JP2008304555A - Sound input apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To input only sound generated from a desired sound source. <P>SOLUTION: In sound pressure O(f, k) which is output from a dead point forming means 2, sound other than the sound generated from the sound source which exists in the dead point, namely, the sound pressure of noise is included, while in sound pressure M(f, k) detected by a sound collecting sensor means 1, both the sound pressure of the sound generated from the sound source which exists in the dead point, and the sound pressure of noise are included. Since only sound pressure S(f, k) of the sound generated from the sound source which exists in the dead point, is extracted by an extracting means 3 of object speaker voice, only the sound generated from the sound source which exists in the dead point is input, even when the noise comes from a direction of the dead point. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an acoustic input device that inputs sound emitted from a sound source.

2. Description of the Related Art Conventionally, various acoustic input devices for inputting only sound emitted from a specific sound source in an environment where ambient noise and reverberation exist, for example, a voice uttered by a person (speaker voice) have been provided. For example, in the conventional example described in Patent Document 1, frequency analysis is performed for two microphones and acoustic signals collected by each microphone to obtain frequency components for two channels, and for each channel frequency component By performing adaptive beamformer processing, the sensitivity other than the direction of the target sound (speaker voice) is reduced to obtain an acoustic signal in which the ambient noise is suppressed, and similarly the sensitivity other than the direction of the ambient noise is reduced. The target sound is suppressed and the target sound direction and the ambient noise direction are estimated from the filter coefficients used in the adaptive beamformer processing and corrected sequentially, and then the former sound signal is processed by spectral subtraction processing. In the latter case, only the target sound (speaker voice) is input by removing the ambient noise component from the acoustic signal.
JP 2000-47699 A

  However, the above conventional example has a problem that the target sound and the noise other than the target sound cannot be separated when the speaker and the noise source are present in the same direction as viewed from the microphone, for example.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide an acoustic input device capable of inputting only sound emitted from a desired sound source.

  In order to achieve the above object, the invention of claim 1 is a collection for detecting a sound pressure, a time differential value of the sound pressure, and a spatial differential value obtained by differentiating the sound pressure in each axial direction of a two-dimensional orthogonal coordinate system. The sound collection sensitivity is minimized by performing the load sum of the predetermined coefficient vector and the low-pass filter processing on the sound sensor means and the sound pressure, time differential value, and spatial differential value detected by the sound collection sensor means. From the target speaker using the dead point forming means for forming the dead point at the position of the target speaker set in advance, and the sound pressure detected by the sound collecting sensor means and the sound pressure output from the dead point forming means. And a target speaker voice extracting means for extracting only the sound pressure of the uttered voice.

  According to the first aspect of the present invention, the sound pressure output from the dead point forming means includes only sound other than the sound emitted from the sound source existing at the dead point, that is, noise pressure, The sound pressure detected by the sensor means includes both the sound pressure of the sound emitted from the sound source existing at the dead point and the sound pressure of the noise. From the sound source existing at the dead point by the target speaker voice extraction means Since only the sound pressure of the emitted sound is extracted, only the sound emitted from the sound source existing at the dead point can be input even when the noise comes from the direction of the dead point.

  The invention of claim 2 is characterized in that, in the invention of claim 1, the target speaker voice extracting means extracts the sound pressure of the sound emitted from the sound source by the spectral subtraction method.

  The invention of claim 3 is characterized in that, in the invention of claim 1, the target speaker voice extraction means extracts the sound pressure of the sound emitted from the sound source by independent component analysis.

  According to a fourth aspect of the present invention, in the third aspect of the present invention, the target speaker voice extracting means performs the principal component analysis before performing the independent component analysis.

  According to a fifth aspect of the present invention, in any one of the first to fourth aspects, the sound collection sensor means includes a plurality of microphones arranged in directions orthogonal to the respective axes of the two-dimensional orthogonal coordinate system. It is characterized by that.

