JPH08250981A - Estimation device for filter coefficient - Google Patents

Abstract

PURPOSE: To realize high speed convergence and reduction in a response residue obtained after the convergence simultaneously when a coefficient for an adaptive acyclic filter simulating a characteristic of an unknown signal transmission system is estimated. CONSTITUTION: The device is made up of a coefficient update amount calculation means 100 calculating an amount required for updating a coefficient in the LMS method, a coefficient update means 110 executing update of the coefficient, and a block length decision means 120 deciding a block length from a shift of the coefficient update amount. Then the coefficient update means 110 updates the coefficient of a cyclic filter based on the coefficient update amount approximated by an estimate error of the coefficient from the coefficient update amount calculation means 100. Furthermore, the block length decision means 120 observes a shift of the coefficient update amount to adjust the block length defined by the LMS method of the coefficient update amount calculation means 100.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a filter coefficient for estimating the coefficient of an adaptive acyclic filter simulating the characteristics of a signal transfer system from a known signal sent to a signal transfer system of unknown characteristics and its response. Estimation device.

The impulse response of a signal transmission system of unknown characteristics can be calculated as a coefficient of an acyclic adaptive filter, and the most general algorithm for the calculation is a learning identification method (Normalized Least Mean Square Algorithm).
ithm) is known.

[0003]

2. Description of the Related Art FIG. 16 is a diagram showing a schematic configuration of a system for estimating characteristics of a signal transmission system. The illustrated system to which the present invention is applied includes a signal transmission system 200 having an unknown characteristic, which is connected to input a reference signal X _j , and an acyclic (F
IR: finite impulse response) filter 210, coefficient updating circuit 220, and signal transmission system 2 having unknown characteristics
Adder 230 for adding the output g _{j of} 00 and the disturbance N _j
And the output Y _j of the adder 230 and the acyclic filter 2
The subtractor 240 outputs the difference E _j from the output G _{j of} 10 to the coefficient updating circuit 220.

Here, the response of the signal transmission system 200 whose characteristics are unknown is

[0005]

## EQU1 ## g _j = Σh _j (i) X _j (i) (1) Where j is the time (sample time index,
iteration), Σ is the addition of i = 1 to I, h _j (i) is the i-th sample value of the impulse response h _j of the signal transmission system 200, and X _j (i) is the sample value of the reference signal X _j . The signal delayed by i sampling periods, and I is the maximum delay of the impulse response of the signal transmission system 200. The output of the adder 230 is

[0006]

## EQU2 ## Y _j = g _j + N _j (2), and the pseudo response synthesized by the acyclic filter 210 is

[0007]

## EQU3 ## G _j = ΣH _j (i) X _j (i) (3) Here, H _j (i) is the i-th sample value of the impulse response H _{j that} the acyclic filter 210 has. The output of the subtractor 240 is

[0008]

## EQU4 ## E _j = Y _j −G _j (4)

According to this system, if the disturbance N _j is not taken into consideration, the same reference signal X _j is input to the signal transfer system 200 and the non-recursive filter 210 whose characteristics are unknown, and signal transfer with unknown characteristics is performed. observing the difference E _j between the response g _j and the pseudo response G _j synthesized in non-recursive filter 210 in system 200, when the difference E _j is the smallest, the coefficient of the FIR filter 210 characteristics Unknown signal transmission system 20
It can be said that the impulse response of 0 was simulated.

It should be noted that an adder 230 for adding the response g _j of the signal transmission system 200 of unknown characteristics and the disturbance N _j , and its output Y _j and the pseudo response G _j synthesized by the acyclic filter 210. The subtractor 240 that generates the difference E _j may not be provided as a concrete device or circuit.

There is also an example in which there is another unknown signal transfer system between the difference E _j and the coefficient updating circuit 220. In that case, the coefficient updating circuit 220 has another input signal to the input side of the reference signal X _j. A filter that simulates the characteristics of one unknown signal transmission system is placed in front. However, the adaptive algorithm (individually-normalized LMS method: Individually Normari) targeted by the present invention
zed Least Mean Square Algorithm or additive normalization L
MS method: Summational Normarized Least Mean Square Al
It was confirmed that gorithm) can be applied to such a system as well (Reference: Kensaku Fujii, Juro Ohga:
"A Study on Individual Normalized LMS Method Suitable for Fixed-Point Arithmetic", IEICE Technical Report, EA94-63, 1994-11, References "Fujii Kensaku, Oga Toshiro:" Continuous updating of adaptive filter coefficient in non-voice noise interval " ”Technical Report, EA94-78, 1994-12”), and the principle of the present invention is modified by the presence of another unknown signal transmission system and the preposition of a filter simulating the characteristic thereof. Not a thing.

