音声が重畳しているが、所望の信号は強調されていた。表３，４より、この実験の相互情報量も式(２８)を用いると、ｄ〜−８５．３％とすることができた。

The estimated directions of arrival θ ₁ and θ ₂ of the sound source are both stable at t = 1.5 sec, converged to near the true value after t = 3 sec, and are estimated with high accuracy. At this time, the direction of arrival of the sound source finally estimated was θ ₁ = 63.4 ° and θ ₂ = 126.4 °. The standard deviations are σ _t1 = 3459.2 and σ _t2 = 3483.3, and σ _t1 : σ _t2 is almost 1: 1, which is consistent with the true value ratio in Table 1. Also, when the separated signal was converted into sound and heard, the other party's sound was somewhat superimposed, but the desired signal was emphasized. From Tables 3 and 4, the mutual information amount of this experiment was also d to −85.3% when the equation (28) was used.

（本発明の第４の実施形態）
以下、本発明の第１の実施形態に係る信号分離装置を、図２３ないし図２７に基づいて説明する。この図２３は本実施形態に係る信号分離装置の全体回路構成図、図２４は図２３に示す無指向性マイクロホンにおける無指向性マイクロホンユニットを説明するための説明図、図２５は図２３に示す信号分離装置のうち第１の空間微分器と第１の積分器によるｘ軸及びｙ軸方向の指向特性を説明するための説明図、図２６は図２３に記載する無指向性マイクロホンの配置構成図、図２７は図２３に記載する信号分離装置の動作フローチャートを示す。

(Fourth embodiment of the present invention)
Hereinafter, a signal separation device according to a first embodiment of the present invention will be described with reference to FIGS. 23 is an overall circuit configuration diagram of the signal separation device according to the present embodiment, FIG. 24 is an explanatory diagram for explaining an omnidirectional microphone unit in the omnidirectional microphone shown in FIG. 23, and FIG. 25 is shown in FIG. FIG. 26 is an explanatory diagram for explaining the directivity characteristics in the x-axis and y-axis directions by the first spatial differentiator and the first integrator in the signal separator, and FIG. 26 is an arrangement configuration of the omnidirectional microphone shown in FIG. FIGS. 27A and 27B show an operation flowchart of the signal separation device shown in FIG.

In each of the drawings, the signal separation device according to the present embodiment is similar to the second embodiment in that the matrix operation circuit 103 (corresponding to the matrix operation circuit 3 in the second embodiment), the separation matrix element calculation circuit 104 ( 2) and a direction / standard deviation calculation circuit 105 (corresponding to the direction / standard deviation calculation circuit 5 in the second embodiment). Non-directional to pick up the source signals P ₁ (t), P ₂ (t), P ₃ (t), P ₄ (t) from the sound sources 201, 202, 203, 204 existing at different positions in the dimension plane Directional microphone 101 and a _first sound pressure gradient relating to source signals P ₁ (t), P ₂ (t), P ₃ (t), and P ₄ (t) collected by this omnidirectional microphone 101 Spatial differentiator 102 and the sound pressure gradient signal f output from the first spatial differentiator 102 _{1 (t), fx 2 (} t), fx 3 (t), fx 4 (t), fy 1 (t), fy 2 (t), fy 3 (t), integrating fy ₄ a (t) time The obtained time integration signals v _x1 (t), v _x2 (t), v _x3 (t), v _x4 (t), v _y1 (t), v _y2 (t), v _y3 (t), v Of _y4 (t), v _x1 (t), v _x4 (t), v _y1 (t), and v _y4 (t) are _input to the average calculator 108, and v _x1 (t), v _x2 (t), v The first outputs _x3 (t), _vx4 (t), _vy1 (t), _vy2 (t), _vy3 (t), and _vy4 (t) to the second spatial differentiator 109, respectively. Integrator 107 and time integration signals v _x1 (t), v _x4 (t), v _y1 (t) output from the first integrator 107,
v _y4 of (t), v _x1 (t) and v _x4 the average of (t) and v _x (t), v _y1 (t) and v _y4 (t) an average of the v _y (t) And the time integration signals v _x1 (t), v _x2 (t), v _x3 (t), v _x4 (v _x1 (t) output from the average calculator 108 output to the matrix calculation circuit 103 and the first integrator 107. t), v _y1 (t), v _y2 (t), v _y3 (t), v _y4 (t), v _x2 (t) and v _x3 (t), v _x1 (t) and v _x4 ( t), v _y2 (t) and v _y3 (t), v _y1 (t) and v _y4 (t), respectively, and a second spatial differentiator 109 for _obtaining the spatial gradient of the sound pressure; The gradient signal output from the second spatial differentiator 109 is time-integrated, and the obtained time integration signals u _xx (t), u _yy (t), u _xy (t), and u _yx (t) are _subjected to the matrix operation. The configuration includes a second integrator 110 that outputs to the circuit 103.
In FIG. 23, the matrix operation circuit 103 and the separation matrix element calculation circuit 104 are shown as one block.