  According to a sixth aspect of the present invention, in any one of the first to fourth aspects, the sound collecting sensor means includes a microphone having a biaxial orthogonal gimbal structure in which a diaphragm is supported at one central point. It is characterized by doing.

  A seventh aspect of the present invention provides the sound source position changing means for changing the sound source position according to any one of the first to sixth aspects, wherein the sound source position changing means is a sound pressure detected by the sound collecting sensor means. The sound pressure, time differential value, and spatial differential value detected by the sound collecting sensor means when the value obtained by dividing the instantaneous power of the sound by the instantaneous power of the sound pressure output from the dead point forming means is equal to or greater than a predetermined threshold value The sound source position is changed to a position estimated based on the above.

  According to the invention of claim 7, even when the position of the sound source changes, the sound source position is changed by the sound source position changing means, so that the position of the sound source does not deviate from the dead point, and as a result, the sound source moves. Also, only sound emitted from the sound source can be input.

  According to the present invention, even when noise comes from the direction of the dead center, it is possible to input only the sound emitted from the sound source existing at the dead point.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In addition, about the literature referred in description of embodiment, it describes with reference literature 1, reference literature 2, ..., and lists each reference literature 1, 2, ... at the end.

(Embodiment 1)
As shown in FIG. 1, the acoustic input device according to the present embodiment collects sound pressure, a time differential value of the sound pressure, and a spatial differential value obtained by differentiating the sound pressure in each axial direction of a two-dimensional orthogonal coordinate system. The sound collection sensitivity is improved by performing a load sum with a predetermined coefficient vector and a low-pass filter process on the sound sensor means 1 and the sound pressure, time differential value, and space differential value detected by the sound collection sensor means 1. Using the dead point forming means 2 for forming the minimum dead point at a predetermined target speaker position, the sound pressure detected by the sound collecting sensor means 1 and the sound pressure output from the dead point forming means 2. And target speaker voice extraction means 3 for extracting only the sound pressure of the sound emitted from the sound source (speaker) existing at the dead point.

As shown in FIG. 1, a plurality of (four in the illustrated example) sound collecting sensor means 1 are arranged in a direction orthogonal to the x-axis and y-axis (positive direction of the z-axis) of the three-dimensional orthogonal coordinate system. Spatiotemporal gradient measurement for directional microphones 10A, 10B, 10C and 10D and output signals f _A (t), f _B (t), f _C (t) and f _D (t) of microphones 10A to 10D And a spatiotemporal gradient measurement processing unit 11 for performing processing. In the spatiotemporal gradient measurement processing unit 11, the in-phase component M (t) of the sound pressure is determined from the output signals f _A (t), f _B (t), f _C (t), and f _D (t) of the microphones 10A to 10D. , Time differential value (time gradient component) M _t (t), x-axis direction spatial differential value (x-axis direction spatial gradient component) M _x (t), y-axis direction spatial differential value (y-axis direction spatial gradient component) M Each _y (t) is obtained from the following equation.

M (t) = f _A (t) + f _B (t) + f _C (t) + f _D (t)
M _t (t) = df _A (t) / dt + df _B (t) / dt + df _C (t) / dt + df _D (t) / dt
M _x (t) = f _A (t) + f _B (t) −f _C (t) −f _D (t)
_{M y (t) = f A} (t) -f B (t) + f C (t) -f D (t)
The dead point forming means 2 includes an in-phase component M (t), a time differential value M _t (t), an x-axis direction spatial differential value M _x (t), and a y-axis direction spatial differential value output from the sound collection sensor means 1. with M _y (t), and forms a dead point by the application of space-time gradient method. Here, in describing the dead point forming process by the dead point forming means 2, the spatiotemporal gradient method will be described in detail first.