Now, the coefficient updating circuit 220 uses the difference E _j.
Coefficient H _j of the non-recursive filter 210 such that
When, for example, the conventional learning identification method is adopted as the adaptive algorithm, the coefficient of the m-th tap is

[0013]

## EQU00005 ## H _{j + 1} (m) = H _j (m) + KE _j X _j (m) / ΣX _j ² (i) (5) is updated. However, K is the step gain (0 <K <2)
And the speed at which the coefficient H _j of the non-recursive filter 210 converges to the impulse response h _j of the signal transmission system 200 with unknown characteristics, and the response residual obtained after the convergence.

[0014]

## EQU6 ## It has a function of determining the size of D _j = g _j -G _j (6). For example, the convergence speed is fastest when K = 1, and conversely, the response residual D _j obtained after convergence becomes smaller as the step gain K approaches 0. Here, as a matter of course, it is desirable that the convergence is fast and the response residual obtained after the convergence is small. However, this cannot be achieved simultaneously with one fixed step gain. Therefore, in the learning identification method, conventionally, many step gain control methods have been proposed in which the step gain K is set to a value close to 1 at the initial stage of convergence and then gradually reduced to improve the convergence characteristic. For example, Yamamoto R. and Kitayama S .: “An Adaptive Echo
Canceller with Variable Step Gain Method ", IECE
Trans., E65, 1, pp. 1-8 (1979-01). Harris RW, Chabries DM and Bishop FA:
`` A Variable Step (VS) Adaptive Filter Algorithm
m ”, IEEE Trans. Acoust., Speech & Signal Proces
s., ASSP-34, 2, pp. 309-316 (1986-04). Tomoko Sawabe, Minoru Matsui, Masafumi Hagiwara, Masao Nakagawa: "Acceleration of variable step adaptive digital filter", Theory of theory (A) , J
70-A, 2, pp. 1858-1860 (1977-11). Ryo Taguchi, Nozomu Hamada: "A Proposal of Variable Step Learning Identification Method", IEICE (A), J71-A, 8, pp. 1663-1665 (1988-0
8). Akihito Kato, Hajime Kubota: "Proposal of variable step algorithm using fuzzy control", IEICE (A), J73-A, 11,
pp. 1844-1850 (1990-11). Kensaku Fujii, Toshiro Oga: "Optimal control of modification constants in learning identification method", IEICE (A), J75-A, 6, pp.975-983 (199)
2-06).

In contrast to this learning identification method, the coefficient updating formula by the individual normalized LMS method of the document derived by the present inventor as an adaptive algorithm suitable for fixed point arithmetic.

[0016]

[Equation 7] H _{n + 1} (m) = H _n (m) + KA _n (m) / P _n (m) (7), but in the application, “the response residual obtained after the convergence is small and the convergence is small. Is faster is still the desired behavior. However, in the individual normalized LMS method, the coefficient update
Do it every sampling period, for the nth update

[0017]

(Equation 8)

[0018]

[Equation 9]

The characteristic feature is that the convergence characteristic can be controlled by calculating the block length.

[0020]

This individual normalized LMS
The method is an adaptive algorithm recently derived by the inventor. Therefore, there is no conventional technique relating to the control of the convergence characteristic in the individual normalized LMS method, which is the object of the improvement of the present invention. However, the control method proposed in the conventional learning identification method can be applied to the device for estimating the characteristics of the signal transmission system, which is the application of the present invention.
However, if the control method is applied as it is, there are the following problems.

First, in the literature, in order to control the step gain, the average power of the disturbance N _j and the estimation error which changes with time

[0022]

It has been shown that the root mean square value of Δ _j (i) = h _j (i) −H _j (i) (10) is required. However, the power of the response g _j of the signal transmission system obtained by applying the reference signal is generally much larger than that of the disturbance N _j , and thus the average of the disturbance N _j buried in the response g _j of the signal transmission system. It is difficult to know the power at the same time as the operation starts. Further, the impulse response h _j (i) is, of course, unknown, so that the estimation error Δ _j (i) cannot be observed. In the conventional method-, therefore, the estimation error Δ _j (i) is substituted by the difference E _j , and the step gain that maximizes the convergence speed is calculated from the observed value.