Further, the average calculator 108 determines the gradient of f ₂ and f ₃ ((f ₃ −f ₂ ) in order to improve the objectivity when obtaining the gradient of the sound pressure in the x direction at f ₅ (origin O). / Δx) and the slope of f ₇ and f ₈ ((f ₈ −f ₇ ) / Δx). Similarly, in obtaining the gradient of the sound pressure in the y direction at f ₅ (origin O), in order to improve the objectivity, the gradient ((f ₃ −f ₈ ) / Δy) between f ₃ and f ₈ and f The average of the slopes of ₂ and f ₇ ((f ₂ −f ₇ ) / Δy) is taken.

The second spatial differentiator 109, the gradient in the x direction of the sound pressure _{((f 5 -f 4) /} Δx- (f 6 -f 5) / Δx) of the x direction of the gradient (((f ₅ - f ₄ ) / Δx− (f ₆ −f ₅ ) / Δx) / Δx), y of the sound pressure gradient in the y direction ((f ₁ −f ₅ ) / Δy− (f ₅ −f ₉ ) / Δy) Direction gradient (((f ₁ −f ₅ ) / Δy− (f ₅ −f ₉ ) / Δy) / Δy), sound pressure gradient in the x direction ((f ₃ −f ₂ ) / Δx− (f ₈ −f ₇ ) / Δx) in the y direction (((f ₃ −f ₂ ) / Δx− (f ₈ −f ₇ ) / Δx) / Δy), the y direction sound pressure gradient ((f ₃ − f ₈ ) / Δy− (f ₂ −f ₇ ) / Δy) in the x-direction gradient (((f ₃ −f ₈ ) / Δy− (f ₂ −f ₇ ) / Δy) / Δx) Yes.

Further, v _x (t) represents the particle velocity in the x direction, and v _y (t) represents the particle velocity in the y direction. By outputting to the matrix operation circuit 103, the signal separation device is placed in the two-dimensional plane. It is intended to be able to handle up to two sound sources. U _xx (t) is the gradient in the x direction of the particle displacement in the x direction, u _xy (t) is the gradient in the y direction of the particle displacement in the x direction, and u _yy (t) is the y direction of the particle displacement in the y direction. , U _yx (t) indicates the gradient in the x direction of the particle displacement in the y direction, which is output to the matrix operation circuit 103 and used together with v _x (t) and v _y (t), The signal separation device can cope with up to four sound sources in a two-dimensional plane.

As shown in FIG. 26, the omnidirectional microphone 101 includes nine omnidirectional microphones (f ₁ , f ₂ , f ₃ , f ₄ , f ₅ , f ₆ , f ₇ , f ₈ , f ₉ ). Thus, four omnidirectional microphones close to each other constitute one omnidirectional microphone unit (a set of f ₁ , f ₂ , f ₃ and f ₅ , a set of f ₃ , f ₅ , f ₆ and f _8. , F ₅ , f ₇ , f ₈ and f _9, and f ₂ , f ₄ , f ₅ and f ₇ ).

In the fourth embodiment, as shown in FIG. 26, the omnidirectional microphone f ₅ is arranged at the origin O (observation point 100), and is at a distance of Δx with respect to the x-axis direction and in the y-axis direction. On the other hand, another omnidirectional microphone is arranged with a distance of Δy.

As shown in FIG. 25, an omnidirectional microphone unit 101a composed of four omnidirectional microphones (f ₁ , f ₂ , f ₃ , f ₅ ) will be described as an example. The source signal from the collected sound source is directed to the directivity 111a in the x-axis direction and the directivity in the y-axis direction by the first spatial differentiator 102a and the first integrator 107a connected to the omnidirectional microphone unit 101a. Is output as a time integration signal having the characteristics 111b. The first spatial differentiator 102a is connected to omnidirectional microphones f ₂ and f ₃ and x
An x-direction spatial differentiator for obtaining an axial spatial gradient (f ₃ −f ₂ ) / Δx and omnidirectional microphones f ₁ and f ₅ are connected, and a spatial gradient in the y-axis direction (f ₁ −f ₅ ) / It is a structure provided with the y-axis direction space differentiator which calculates | requires (DELTA) y.

