JP2000276193A - Signal source separating method applied with repetitive echo removing method and recording medium where same method is recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily provide a reverse filter which includes a zero point. SOLUTION: Methods 'adaptive notch filter' and 'beam homing' which separate a sound source by using a 'microphone array' having multiple microphones 1,..., i,..., M are applied with a repetitive echo removing method' which optimizes an update coefficient by using a 'learning fixing method' as a method for finding approximately a reverse filter correcting the delay between microphones.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a plurality of input devices, a signal source separation method having a plurality of input signals and separating a desired signal from the plurality of signals, and a recording medium on which the method is recorded. It is.

[0002]

2. Description of the Related Art As a signal source separation method for extracting and separating only a desired signal from a plurality of signal sources coming from a plurality of different directions, "source separation using an adaptive notch filter" and "beam forming are used. Although there is "sound source separation", the husband and wife will be briefly described.

[Sound Source Separation Using Adaptive Notch Filter] When sound source separation is performed using an adaptive notch filter, a signal not including each sound source signal from a signal observed by a microphone array in a situation where a plurality of sound sources exist. Create Next, a delay component between the sound source signal and the microphone is separated by z-transforming these signals. The intended sound source is separated by using an inverse filter for correcting a delay between the obtained signal and the microphone.

[Source Separation Using Beamforming]
When sound source separation is performed using beamforming, first, synchronous addition is performed from a signal observed by a microphone array in a situation where a plurality of sound sources exist to increase the sensitivity in a specific sound source direction. Next, a signal containing no signal component of a specific sound source is created using the reference microphone and other microphones, and signals from the respective directions are obtained from these signals using an inverse filter that corrects a delay between the microphones. . Further, when a signal obtained by subtracting these signals from the signal subjected to the synchronous addition is output, a directional characteristic having high sensitivity in a specific sound source direction is obtained.

[0005]

In both the case of using an adaptive notch filter and the case of using beamforming, it is necessary to find an inverse filter including a zero point. This inverse filter has a problem that even if it can be expressed formally, it often becomes an unstable filter and cannot be realized.

[0006]

In order to solve the above-mentioned problems,
The signal source separation method of the present invention employs a “repeated echo removal method” for optimizing an update coefficient using a “learning fixed method” as a method for approximately realizing the characteristics of an inverse filter. .

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The signal source separation method according to the first aspect of the present invention employs an adaptive notch as a method for performing sound source separation when extracting a signal using a microphone array in an environment where a plurality of sound sources exist. In a system that uses either the "filter" or the "beamforming", the inverse filter is based on the "iterative echo elimination method" that uses a "learning fixed method" to optimize the update coefficient. This has the effect that an inverse filter including difficult zeros can be created.

In the recording medium according to the present invention, when extracting a signal using a microphone array in an environment where a plurality of sound sources exist, any one of an "adaptive notch filter" and a "beam forming" is used as a method of separating sound sources. In a system that uses か, a signal source separation method that applies the “iterative echo removal method” that optimizes the update coefficient using the “learning fixed method” for the inverse filter is recorded, and it is difficult to realize This has the effect that an inverse filter including zeros can be created.

According to a third aspect of the present invention, there is provided a signal source separation method for extracting a signal using an antenna array in an environment where a plurality of signal sources are present. In this system, the inverse filter is based on an iterative echo elimination method that optimizes the update coefficient by using a learning fixed method. It has the effect that a filter can be created.

According to the recording medium of the present invention, when extracting a signal using an antenna array in an environment where a plurality of signal sources are present, an "adaptive notch filter" and a "beam forming" can be used as a method of performing signal source separation. In the system using any one of the above, the signal source separation method applying the "iterative echo removal method" for optimizing the update coefficient using the "learning fixed method" for the inverse filter is recorded and realized. Has the effect that it is possible to create an inverse filter including zeros, which is difficult.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. (Embodiment 1) [Sound Source Separation Using Adaptive Notch Filter] A case where both the number of sound sources and the number of microphones shown in FIG. 1 are M will be described as an example. The received signal is a plane wave, and the direction of the sound source is known. In FIG. 1, 1, ...
, I,..., M are microphones.

Assuming that a signal coming from the sound source j and received by the microphone 1 is (Equation 1), the microphone i
(Equation 2) is as shown in (Equation 3).