The problem is that the equation giving the difference E _j

[0024]

As can be seen from E _j = ΣΔ _j (i) X _j (i) + N _j (11), the estimation error Δ _j (i) becomes small and the step gain control is essential. The estimation error Δ _j (i) is sometimes buried in the disturbance N _j, and the magnitude of the estimation error necessary for controlling the step gain cannot be observed.

FIG. 17 is a diagram showing the estimation accuracy when the step gain K = 1/16. In this figure, (a) is an estimation error Δ _j calculated by giving the tap number I of the acyclic filter 210 = 512, the step gain K = 1/16, and the power ratio of the response of the signal transmission system 200 to the disturbance of 30 dB. (i) short-time root mean square value, (b) calculated under the same conditions

[0026]

(Equation 12)

The transition of the short-time root mean square value of is shown.
Here, instead of the difference E _j , the short-time root mean square value of the equation (12) is used as the control signal of the step gain because the influence of the reference signal X _j contained in the difference E _j is canceled by the numerator and the denominator. , To suppress.

FIG. 18 is a diagram showing the estimation accuracy when the step gain K = 1/32. FIG. 18 shows the result of calculation under the same conditions as in FIG. 17 with a step gain of 1/32. As is clear from FIGS. 17 and 18, by comparing the convergence characteristics of (a), it can be seen that the estimation accuracy is improved by reducing the step gain, but in (b) which can be actually observed, Even if the step gain is reduced after the saturation, the root mean square value of the estimation error is the same, and therefore it is impossible to observe the change. This means that the effect of controlling the step gain cannot be observed using the difference E _j . Naturally, the calculation of the next control value will be impossible unless this effect is observed. The above methods (1) to (3) do not take this fact into consideration.

In this respect, the method of the literature is based on the difference E _j
The difference is that the power ratio between the response and the response Y _j including the disturbance N _j (approximated to the short-time root mean square value) is used for control, but after saturation, the theoretically calculated optimum step gain is obtained. Therefore, it is possible to control after saturation. However, this method has a problem in that it is necessary to calculate the logarithm of the power ratio and the optimum step gain value, and an increase in the processing amount cannot be avoided.

The present invention has been made in view of the above circumstances, and provides a filter coefficient estimating device for estimating the coefficient of a filter that simultaneously realizes speeding up of convergence and reduction of a response residual obtained after convergence. The purpose is to do.

[0031]

FIG. 1 is a principle diagram of the present invention for achieving the above object. The filter coefficient estimation apparatus according to the present invention calculates coefficient update amount calculation means 100 for calculating the amounts A _n (m) and P _n (m) required for coefficient update in the individual normalized LMS method, and formulas for updating coefficients ( 7) according to 7), and the block length determining means 120 for determining the block length from the transition of the coefficient update amount A _n (m) / P _n (m).

[0032]

According to the above-mentioned means, the filter coefficient estimating device of the present invention has the characteristic that the size of the estimation error can be observed even after the convergence if the block length is increased in accordance with the decrease of the estimation error. It is used to control the convergence speed.
The block length determining means 120 is the coefficient update amount calculating means 100.
The transition of the sum of squares of the coefficient update amount calculated in step 1 is observed, and the block length J is determined according to the size. As a result, the speeding up of the convergence and the reduction of the response residual after the convergence are realized at the same time.

[0033]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings, but before that, the operating principle of the present invention will be described first.

For the purpose of this description, first the components of equation (8) are

[0035]

[Equation 13] E _j X _j (m) = [ΣΔ _j (i) X _j (i) + N _j ] X _j (m) = Δ _j (m) X _j ² (m) + [Σ _m Δ _j ( i) X _j (i) + N _j ] X _j (m) ... (13) Here, Σ _m is the addition of i = 1 to I excluding i = m. That is, when the equation (8) is expressed as described above, the information necessary for updating the m-th coefficient of the acyclic filter 210 is the first term thereof.

[0036]

(14) It becomes clear that Δ _j (m) X _j ² (m) = [h _j (m) −H _j (m)] X _j ² (m) (14). Next, the coefficient update is performed every J sampling periods, so that time j = nJ + 1 ...
The coefficient is fixed at H _n (m) during (n + 1) J, and the characteristics of the signal transmission system (impulse response of echo path)
Pays attention to that the coefficient of the adaptive filter can be regarded as constant until it converges [h (m) = h _j (m)],

[0037]

(Equation 15)

Rename as Obviously, the necessary information [h (m) -H _j (m)] common to the coefficient update and the convergence property control is
By calculating the ratio with the power P _n (m) of the reference signal given to equation (9),

[0039]

[Equation 16]