  The spatiotemporal gradient method was originally proposed as one of the methods for determining an optical flow that is an apparent velocity field in a moving image (see Reference 1). Assumption (f (x, y, t) = f (x + δx, y +) that the image function f (x, y, t) representing the characteristics of the light and shade pattern in the moving image is kept unchanged during the movement. δy, t + δt)), the analysis method is based on an equation that relates the velocity of the optical flow at a certain point (x, y) to the spatial gradient and temporal gradient of the gray-scale distribution of the moving image. The method will be explained in detail below.

At time t + δt, when the density pattern f (x + δx, y + δy, t + δt) at coordinates (x + δx, y + δy) is tailored around (x, y, t),
f (x + δx, y + δy, t + δt) = f (x, y, t) + f x δx + f y δy + f t δt + O (δx + δy + δt) ... (1)
It becomes. Here, O (δx + δy + δt) is a second-order or higher term of δx, δy, δt, but is ignored in the following because it is a minute amount. At this time, when the grayscale pattern at the coordinate (x, y) at time t moves with the density value distribution kept constant at the coordinate (x + δx, y + δy) after δt time has passed, The following equation holds from the attachment.

f (x, y, t) = f (x + δx, y + δy, t + δt)
= f (x, y, t ) + f x δx + f y δy + f t δt ... (2)
_{f x δx + f y δy +} f t δt = 0 ... (3)
Dividing both sides of equation (3) by δt,
_{f x δx / δt + f y} δy / δt + f t = 0 ... (4)
Get. Here, assuming that δt is infinitesimal, assuming that δt → 0, the following equation is obtained.

_{f x dx / dt + f y} dy / dt + f t = 0 ... (5)
Using optical flow velocity v = (u, v) = (dx / dt, dy / dt), equation (5) becomes
_{uf x + vf y + f t} = 0 ... (6)
Equation (6) is an equation that correlates the gradient of time and space of the gray value of the moving image with the optical flow velocity v.

  Next, an assumption is made that “the velocity field can be approximated to be almost constant in the vicinity region Γ of a certain point of interest”. At this time, equation (6) must be established everywhere in the region Γ. Therefore, evaluation is performed using the square integral of the left side of the equation (6) (the following equation (7)), and the velocity field is obtained by the method of least squares.

Differentiating equation (7) with respect to u and v and setting it to 0,
uS _xx + vS _xy + S _xt = 0, uS _xy + vS _yy + S _yt = 0 (8)

Is obtained. Solving equation (8) gives the velocity vector (u, v)
u = (S _yt S _xy -S _xt S _yy ) / (S _xx S _yy -S ² _xy ), v = (S _xt S _xy -S _yt S _xx ) / (S _xx S _yy -S ² _xy )… (10)
It is required as follows.

  Next, by applying the spatiotemporal gradient method for obtaining the optical flow velocity in the moving image described above, based on the linear relationship established between the sound pressure at one point in the sound field created by the sound source in the space and the spatiotemporal gradient. A method for localizing the sound source position will be described (see Reference 2).

As shown in FIG. 2, it is assumed that a three-dimensional orthogonal coordinate system having an observation point as the origin is taken, and there are a plurality of point sources that are uncorrelated with each other in front (z> 0). The speed of sound is c, the coordinates of the i-th sound source are (x _i , y _i , z _i ), the distance between the sound source and the observation point is R _i = (x ² _i + y ² _i + z ² _i ) ^1/2 , Assuming that the sound source sound is g ⁱ (t) and the sound field formed by each sound source at the observation point is f ⁱ (t), the synthesized sound field f formed at the observation point is the sum of the spherical waves from them,

It is expressed. By performing partial differentiation, the x, y differentiation and time differentiation of the sound field at the observation point are expressed by the following equations (12), (13), (14).

here,
ξ ⁱ _x = x _i / R ² _i , ξ ⁱ _y = y _i / R ² _i … (15)
Is called the intensity gradient,
τ ⁱ _x = x _i / cR _i , τ ⁱ _y = y _i / cR _i (16)
Is called the time gradient in the x and y directions.