[0013]

(Equation 1)

[0014]

(Equation 2)

[0015]

(Equation 3)

Here, t represents time, and (Equation 4) represents a time delay or advance of the microphone i corresponding to the sound source j with respect to the microphone 1. Here, if a signal component that does not include the component of the sound source signal (Equation 5) is represented by (Equation 6), (Equation 6) can be given by (Equation 7).

[0017]

(Equation 4)

[0018]

(Equation 5)

[0019]

(Equation 6)

[0020]

(Equation 7)

However, when i = M, i + 1 is replaced with 1. When both sides of (Equation 7) are z-transformed, (Equation 8) is obtained.

[0022]

(Equation 8)

Where (Equation 9) is the sampling frequency. Here, if it replaces like (Equation 10),

[0024]

(Equation 9)

[0025]

(Equation 10)

(Equation 8) can be rewritten as (Equation 11).

[0027]

[Equation 11]

The relationship of (Equation 11) is obtained by using a matrix (Equation 1).
It can be expressed as 2).

[0029]

[0030]

(Equation 12)

Here, when the vector on the left side of (Equation 12) is represented by (Equation 13), the matrix on the right side is represented by (Equation 14), and the vector on the right side is represented by (Equation 15), as shown in (Equation 16).

[0032]

(Equation 13)

[0033]

[Equation 14]

[0034]

(Equation 15)

[0035]

(Equation 16)

By multiplying both sides of (Equation 16) by (Equation 17), (Equation 15) is obtained as in (Equation 18).

[0037]

[Equation 17]

[0038]

(Equation 18)

Accordingly, (Formula 19) to (Formula 20) can be obtained by inverse z-transformation of both sides of (Formula 18) formally. However, if (Equation 18) is (Equation 21), (Equation 2)
A filter having the characteristic of 2) is included. Such an inverse filter is often impossible to realize even if it can be expressed formally. This is because (Expression 22) becomes an unstable filter when the zero of (Expression 23) is not included in the unit circle on the complex z plane.

[0040]

[Equation 19]

[0041]

(Equation 20)

[0042]

(Equation 21)

[0043]

(Equation 22)

[0044]

(Equation 23)

Therefore, in the present invention, as a method of approximately realizing the characteristics of the inverse filter, the method is realized by applying a “repetitive echo removal method” for optimizing an update coefficient using a “learning fixed method”. .

Hereinafter, description will be made using equations.

Assuming that the original sound in the frequency domain is (Equation 24), the impulse response is (Equation 25), the observation signal is (Equation 26), and the observation noise is (Equation 27), these relational expressions are represented by (Equation 2).
It becomes like 8).

[0048]

(Equation 24)

[0049]

(Equation 25)

[0050]

(Equation 26)

[0051]

[Equation 27]

[0052]

[Equation 28]

Thereafter, in the ideal case of (Equation 29), the calculation is expanded for an arbitrary (Equation 30).

[0054]

(Equation 29)

[0055]

[Equation 30]

If the difference between the estimated value of the observed signal and the observed signal (Equation 26) when the estimated value (Equation 31) is used instead of the original sound (Equation 24) in (Equation 28) is (Equation 32) , (Equation 32) is represented by (Equation 33).

[0057]

(Equation 31)

[0058]

(Equation 32)

[0059]

[Equation 33]

The square of (Equation 32) is partially differentiated with the estimated value of the original sound (Equation 31) to obtain (Equation 34).

[0061]

(Equation 34)

Using (Equation 34), the update equation at the k-th iteration is expressed by (Equation 35).

[0063]

(Equation 35)

Here, μ is a step gain.

Next, μ that minimizes the square of the error is determined.
(Equation 3) when (Equation 36) obtained by (Equation 35) is used
Let (Formula 32) of 3) be (Formula 37).

[0066]

[Equation 36]

[0067]

(37)

[0068]

(38)

From (Equation 38), in the case of (Equation 39), (Equation 4)
0), and the error square is minimized.

[0070]

[Equation 39]

[0071]

(Equation 40)

By substituting this μ into (Expression 35), a complex learning identification method is obtained.

[0073]

[Equation 41]

Here, ^* represents a complex conjugate. (Equation 41)
If μ is newly set as (Equation 42) to make it easier to see, and an update expression is described including (Equation 32), (Equation 43), (Equation 4)
It becomes like 4).