It is extracted as the first term of. Furthermore, the first term H _n (m) in the equation (7) becomes the second term in the equation (1).
Paying attention to the fact that it is also included in 6), the individual normalized LM
When the equation (7) for updating the coefficient that constitutes the S method is arranged,

[0041]

[Equation 17] H _{n + 1} (m) = [h (m) −R _n (m)] K + (1−K) H _n (m) (17) To be However,

[0042]

(Equation 18)

It is FIG. 2 is a diagram showing a first-order cyclic filter expression of the individual normalized LMS method. According to this representation, the impulse response h of the signal transmission system whose characteristics are unknown
A first-order recursive filter whose input is the m-th sample value h (m) of (i) and whose disturbance is R _n (m) consisting of estimation errors Δ _n (i) related to the sample value of i ≠ m Are configured. Where 30
0 is a subtractor that subtracts R _n (m) from h (m), 310 is a multiplier that multiplies the output of the subtractor 300 by a step gain K, 320 is an adder of the first-order cyclic portion, and 330 is one sampling period. The delay element, and 340, is a multiplier that multiplies the previous coefficient by the cyclic coefficient (1-K).

Clearly, from this expression, the convergence characteristic of the coefficient H _n of the non-recursive filter 210 is given as the response of this first-order recursive filter, and its convergence speed is the cyclic coefficient (1- K) and the operation interval (block length J), that is, the individual normalized LM
It is found that the convergence speed shown by the S method can be adjusted by the block length J in addition to the step gain K known in the conventional learning identification method.

The relationship between the convergence speed and the step gain K and the block length J shown by the individual normalized LMS method is shown below. FIG. 3 is a diagram showing the relationship between the convergence speed and the step gain, and FIG. 4 is a diagram showing the relationship between the convergence speed and the block length.

In FIG. 3, for example, when the block length J is equal to the number of taps I of the acyclic filter 210, I = J = 5.
Fixed to 12, step gain K = 1/2, 1/8,
It is a convergence characteristic (response residual reduction characteristic) calculated by changing it to 1/32. Further, in FIG. 4, the same tap number I = 512
With the step gain fixed to 1/8, the block length J is 128 (= I / 4), 512 (= I), 2048
It is the convergence characteristic obtained by changing it to (= 4I). From this result, it is apparent that the block length J in the individual normalized LMS method is
It is confirmed that gives the same effect as the step gain K for adjusting the convergence speed, for example, doubling the block length is equivalent to halving the step gain. From this, it can be seen that it is also effective to control the convergence characteristic by the block length.

Here, the block length J and the step gain K
The difference between these two controls is that the above-mentioned R _n (m) (equation (16) second which disturbs the estimation of the coefficient H _n (m) when the block length J is lengthened).
Term) is reduced by the arithmetic mean effect.

Next, let us consider how the short-time root mean square of the estimation error, which does not change even if the step gain K is changed, changes when the block length J is changed. FIG. 5 is a diagram showing the estimation accuracy when the block length J = 512, and FIG. 6 is a diagram showing the estimation accuracy when the block length J = 1024.

In these figures, (a) is an individual normalized LMS
Estimation error obtained by the method

[0050]

[Formula 19]

This is the transition of the short-time root mean square value of. (B)
Is the coefficient update amount obtained for each J sampling period required for controlling the convergence characteristic, that is, m = 1 to 1 of the squared value of Expression (16).
The added value of I,

[0052]

(Equation 20)

Is a transition of. However, the tap number I, the step gain K, and the like have the same values as in FIG. 17, and the same transition is calculated by multiplying the block length J by 512 in FIG. 5 and multiplying the block length J by 1024 in FIG. Apparently, when the block length J is doubled as compared with the result of FIG. 18 in which the step gain K is reduced, the sum of squares given to the equation (20) increases, that is, the influence of the disturbance N _j . It is confirmed that it is getting smaller. This means that the estimation error [h
If the block length J is increased in accordance with the decrease of (m) −H _n (m)], the magnitude of the estimation error necessary for controlling the convergence speed can be observed even after the convergence. This means that the convergence speed can be controlled.

The present invention uses this characteristic to observe the transition of the sum of squares of the coefficient update amount, and controls the block length J in accordance with the size of the sum. This makes the convergence faster and converges. Later, it is possible to simultaneously reduce the response residual.

Next, eight methods will be described below, which consider a method of associating the transition of the sum of squares of the coefficient update amount and the block length J, which is necessary as a means for specifically carrying out the present invention.