Next, for the sake of simplicity, a sound source localization method in the case of one sound source will be described. In the case of one sound source, equations (12) and (13) are
f _x = -ξ _x f-τ _x f _t , f _y = -ξ _y f-τ _y f _t (17)
Thus, τ _x , τ _y , ξ _x , ξ _y are obtained by applying the method of least squares as in the equation (1). The evaluation function in the short time window Γ
J = ∫ _Γ {(f _x + ξ _x f + τ _x f _t ) ² + (f _y + ξ _y f + τ _y f _t ) ² } dt… (18)
And When the equation (18) is partially differentiated with respect to τ _x , τ _y , ξ _x , ξ _y and set to 0, the following equation is obtained.

∂J / ∂τ _x = ∫ _Γ 2 (f _x + ξ _x f + τ _x f _t ) ・ f _t dt = 0, ∂J / ∂τ _y = ∫ _Γ 2 (f _y + ξ _y f + τ _y f _t ) ・ f _t dt = 0… (19)
∂J / ∂ξ _x = ∫ _Γ 2 (f _x + ξ _x f + τ _x f _t ) ・ f _t dt = 0, ∂J / ∂ξ _y = ∫ _Γ 2 (f _y + ξ _y f + τ _y f _t ) ・ f _t dt = 0… (20)
Where the covariance matrix estimated from the observation window Γ is

Then, equations (19) and (20) are
S _xt + ξ _x S _t + τ _x S _tt = 0, S _yt + ξ _y S _t + τ _y S _tt = 0… (22)
S _x + ξ _x S + τ _x S _t = 0, S _y + ξ _y S + τ _y S _t = 0… (23)
Rewritten. By solving the equations (22) and (23), τ _x , τ _y , ξ _x , and ξ _y are obtained as follows.

τ _x = (S _x S _t -SS _xt ) / (SS _tt -S ² _t ), τ _y = (S _y S _t -SS _yt ) / (SS _tt -S ² _t ) (24)
_{ξ x = (S xt S t} -S x S tt) / (SS tt -S 2 t), ξ y = (S yt S t -S y S tt) / (SS tt -S 2 t) ... (25 )
The azimuth angle (x / R, y / R) = (cτ _x , cτ _y ) of the sound source can be obtained from equations (21) and (24). The distance R to the sound source can be obtained by applying the least square method from the equations (15) and (16). Evaluation function

And this is a partial differential with 1 / R and set to 0

It becomes. Solving this
R = c (τ ² _x + τ ² _y ) / (τ _x ξ _x + τ _y ξ _y )… (28)
The distance to the sound source is required.

  Next, a method for directivity control using the spatiotemporal gradient of the sound field will be described (see References 3 to 5). Assuming the case of one sound source, the spatial gradient in the x and y directions of the sound pressure signal f (t) at the observation point is given by equations (12) and (13).

It becomes. When this equation is rewritten using the vector r = (x, y, z) from the sound source to the observation point,

It becomes. Then _{f (t), f t (} t), when ∇f (t) is observed, these weighted sum is

It is expressed. Here, u and u _t are real constants, and w = (w _x , w _y , 0) is a unit vector with the observation point as the origin and pointing in an arbitrary direction. Substituting equation (30) into equation (31),

It becomes. Therefore, the load sum of the spatiotemporal gradient is expressed as the sum of filters having different directivity characteristics H (r) and H _t (r) for f (t) and f _t (t), respectively. When H (r) = α, equation (33) becomes

And can be transformed. Here, if the angle formed by the two vectors a and b is θ, the following formula holds.

Using the formula of equation (38), equation (36) can be rewritten as

Here, from | w | = 1,

It is expressed by the sphere equation. When u + α = 0, equation (35) becomes r · w = 0 (42)
It becomes. Also, when H _t (r) = α, equation (34) becomes

Therefore, if the angle between vectors r and w is θ (r), | w | = 1

It becomes. Therefore, equation (43) becomes

It becomes.