[0075]

(Equation 42)

[0076]

[Equation 43]

[0077]

[Equation 44]

(Equation 44) is an update equation of a general learning identification method. When (Equation 42) is used as the step gain, μ becomes a very large value at the zero point of (Equation 25) and its vicinity, and The influence of the observation noise component on the estimated value also increases.

Therefore, without fixing μ, the upper limit of μ is determined as shown in (Equation 45), (Equation 46), and (Equation 47) using the value of SNR.

[0080]

[Equation 45]

[0081]

[Equation 46]

[0082]

[Equation 47]

However, the norm interval is [0,2π), max (•, •)
Selects the larger value. (Equation 4
5), (Equation 46) and (Equation 47) are intended to suppress the influence of observation noise. Therefore, the SNR is more than the average power of (Equation 25) so that the value of μ does not increase excessively.
The threshold value γ is provided at a level lower by the value of (Expression 25), and is set to (Expression 48) only when the power of (Expression 25) is less than γ.

[0084]

[Equation 48]

The repetition termination condition is set as shown in (Expression 49).

[0086]

[Equation 49]

Note that the norm interval is [0, 2π).

As described above, in the sound source separation using the adaptive notch filter, there is an effect that an inverse filter including a zero point, which is difficult to realize, can be created by a simple method.

[Source Separation Using Beamforming] The number M of microphones and the number M of directions to be controlled shown in FIG.
An example will be described. It is assumed that the received signal is a plane wave. The directional characteristic is controlled by having the maximum sensitivity to the signal arriving from the direction of the control point 1 and lowering the sensitivity in the directions of the control points 2 to M. A signal arriving from the direction of the control point j and received by the microphone 1 is represented by (Equation 5).
0), the output of the microphone i (Equation 51) becomes as shown in (Equation 52).

[0090]

[Equation 50]

[0091]

(Equation 51)

[0092]

(Equation 52)

Here, t represents time, and (Equation 53) represents time delay or advance of the microphone i corresponding to the control point j with respect to the microphone 1.

[0094]

(Equation 53)

First, in order to increase the sensitivity in the direction of the control point 1, synchronous addition of signals received by all microphones is performed.

[0096]

(Equation 54)

When both sides of (Equation 54) are z-transformed, (Equation 5)
It becomes like 5).

[0098]

[Equation 55]

Here, (Equation 56) is a sampling frequency.

[0100]

[Equation 56]

Here, by replacing as shown in (Equation 57),

[0102]

[Equation 57]

(Equation 55) can be rewritten as (Equation 58).

[0104]

[Equation 58]

The relationship of (Equation 58) can be expressed as (Equation 59) using a vector.

[0106]

[Equation 59]

Here, a vector having an element of (Equation 60) on the right side of (Equation 59) as an element (Equation 61) and a vector having an element of (Equation 62) on the right side as (Equation 63) are expressed as (Equation 63). 64)
become that way.

[0108]

[Equation 60]

[0109]

[Equation 61]

[0110]

(Equation 62)

[0111]

[Equation 63]

[0112]

[Equation 64]

On the other hand, a signal component not including the component of the signal (Equation 66) from the direction of the control point 1 obtained by using the signals received by the microphone 1 and the microphone (Equation 65) is represented by (Equation 67). Then, (Equation 67) becomes (Equation 6)
8).

[0114]

[Equation 65]

[0115]

[Equation 66]

[0116]

[Equation 67]

[0117]

[Equation 68]

When both sides of (Equation 68) are z-transformed, (Equation 6)
It becomes like 9).

[0119]

[Equation 69]

Here, by replacing as shown in (Equation 70),

[0121]

[Equation 70]

(Equation 69) can be rewritten as (Equation 71).

[0123]

[Equation 71]

The relationship of (Expression 71) is obtained by using a matrix (Expression 7).
It can be expressed as 2).

[0125]

[Equation 72]

Here, when the vector on the left side of (Equation 72) is expressed by (Equation 73), the matrix on the right side is expressed by (Equation 74), and the vector on the right side is expressed by (Equation 75), the following expression is obtained.

[0127]

[Equation 73]

[0128]

[Equation 74]

[0129]

[Equation 75]

[0130]

[Equation 76]

From (Equation 64) multiplied by (Equation 78), (Equation 78) is subtracted from (Equation 76) multiplied by (Equation 79) and (Equation 80). Is expressed as (Expression 81), as shown in (Expression 82).