Note that one omnidirectional microphone unit shown in FIG. 25 is replaced with two omnidirectional microphone units 101b and 101d in the x direction and two omnidirectional microphone units 101a and 101c in the y direction, as shown in FIG. Are placed close to each other in parallel to form the omnidirectional microphone 101 in FIG. Further, omnidirectional microphones (f ₂ , f ₃ , f ₅ , f ₇ , f ₈ ) that are in contact with each other between adjacent omnidirectional microphone units are shared between adjacent omnidirectional microphone units.

Next, the separation processing operation of the signal separation device according to the present embodiment based on the above configuration will be described.
First, source signals P ₁ (t), P ₂ (t), P ₃ (t), and P ₄ (t) from sound sources 201, 202, 203, and 204 arranged in a homogeneous two-dimensional plane are observed. Sound is collected by an omnidirectional microphone (f ₁ , f ₂ , f ₃ , f ₄ , f ₅ , f ₆ , f ₇ , f ₈ , f ₉ ) near the point 100 (step 27).

The collected source signals P ₁ (t), P ₂ (t), P ₃ (t), and P ₄ (t) are spatially differentiated in the x-axis and y-axis directions by the first spatial differentiator 102. Spatial gradient fx ₁ (t), fx ₂ (t), fx ₃ (t), fx ₄ (t), fy ₁ (t), fy ₂ (t), fy ₃ (t), fy ₄ (t) Is calculated (step 28).

This spatial gradient fx ₁ (t), fx ₂ (t), fx ₃ (t), fx ₄ (t), fy ₁ (t), fy ₂ (t), fy ₃ (t), fy ₄ (t) Is time-integrated by the first integrator 107 (step 29).

This time-integrated spatial gradient fx ₁ (t), fx ₂ (t), fx ₃ (t), fx ₄ (t), fy ₁ (t), fy ₂ (t), fy ₃ (t), fy ₄ (t) time integration signals v _x1 (t), v _x2 (t), v _x3 (t), v _x4 (t), v _y1 (t), v _y2 (t), v _y3 (t), Of v _y4 (t), the average of v _x1 (t) and v _x4 (t) and the average of v _y1 (t) and v _y4 (t) are calculated by the average calculator 108 (step 30). .

On the other hand, the spatial gradients fx ₁ (t), fx ₂ (t), fx ₃ (t), fx ₄ (t), fy ₁ (t), fy ₂ (t), fy ₃ time-integrated in step 29 above. (t), fy ₄ (t) time integration signals v _x1 (t), v _x2 (t), v _x3 (t), v _x4 (t), v _y1 (t), v _y2 (t), v _{Of y3} (t) and v _y4 (t), v _x2 (t) and v _x3 (t), v _x1 (t) and v _x4 (t), v _y2 (t) and v _y3 (t), v The second spatial differentiator 109 calculates _y1 (t) and _vy4 (t), and the gradient of the spatial gradient based on the spatial differentiation in the x-axis and y-axis directions (step 31).
The gradient signal output from the second spatial differentiator 109 is time-integrated (step 32).

The average time integration signals v _x (t), v _y (t) obtained in step 30 and the time integration signals u _xx (t), u _yy (t), u _xy (t), obtained in step 32, u _yx (t) is multiplied by the inverse matrix of the mixing matrix by the matrix operation circuit 103. Thereafter, the separation signals P ₁ (t), P ₂ (t), P ₃ (t), and P ₄ (t) are output and the source at the observation point 100 in the same manner as in the second embodiment. Standard directions of arrival directions θ ₁ , θ ₂ , θ ₃ , θ _{4 of} signals P ₁ (t), P ₂ (t), P ₃ (t), P ₄ (t) and sound sources 201, 202, 203, 204 σ ₁ , σ ₂ , σ ₃ , σ ₄ are output to the display device or the like.

In this way, the signal separation device according to the fourth embodiment is based on the spatiotemporal gradient analysis, and in a homogeneous space, the spatial gradient at any observation point 100 is the temporal gradient at the sound sources 201, 202, 203, and 204. The fact that it is expressed by a linear mixture of That is, the features of this embodiment are summarized as follows. By measuring the spatial gradient at only one observation point 100 and its vicinity, the instantaneous mixed sum of the time gradients of the source signals P ₁ (t), P ₂ (t), P ₃ (t), and P ₄ (t) Can be obtained. Therefore, the problem can be reduced to the simplest instantaneous mixed blind signal separation problem. Furthermore, since the observed quantity is not a scalar quantity (sound pressure) but a vector quantity (sound pressure spatial gradient), source signals P ₁ (t), P ₂ (t), P ₃ (t) including the direction of arrival of the sound source are included. , P ₄ (t) can be separated. By using this spatiotemporal gradient analysis method, it is not necessary to consider the difference in arrival time of signals generated between the observation points 100, and the spatial gradient of the observation signal can be regarded as an instantaneous linear mixed signal. The separation signals P ₁ (t), P ₂ (t), P ₃ (t), and P ₄ (t) can be estimated by processing for linear mixing.