From the equations (41), (42), and (45), the following properties are obtained for H (r) and H _t (r).
1) The two directivity characteristics H (r) and H _t (r) have a rotationally symmetric body with w as the axis. 2) When H (r) = 0, the distribution of r has a diameter 1 / u (u ≠ 0). ) To form a spherical surface or plane (u = 0) 3) When H _t (r) = 0, the distribution of r is a conical surface or plane (u _t = 0) with apex angle 2cu _t (u _t ≠ 0). 4) The intersection of the distributions of r when H (r) = 0 and H _t (r) = 0 is a circle or plane.

Get. Therefore, the frequency response T (r, w) from the sound source r to s (t) is
T (r, w) = H (r) + jωH _t (r) (47)
If H (r) and H _t (r) are real numbers, if T (r, w) = 0, H (r) = 0, H _t (r) = 0 (48)
It becomes. Therefore, it can be seen from equation (47) that the zero distribution where S (ω) = 0 does not depend on the frequency ω, but only on the sound source position r. Therefore, the time slope and x of the sound pressure at the observation point, when the spatial gradient in the y-direction is obtained, to form the zero sensitivity region (dead center), at a certain moment f, f _t, f _x, the f _y It is only necessary to take the load sum and perform compensation filter processing (low-pass filter processing).

Thus, in the dead point forming means 2 in this embodiment, the in-phase component M (t), the time differential value M _t (t), and the x-axis direction spatial differential value M _x ( t), y-axis spatial differential value M _y (t) respectively above f, f _t, f _x, replaced by f _y, the vector M = (M (t of these values as elements) M _t (t ) defines the _{M x (t) M y (} t)) T, after calculating the weighted sum of the coefficient vector and a load element _{W = (W W t W x} W y) T to these, the low-pass By performing the filtering process, a dead point is formed at a predetermined arbitrary position. Specifically, the directivity characteristics H (r) and H _t (r) described above are replaced with H ₁ (r _i ) and H ₂ (r _i ), respectively, and defined as follows (where r _i is a sound source) i's position vector, n _ix and n _iy are the x and y components of r _i / | r _i |, respectively).

Further, the filter coefficients p = W ^H H ₁ (r _i ) and q = W ^H H ₂ (r _i ) of the low-pass filter 20 represented by (p + jωq) ⁻¹ in FIG. By selecting such a coefficient vector W, a dead point independent of the frequency can be formed at the position r _i of the sound source i. The output O (t) of the dead point forming means 2 does not include the sound pressure of the sound emitted from the sound source existing at the dead point. In other words, ambient noise and reverberation excluding the voice of the speaker at the dead point. Only sound (hereinafter referred to as noise) is included.

  The target speaker voice extraction means 3 has a sound pressure output from the sound collection sensor means 1 (target sound <speaker voice at the dead point> and sound pressure including noise) M (t), and dead point formation. The target sound S (t) is extracted from the noise component O (t) output from the means 2 by a conventionally known spectral subtraction method (see Reference 6). First, the target speaker voice extraction means 3 divides the in-phase component M (t) and the noise component O (t) for each unit time (frame time) by the frame dividing unit 30, and the divided sound pressure M (t, k) and the noise component O (t, k) are converted from the time domain to the frequency domain by the fast Fourier transform unit 31 (where k indicates a frame number). Then, the average amplitude μ (= E {| O (f, k |}) of the noise component O (f, k) is calculated by the noise average amplitude calculator 32 and the sound pressure M (f calculated by the amplitude calculator 33 is calculated. , k) amplitude value | M (f, k) | is subtracted from the average amplitude μ of the noise component O (f, k) and the subtracted value (| M (f, k) | −μ) By multiplying the phase of the sound pressure M (f, k) calculated by the calculation unit 34 (= exp {j∠M (f, k)}), the output S (f, k) = (not including noise) | M (f, k) | −μ) · exp {j∠M (f, k)} is taken out, and this output S (f, k) is inversely transformed by fast Fourier transform to return from the frequency domain to the time domain. Only the target sound S (t) that does not contain noise can be obtained.