[0132]

[Equation 77]

[0133]

[Equation 78]

[0134]

[Expression 79]

[0135]

[Equation 80]

[0136]

[Equation 81]

[0137]

(Equation 82)

By dividing both sides of (Equation 82) by (Equation 83),
(Equation 84), which is a z-converted value of a signal that emphasizes the signal (Equation 66) coming from the direction of the control point 1, is expressed by (Equation 85).

[0139]

[Equation 83]

[0140]

[Equation 84]

[0141]

[Equation 85]

Therefore, when formally (formula 86) obtained by inverse z-transformation of both sides of (formula 85) is output, directional characteristics with high sensitivity can be formed in the direction of the control point 1. However, as described for the adaptive notch filter, (Equation 85) can often be expressed in a formal manner, but often cannot be realized.

[0143]

[Equation 86]

Therefore, in the present invention, as a method of approximately realizing the characteristics of the inverse filter, the method is realized by applying the “iterative echo removal method” for optimizing the update coefficient using the “learning fixed method”. .

Note that the method of realization is the same as the method of realizing the adaptive notch filter, and a description thereof will be omitted.

As described above, in sound source separation using beamforming, there is an effect that an inverse filter including zero points, which is difficult to realize, can be created by a simple method.

As described above, according to the first embodiment, there is an effect that an inverse filter including zero points, which is difficult to realize, can be created. (Embodiment 2) FIG. 2 is a view similar to FIG. 1 used in Embodiment 1.
In the above, "sound source" was changed to "signal source" and "microphone array" was changed to "antenna array".
The method of realizing this is as follows. In “Embodiment 1”, “sound source” is used as “signal source”, and “microphone” is used as “antenna”.
And the description is omitted.

According to the second embodiment, in the signal source separation using the adaptive notch filter, there is an effect that an inverse filter including a zero point, which is difficult to realize, can be created by a simple method.

In the signal source separation using the beam forming, there is an effect that an inverse filter including a zero point, which is difficult to realize, can be created by a simple method.

[0150]

As described above, according to the present invention,
In systems using a microphone array or antenna array, when performing signal separation using an adaptive notch filter and beamforming, the `` learning fixed method '' is used as a method to approximately realize the characteristics of an inverse filter including zeros, which is difficult to realize. A practical inverse filter can be created by applying the “iterative echo removal method” for optimizing the update coefficient using.

[Brief description of the drawings]

FIG. 1 is a configuration diagram for explaining a signal separation method using a microphone array according to a first embodiment of the present invention;

FIG. 2 is a configuration diagram for explaining a signal separation method using an antenna array according to a second embodiment of the present invention.

[Explanation of symbols]

 1 microphone

 ──────────────────────────────────────────────────の Continued on the front page (72) Inventor Akira Koda 497-12 Nagayoshicho, Kagoshima City, Kagoshima Prefecture Takashi Takataka 6-11-15 Kuroge, Kumamoto City, Kumamoto Pref. Inside Electric Industrial Co., Ltd. (72) Inventor Shogo Iwamura 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. F-term (reference) 2G064 AB16 BD02 CC02 CC13 5D020 CC03 CE02

Claims (4)

[Claims] When extracting a signal using a microphone array in an environment in which a plurality of sound sources exist, a system using any one of an "adaptive notch filter" and a "beam forming" as a method of performing sound source separation. A signal source separation method that applies a "repeated echo cancellation method" that optimizes update coefficients using a "learning fixed method" as a filter. When extracting a signal using a microphone array in an environment in which a plurality of sound sources exist, a system using either an "adaptive notch filter" or "beam forming" as a method of separating sound sources is used. A recording medium in which a signal source separation method is applied, which employs a "repeated echo removal method" for optimizing an update coefficient using a "learning fixed method" as a filter. 3. A system using one of an "adaptive notch filter" and a "beam forming" as a method of performing signal source separation when extracting a signal using an antenna array in an environment where a plurality of signal sources exist. A signal source separation method applying a "repeated echo removal method" for optimizing an update coefficient using a "learning fixed method" for an inverse filter. 4. A system using either an "adaptive notch filter" or "beamforming" as a method of performing signal source separation when extracting a signal using an antenna array in an environment where a plurality of signal sources exist. A recording medium in which a signal source separation method is applied, which employs an "iterative echo removal method" for optimizing an update coefficient using an inverse filter using a "learning fixed method".