First, as the first method, the sum of squares of the coefficient update amounts obtained in the previous coefficient update [Equation (2
0)] is stored as EP and compared with the sum of squares ER at each time point obtained while extending the width of the block in sampling cycle units.

[0057]

[Equation 21] ER≤EP ...... When the condition (21) is satisfied, the expansion of the block width is stopped, and EP = E.
A method is conceivable in which the coefficient is replaced with R and the coefficient is updated (claim 4).

In this case, since the estimation error is always guaranteed to be smaller than that at the time of the previous update by the equation (21), the coefficient update is stably executed. FIG. 7 is a diagram showing convergence characteristics by the learning identification method and the individual normalized LMS method.

In this figure, (a) shows the number of taps I of the acyclic filter I = 512, and the step gain K = 1/16.
The convergence characteristics obtained by the fixed learning identification method are shown. (B) is a convergence characteristic obtained in the individual normalized LMS method in which the block length J = I / 4 = 128 and the step gain K = 1/16 is fixed, but the minimum width of the block is I / 8 = 64. I have a limit. (C) shows the convergence characteristics obtained in the same individual normalized LMS method, but shows the case where the block length is fixed to J = I / 4 = 128 and control is not performed.

These learning identification methods and individual normalized LM
From the result of calculating the response residual reduction characteristics obtained in the S method, the method of controlling the block length in (b) shows a convergence speed in the initial stage of convergence as compared with the case where the block length in (c) is not fixed and controlled. It is shown that there is a case where it is reversed in the middle of the learning identification method of FIG. However, in the method of controlling the block length, firstly, the convergence characteristic by the learning identification method is saturated, while the estimation error thereafter continues to decrease. Secondly,
Even if the power of the disturbance N _j is actually larger than the value assumed at the time of design, the estimation error can be automatically reduced to a required magnitude. Thirdly, if it is small, the convergence speed is automatically set. Fourthly, there is an advantage that, when the power of the disturbance N _j sharply increases after the convergence, the block length is automatically lengthened and the estimation error is controlled to a required magnitude.

As a second method for making the block length correspond, a method of setting a lower limit on the extension unit of the block in order to prevent the reversal seen in the first method and improving the convergence characteristic as compared with the learning identification method is considered. (Claim 5).

As the rate of the disturbance N _j included in the coefficient update amount used for determining the block length increases as the coefficient update progresses, the influence of the disturbance N _j on the determination gradually increases. It increases the possibility that the convergence speed will decrease due to the block length decision error. It is considered that this is the reason why the learning identification method is reversed in the calculation result of FIG. 7. That is, in order to suppress the possibility of this reversal, the proportion of the disturbance N _j included in the equation (18) may be reduced so that the estimation error is well observed. The proportion of the disturbance N _j can be reduced by setting a lower limit on the extension unit of the block.

FIG. 8 is a diagram showing the convergence characteristic when the lower limit is set for the extension unit of the block. In the illustrated example, (b)
The response residual reduction characteristic is calculated by setting the lower limit of the extension unit of the block to 64 under the condition of FIG. As is clear from this figure, in the initial stage of convergence, the learning identification method (a) is inferior to the individual normalized LMS method (c) in which the block length J = I / 4 = 128 and the step gain K = 1/16 are fixed. ) Has not been reversed. The convergence speed is almost double that of the learning identification method.

As the third method for making the block length correspond, the cause of the decrease in the convergence speed in the initial stage of convergence, which is observed in the second method, is considered as follows. That is, the sum of squares F _n of the coefficient update amount calculated using the equation (20) may appear small due to the influence of the disturbance N _j . In this case, the next coefficient update must be below this reduced value before it is executed. Once this sum of squares F _n becomes small,
A block with a long width is required for subsequent coefficient update, and as a result, the convergence speed decreases. This drawback can be solved by suppressing the influence of the disturbance N _{j on} the sum of squares F _n , and the suppression can be achieved by using the output after passing the sum of squares F _n through a low-pass filter ( Claim 6).

FIG. 9 is a diagram showing the convergence characteristics when the influence of disturbance is suppressed by the low-pass filter. In this figure, (a) is the reduction characteristic of the estimation error by the learning identification method. In (d) and (e), the lower limit of the extension unit of the block is 64 with respect to the value used in the calculation of FIG. 7, and the cyclic coefficient (1-K) of the low-pass filter is (1-1 / 16), respectively.
And (1-1 / 8) are the estimation error reduction characteristics obtained. Obviously, it is confirmed that the fluctuation due to the disturbance N _j is suppressed and the decrease in the convergence speed in the initial stage of the convergence is prevented.