  Thus, according to the acoustic input device of the present embodiment, the sound pressure (in-phase component) O (f, k) output from the dead point forming means 2 is a sound other than the sound emitted from the sound source existing at the dead point. That is, only the sound pressure of noise is included, while the sound pressure (in-phase component) M (f, k) detected by the sound collection sensor means 1 is the sound pressure of the sound emitted from the sound source existing at the dead point. And the sound pressure of the noise are included, and only the sound pressure S (f, k) of the sound emitted from the sound source existing at the dead point is extracted by the target speaker voice extraction means 3, so that the noise is the dead point. Even when coming from the direction of, only the sound emitted from the sound source existing at the dead point can be input. FIG. 3 shows a sound pressure M (f) including the target sound and noise, a sound pressure O (f) including only the noise not including the target sound, and a sound pressure S (f) of the target sound extracted by the target speaker voice extraction means 3. ) Shows an example of frequency characteristics, and it can be seen that the noise component included in the sound pressure S (f) is sufficiently suppressed.

  Here, if the acoustic input device of this embodiment is installed in an interphone device (door phone slave), it is possible to make a call by extracting only the voice of a speaker (visitor) even in an environment with a high ambient noise.

Incidentally, in the sound collection sensor means 1 in the present embodiment, the four microphones 10A to 10D are arranged on the xy plane, but as shown in FIG. May be used instead of the microphones 10A to 10D. The diaphragm 13 has a thin disk shape as a whole, and double grooves 14 and 15 are formed on concentric circles in the central portion thereof, and a pair of beams 14a and 14a partitioning the inner groove 14 and an outer groove 15 are formed. It has a pair of beams 15a and 15a for partitioning, and can rotate around the x-axis and the y-axis by twisting each beam 14a and 15a with a point (center) supported by the support rod 16 as a fulcrum (reference) References 7 and 8). Therefore, if four observation points A, B, C, and D are set on the diaphragm 13, and the displacement at each observation point is replaced with the output of the microphones 10A to 10D, the in-phase component M (t) and the time differential value M _{t (t),} x-axis spatial differential value M _x (t), it is possible to detect the y-axis direction spatial differential value M _y (t).

(Embodiment 2)
The present embodiment is characterized in that an independent component analysis technique is used instead of the spectrum subtraction method as the extraction process in the target speaker voice extraction means 3, and other configurations and operations are the same as those in the first embodiment. Therefore, the same components are denoted by the same reference numerals, and illustration and description thereof are omitted.

The purpose of Independent Component Analysis (ICA) is to represent multiple observed variables as a linear combination of statistically independent variables, and the independent variables calculated from the observed variables are independent components. It is. For example, assuming that the observed variable vector is X and this observed variable vector X is given by a linear combination of the unknown independent variable vector S, the unknown mixing matrix is A and the relationship X = AS is established. The independent component analysis is such that each component of the reconstructed data vector Y obtained by Y = WX using only the observation data and the separation matrix W without using any knowledge about the independent component and the mixing matrix becomes independent. This is a technique for obtaining the separation matrix W. Ideally, the separation matrix W may be an inverse matrix of the mixing matrix A (W = A ⁻¹ ). Here, FIG. 5 shows a model of independent component analysis when the observation data is two-dimensional.

Here, an observation matrix having a sound pressure M (f, k) including a target sound and noise and a sound pressure O (f, k) of noise output from the dead point forming means 2 as components is X (f, k). ) = [M (f, k) O (f, k)] ^T, and a matrix having the sound pressure S (f, k) of the target sound and the sound pressure N (f, k) of the noise source as components. f, k) = [S (f, k) N (f, k)] ^T and the spatial transfer matrix is A,

It is expressed. The target speaker voice extraction means 3 adaptively identifies the target matrix by identifying the separation matrix W satisfying S (f, k) = W _{i + 1} (f, k) X (f, k). And only the target sound S (f, k) is extracted from the sound pressure O (f, k) including noise. This separation matrix W _{i + 1} (f) is