In particular, in the first embodiment, the second embodiment, and the third embodiment, the signal separation device can only support a maximum of two sound sources in a two-dimensional plane. In the fourth embodiment, the time integration signal of the sound pressure gradient signal from the first integrator 107 is converted by the average calculator 108 using the omnidirectional microphone 101 including nine omnidirectional microphones. The particle velocity v _x (t) in the direction and the particle velocity v _y (t) in the y direction are output to the matrix operation circuit 103, and the particles in the x direction are output by the second spatial differentiator 109 and the second integrator 110. X-direction gradient of displacement u _xx (t), x-direction particle displacement y-direction gradient u _xy (t), y-direction particle displacement y-direction gradient u _yy (t) and y-direction particle displacement By outputting to the matrix operation circuit 103 as the gradient u _yx (t) in the x direction, The signal separation device can support up to four sound sources in a two-dimensional plane.

(Operational logic of each embodiment of the present invention)
The arithmetic logic of the fourth embodiment will be described below.

The arithmetic logic of the fourth embodiment is substantially the same as the arithmetic logic up to the equation (8) described in the first embodiment, the second embodiment, and the third embodiment. Therefore, description of the arithmetic logic up to the equation (8) is omitted.
The spatial gradient of f (x, y, t) is expressed as an instantaneous linear mixed signal when P _t1 (t), P _t2 (t), P _t3 (t), and P _t4 (t) are used as source signals. .

両辺を時間で積分すると、次式が得られる。

Integrating both sides over time gives the following equation:

ここで、∇ｆ（ｘ，ｙ，ｔ）は、粒子速度ｖの∂ｖ／∂ｔに比例しており、式（３８）に代入すると、次式が得られる。

Here, ∇f (x, y, t) is proportional to ∂v / ∂t of the particle velocity v, and the following equation is obtained by substituting into equation (38).

また、粒子速度ｖを式（４０）とすると、式（４１）が得られる。

Further, when the particle velocity v is expressed by equation (40), equation (41) is obtained.

したがって、行列Ｂ（６行４列）の擬似逆行列Ｂ^-1を既存の方法で求めることで、源信号Ｐ₁(t)、Ｐ₂(t)、Ｐ₃(t)、Ｐ₄(t)を求めることができる。

Therefore, the source signals P ₁ (t), P ₂ (t), P ₃ (t), P ₄ (t) are obtained by obtaining the pseudo inverse matrix B ⁻¹ of the matrix B (6 rows × 4 columns) by an existing method. ).

In this case, the 1st and 2nd row portions of the matrix B are a mixed matrix of A relating to θ ₁ , θ ₂ , θ ₃ , and θ ₄ , and the 3rd to 6th row portions of B are σ ₁ , σ ₂ , It is a new matrix of σ ₃ and σ ₄ .

Claims (7)

In a signal separation device for separating superimposed signals on which output signals output from output sources at different positions in a two-dimensional plane or a three-dimensional space are superimposed,
Spatial gradient detection means for detecting the superimposed signal as an instantaneous mixed sum of time gradients for each output signal;
Signal separation means comprising signal separation means for directly capturing the signals detected by the spatial gradient detection means without storing them and separating the output signals based on the instantaneous mixed sum of the time gradients of the output signals. apparatus. The signal separation device according to claim 1,
The signal separation device includes an integrator that time-integrates a spatial gradient signal output from the gradient detection unit. The signal separation device according to claim 2, wherein
The signal separating means includes
A first integrator for time-integrating a spatial gradient signal output from the gradient detection means;
Among the time integration signals output from the first integrator, an average calculator that calculates the average of two signals;
A spatial differentiator for calculating a gradient of a spatial gradient by spatial differentiation of the time integration signal output from the first integrator;
And a second integrator for time-integrating the spatial gradient signal output from the spatial differentiator. In the signal separation device according to any one of claims 1 to 3,
The signal separation device, wherein the spatial gradient detecting unit is configured as a microphone array that detects each output signal as each audio signal, and the microphone array detects the sound pressure gradient of the audio signal. The signal separation device according to claim 4, wherein
A signal separation device, wherein a sound pressure gradient detected by the microphone array is obtained by time differentiation of particle velocity. In the signal separation device according to claim 4 or 5,
The signal separation device, wherein the microphone array is configured by a bidirectional microphone that measures a sound pressure difference at least in the x-axis direction or the y-axis direction at an observation point with respect to a different position from which each audio signal is output. . In the signal separation device according to any one of claims 1 to 6,
The signal separation means is characterized in that the signal separation means separates the superimposed signal into output signals and outputs the direction and / or standard deviation of each output signal at each output source.

Images