The influence of this disturbance N _j is expressed by the equation (1
According to 5), it is clear that this can also be suppressed by lengthening the block. FIG. 10 is a diagram showing the convergence characteristic when the block length is extended.

In this figure, (f) shows the result when the extension unit of the block is set to 512, which is equal to the number of taps of the non-recursive filter, from 64 given in FIGS. 7 and 8.
However, the step gain K can be made larger than the block length I / 8 [see literature], and it is set to 0.5. That is, according to the result given in FIG. 10, the influence of the disturbance N _j is suppressed as a result of increasing the lower limit of the block width, and the block length is 512 at the initial stage of convergence.
It is confirmed that the convergence speed is stably improved.
However, increasing the extension unit of the block has the effect of roughening the adjustment of the block length, and when the block length control becomes necessary, the roughness becomes a drawback and the convergence speed is reversed by the learning identification method. It is supposed to be.

As a fourth method of making the block length correspond, the drawback found in the third method is solved by controlling the step gain in addition to the block length (claim 7). .

The step gain can be calculated, for example, as follows. That is, when the block length M _T is required to fall below the minimum value EP of the sum of squares of the coefficient update amounts obtained in the previous coefficient update, the block length is relative to the basic block length M _B initially determined. (M _T /
M _B ). When this extension effect is converted into a step gain, the basic block length M _B defined at the beginning is calculated.
Equal to the step gain (M _B / M _T) multiplied by that relative in step gain K _B to ensure stable convergence. Therefore, the step gain K _T changed at this point
To

[0070]

[Equation 22] By converting K _T = K _B M _B / M _T (22) and updating the coefficient, control of the convergence characteristic is realized.

FIG. 11 is a diagram showing the convergence characteristic by the method to which the step gain control is added. In this figure,
(G) sets the initially determined basic block length M _B to I = 512, which is equal to the number of taps of the adaptive filter, and also provides a basic step gain K that guarantees stable convergence for the block length M _{B. This} is the response residual reduction characteristic calculated by giving _B as 0.5. In this case, the step gain is

[0072]

## EQU23 ## K _T = K _B I / M = 256 / M (23) Obviously, it can be seen that the convergence is faster than the decrease characteristic obtained by the second method shown in FIG. 8 and that the convergence speed is almost equal to that of the third method.

Next, as one of the calculation methods of the coefficient update amount by the individual normalization LMS method, "the number of coefficients updated per sampling period is set to be one, which has the same processing amount as the learning identification method. ] There is a distributed update method. However, with this calculation method, there is a difficulty in that the time at which it becomes possible to determine whether to execute / postpone the coefficient update is the time when the coefficient update ends for all coefficients. A convergence speed control method suitable for this calculation method is given below.

First, when the number of coefficients updated per sampling period in the individual normalized LMS method is one, equations (8) and (9) are

[0075]

[Equation 24]

[0076]

(Equation 25)

It is rewritten as [reference]. That is, the normalized power becomes common to all the coefficients, and the coefficient update is performed once in one sampling period, so that the calculation amount can be reduced. However, in this updating method, the coefficient updating amount used for controlling the convergence speed can be known only for one coefficient in one sampling period, and therefore, when the coefficient updating amount is calculated for all coefficients, the last coefficient updating amount is calculated. It means that the update has already been completed except for. Naturally, this problem does not occur if the converging speed is controlled in units of coefficients. However, the variance of the coefficient update amount calculated by the equation (16) per coefficient is large, and its control is actually difficult. That is, the coefficient update amount needs to be treated as an average of all coefficients. Hereinafter, a method of treating the average will be described.

In the fifth method, first, the coefficient update having the common normalized power P _n (m) given in the equation (25) is considered as one block, and one coefficient is obtained for each sampling period. The result C of sequentially adding the squared values of the coefficient update amounts
The minimum value D of the coefficient update amount obtained by the coefficient update up to the previous time is sequentially compared, and control is performed as follows. That is, (i) When the addition result exceeds the minimum value, the coefficient updating thereafter is suspended. However, regarding the coefficient update amount, the sum of all coefficients is obtained as C. In this case, the estimation error is C
It can be considered that it has become / D times. Therefore, the step gain K _b in the next coefficient update is set with respect to the current step gain K _a , for example,

[0079]

## EQU26 ## By rewriting as K _b = K _a D / C (26), the increase of the estimation error can be suppressed (claim 8).

Further, in the distributed updating method, which is a method of reducing the calculation amount in the individual normalized LMS method, the block length can be set only to an integral multiple of the number of taps, so that the block length cannot be finely adjusted. Therefore, the block length is changed by changing the block length M _a to the basic block length M _B when the changed step gain is less than (1/2) ^k times the basic step gain K _B defined for the basic block length. Against, for example,

[0081]

[Number 27] M _a = (k + 1) to change the M _B ...... (27). That is, the block length is changed to a length converted from the obtained corrected step gain (claim 9). (Ii) When it does not exceed, the current block length and step gain are maintained.