(See Reference 7). Here, diag indicates a diagonal matrix, and Φ (Y) is a sigmoid function represented by φ (y) = (1 + exp (−y)) ⁻¹ , or φ (y) = 3 / 4y ¹¹ + It is a nonlinear vector function approximated by a polynomial such as 25 / 4y ⁹ + 14 / 4y ⁷ + 47 / 4y ⁵ + 29 / 4y ³ . For example, when the above equation is approximated by a sigmoid function,

It becomes.

  By the way, if the number of observation variables is larger than the number of independent variables, the observation variables are linearly dependent, and the separation matrix is a matrix for reducing the order. Also, if the variables are independent of each other, they are uncorrelated, so the separation matrix is also a matrix that decorrelates the variables. Principal component analysis (PCA) is a statistical method that simultaneously performs decorrelation and associated reduction in dimensions. Principal component analysis is sometimes used as preprocessing for independent component analysis.

  Therefore, also in this embodiment, when there is reflection or reverberation as noise, principal component analysis may be performed as preprocessing of independent component analysis in order to reduce the dimension of the observation matrix X (f, k). If the variables are independent of each other, they are also uncorrelated, so that decorrelation and reduction in dimensions can be performed simultaneously by principal component analysis.

  For example, when the number of sound sources is r, the singular value decomposition of the m-dimensional observation matrix X (f, k) is as follows.

When each sound source spectrum S ₁ , S _{2 ,.}

here,

Considering this transformation, the variance is

It becomes. Since the variance matrix is a diagonal matrix, the converted variables are uncorrelated with each other. In addition, the m-dimensional observation matrix X can be compressed to r dimensions, and the amount of processing in the subsequent independent component analysis can be reduced.

(Embodiment 3)
In the first and second embodiments, assuming that the position of the sound source (for example, a speaker) of the target sound is known, only the target sound emitted from the sound source is collected by forming a dead point at the position. . However, there are many cases where the position of the sound source (visitor) of the target sound is not uniquely determined as in a door phone sub-unit. On the other hand, in an environment where there is very little noise (ambient noise and reverberation sound), the position of the sound source can be estimated using the sound source localization technique described above using the spatiotemporal gradient method, and the position of the sound source is uniquely Even if it is not fixed, it is possible to input only the target sound emitted from the sound source by estimating the position of the sound source and forming a dead point at the position.

Therefore, in the present embodiment, the instantaneous power P _M (t) of the in-phase component M (t) detected by the sound collecting sensor means 1 is used as the instantaneous power P of the in-phase component O (t) output from the dead point forming means 2. Sound pressure (in-phase component M) detected by the sound collecting sensor means 1 when the value divided by _O (t) (= P _M (t) / P _O (t)) is equal to or greater than a predetermined threshold value δ. , and a sound source position changing means for changing the sound source position to a position where the time differential value M _t, spatial differential values M _x, is estimated based on M _y. Since this sound source position changing means performs almost the same processing as the dead point forming means 2, the dead point forming means 2 can also be used. Then, when the above condition is satisfied, the dead point forming means 2 estimates the sound source position. If the estimated sound source position is different from the position of the dead point at that time, the dead point is set at the estimated sound source position. Change it.

As described above, according to the present embodiment, even when the position of the sound source changes, the sound source position is changed by the sound source position changing means, so that the position of the sound source does not deviate from the dead point, and as a result, the sound source moves. Only the sound emitted from the sound source can be input.
<List of references>
Reference 1: Shigeru Ando “Velocity Vector Distribution Measurement System Using Spatiotemporal Differential Arithmetic of Images” Transactions of the Society of Instrument and Control Engineers 22-12, 1330/1336 (1986)
Reference 2: Shigeru Ando, Hiroyuki Shinoda, Katsuya Ogawa, Satoshi Mitsuyama "Three-dimensional sound source localization sensor system based on spatiotemporal gradient method" Vol. 29, No. 5, p520-528, 1993
Reference 3: N. Ono, T. Arita, Y. Senjo, and S. Ando, “Directivity steering principle for biomimicry silicon microphone”, Proc. Int. Conf. Solid State Sensors, Actuators, and Microsystems (Transducers '05) , pp. 792-795, 2005.
Reference 4: Ono, Ando, “Measurement of sound field and directivity control, 22nd Sensing Forum document, pp. 305-310, 2005.
Reference 5: Ono, Arita, Chiaki, Ando, “Theory of directivity control and sound source separation based on spatiotemporal gradient measurement, Proc. Of the Spring Meeting of the Acoustical Society of Japan 2005, 2-6-13, pp. 607-608, 2005.
Reference 6: SFBoll "Suppression of Acoustic Noise in Speech. Using Spectral Subtraction" IEEE Trans.on. Acoustics, Speech and Signal Processing Vol. ASSP-27, No. 2, pp. 113-1, 1979
Reference 7: Junji Ono, Akihito Saito, Shigeru Ando “Theory and Experiments of Localization Sensors for Miniaturized Sound Sources Simulating Drosophila (2nd Report)”, 19th Sensing Forum, pp.379-382,2002
Reference 8: Junji Ono, Akihito Saito, Shigeru Ando, “Theory and Experiment of Differential Detection Type Sound Source Localization Sensor Imitating Mistlefly”, Auditory Society, pp.187-192,2002

It is a block diagram which shows Embodiment 1 of this invention. It is explanatory drawing for demonstrating the spatiotemporal gradient method in the same as the above. It is explanatory drawing same as the above. The gimbal structure type microphone is shown, wherein (a) is a plan view of the diaphragm and (b) is a cross-sectional view. It is explanatory drawing of the target speaker audio | voice extraction means in Embodiment 2 of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Sound collection sensor means 2 Dead point formation means 3 Target speaker voice extraction means

Claims (7)

  Sound collecting means for detecting sound pressure, time differential value of the sound pressure, spatial differential value obtained by differentiating the sound pressure in the direction of each axis of the two-dimensional orthogonal coordinate system, and sound pressure detected by the sound collecting sensor means The dead point at which the sound collection sensitivity is minimized is formed at a preset target speaker position by performing a load sum with a predetermined coefficient vector and a low-pass filter process on the time differential value and the spatial differential value. Target speaker voice extraction for extracting only the sound pressure of the voice uttered from the target speaker using the dead point forming means and the sound pressure detected by the sound collecting sensor means and the sound pressure output from the dead point forming means And a sound input device.   2. The sound input device according to claim 1, wherein the target speaker voice extraction means extracts the sound pressure of the sound emitted from the sound source by a spectral subtraction method.   The sound input device according to claim 1, wherein the target speaker voice extraction unit extracts a sound pressure of a sound emitted from the sound source by independent component analysis.   4. The acoustic input device according to claim 3, wherein the target speaker voice extraction means performs principal component analysis before performing independent component analysis.   The sound input device according to any one of claims 1 to 4, wherein the sound collection sensor means includes a plurality of microphones arranged in a direction orthogonal to each axis of the two-dimensional orthogonal coordinate system. .   5. The sound input according to claim 1, wherein the sound collection sensor means includes a microphone having a biaxial orthogonal gimbal structure in which a diaphragm is supported at one central point. apparatus.   The sound source position changing means for changing the sound source position, the sound source position changing means is a value obtained by dividing the instantaneous power of the sound pressure detected by the sound collecting sensor means by the instantaneous power of the sound pressure output from the dead point forming means. The sound source position is changed to a position estimated based on a sound pressure, a time differential value, and a spatial differential value detected by the sound collection sensor means when the value becomes equal to or greater than a predetermined threshold value. The acoustic input device according to any one of 1 to 6.