FIG. 12 is a diagram showing a convergence characteristic when the step gain control is added to the distributed updating method. In this figure, (g) shows the response residual reduction characteristic obtained by this control. As is clear from the figure, it is confirmed that the result close to that of FIG. 11 obtained by the basic calculation method of the individual normalized LMS method is obtained.

Next, the sixth method will be described. In the fifth method described above, the coefficient update amount obtained at a rate of one coefficient for each sampling cycle is added in order, and the result C exceeds the minimum value D of the coefficient update amount obtained by the previous coefficient update. At this time, a method of suspending the subsequent coefficient updating is adopted, but if the coefficient updating is executed up to the m-th coefficient with respect to the tap number I, the step gain is

[0084]

Equation 28 It is also possible to _{K b = K a m / I} ...... (28) and by reducing to continue subsequent coefficient update (claims 10, 11).

Further, as in the fifth method, it is possible to suspend the coefficient update and correct the step gain in the fifth method by the tap number ratio (claim 9). Further, as the seventh method, it is possible to control only the step gain, which will be described. If the convergence characteristic after saturation is allowed to deteriorate as compared with the case where the block length is adjusted, control with only the step gain (block length is fixed) is also possible.

FIG. 13 is a diagram showing a convergence characteristic obtained only by controlling the step gain. In this figure, (h)
Is a response residual reduction characteristic obtained by the control that the step gain gives by the sixth method. According to this characteristic, it is confirmed that the convergence speed is controlled, though it is inferior to the results of FIGS. 11 and 12.

Finally, the eighth method will be described. By the way, it is fully expected that the power of the disturbance N _j fluctuates. The present invention can automatically respond to this variation. That is, with respect to the increase of the power, the coefficient update is not executed until the relation of the equation (21) is satisfied, and the block length or the step gain is automatically adjusted so that the estimation error becomes equal to or smaller than the magnitude before the increase of the power. Adjusted to. Further, as the power of the disturbance N _j decreases, the coefficient update amount given to the equation (16) decreases, so that the block length M _T is reduced in the fourth method, and the coefficient update amount is reduced in the fifth method. From the increase of D / C, the step gain is automatically increased from the increase of m / I in the sixth method, and the block length is determined corresponding to the step gain. It will be automatically shortened.

FIG. 14 is a diagram showing changes in the convergence characteristics due to fluctuations in the power of disturbance. This figure shows a decrease characteristic of the response residual when the power of the disturbance N _j is increased (the power ratio of the response to the disturbance is set to 10 dB) and then returned again to the seventh method. This is the result of calculation. In the conventional learning identification method shown in (a), the response residual increases as the power of the disturbance N _j increases, whereas in the present invention, the response residual increases despite the power of the disturbance N _j. It is confirmed that the response residual is starting to decrease again when the power is restored to the original level. From this,
It can be seen that it has a strong characteristic against the fluctuation of the disturbance.

FIG. 15 is a diagram showing the configuration of an acoustic echo canceller to which the filter coefficient estimating device is applied. In this figure, 400 is a microphone, 410 is a speaker,
420 is a subtractor, 430 is a non-recursive filter, and 440 is a coefficient estimation circuit. Here, the echo path between the speaker 410 and the microphone 400 is an unknown signal transmission system, and the coefficient estimation circuit 440 corresponds to the filter coefficient estimation device shown in FIG.

Far-end speaker signal X corresponding to the reference signal_jIs
It is sent from the peaker 410 to the echo path and its response g_j
Is the disturbance N that is the ambient noise_jMicrophone 4
00 is input. Its output Y_j= G_j+ N_jIs the subtractor 4
It is input to 20. The subtractor 420 is the acyclic filter 4
Pseudo echo G made with 30_jDifference E with_j= Y_j+ G _j
Is output. The coefficient estimation circuit 440 outputs the signal X_jAnd the difference E_j
And the filter coefficient is estimated by the individual normalized LMS method.
And update the filter coefficient of the acyclic filter 430.
It

According to this acoustic echo canceller, for example, when a large door opening / closing sound is generated as noise during a call, the coefficient estimating circuit using the conventional learning identification method can be used every time the door is opened / closed. Although the coefficient may be disturbed to reduce the amount of echo cancellation and howling may occur, the coefficient estimation circuit according to the present invention is strong against external disturbance as shown in FIG. 14, and therefore howling is not caused by surrounding noise. Absent.

In the above description, the individual normalized LMS method is shown as the adaptive algorithm for calculating the coefficient update amount of the adaptive filter, but the addition normalized LMS method (reference) can be similarly applied. . Individual normalization L
The formula for updating the coefficient in the MS method is

[0093]

[Equation 29]

On the other hand, the equation for updating the coefficient in the addition-normalized LMS method is

[0095]

[Equation 30]

It is represented by In these expressions, focusing on the denominator of the second term representing the coefficient update amount, the individual normalized L
While the MS method takes out and adds only the m-th tap, the addition-normalized LMS method adds the taps by adding several taps.

[0097]

As described above, according to the present invention, the block length is controlled when updating the coefficient of the non-recursive filter. For this reason, compared to the conventional case where control is performed only with the magnitude of the step gain, the control by the step gain is replaced with the adjustment of the block length, or both the block length and the step gain are controlled to achieve higher speed. In addition, a filter coefficient estimation device that can update the adaptive filter coefficient with high accuracy is realized.

[Brief description of drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a diagram showing a first-order recursive filter expression of an individual normalized LMS method.

FIG. 3 is a diagram showing a relationship between a convergence speed and a step gain.

FIG. 4 is a diagram showing a relationship between a convergence speed and a block length.

FIG. 5 is a diagram showing estimation accuracy when a block length J = 512.

FIG. 6 is a diagram showing estimation accuracy when a block length J = 1024.

FIG. 7 is a diagram showing convergence characteristics by a learning identification method and an individual normalized LMS method.

FIG. 8 is a diagram showing a convergence characteristic when a lower limit is set for an extension unit of a block.

FIG. 9 is a diagram showing a convergence characteristic when the influence of disturbance is suppressed by a low-pass filter.

FIG. 10 is a diagram showing convergence characteristics when the block length is extended.

FIG. 11 is a diagram showing a convergence characteristic by a method in which step gain control is added.

FIG. 12 is a diagram showing a convergence characteristic when step gain control is added to the distributed updating method.

FIG. 13 is a diagram showing a convergence characteristic obtained only by controlling a step gain.

FIG. 14 is a diagram showing a change in convergence characteristic due to power fluctuation of disturbance.

FIG. 15 is a diagram showing a configuration of an acoustic echo canceller to which a filter coefficient estimating device is applied.

FIG. 16 is a diagram showing a schematic configuration of a system for estimating characteristics of a signal transmission system.

FIG. 17 is a diagram showing estimation accuracy when a step gain K = 1/16.

FIG. 18 is a diagram showing the estimation accuracy when the step gain K = 1/32.

[Explanation of symbols]

 100 coefficient update amount calculation means 110 coefficient update means 120 block length determination means 400 microphone 410 speaker 420 subtractor 430 non-recursive filter 440 coefficient estimation circuit

Claims (12)

[Claims] 1. A coefficient of a non-recursive filter that outputs a response equivalent to an impulse response of the signal transmission system from a signal transmitted to a signal transmission system of unknown characteristics and its response, of filter coefficients for estimating by a LMS method. In the estimation device, a coefficient update amount calculation means for calculating an amount necessary for coefficient update in the LMS method, and a coefficient for the non-cyclic filter based on the calculated coefficient update amount approximated to the coefficient estimation error. Coefficient updating means for updating the LMS, and observing the transition of the calculated coefficient updating amount
And a block length determining unit that adjusts the block length used for calculating the coefficient update amount according to the transition, and a filter coefficient estimating device. 2. The filter coefficient estimating apparatus according to claim 1, wherein the LMS method is an individual normalized LMS method. 3. The filter coefficient estimating apparatus according to claim 1, wherein the LMS method is an addition-normalized LMS method. 4. The block length determined by observing the transition of the coefficient update amount is until the time when a coefficient update amount less than the minimum coefficient update amount obtained in the previous coefficient update block is obtained. The filter coefficient estimating device according to claim 1, wherein the filter coefficient estimating device comprises: 5. The filter coefficient estimation device according to claim 4, wherein a lower limit is set for an extension unit of the block. 6. The filter coefficient estimation according to claim 1, further comprising a low-pass filter that receives the coefficient update amount and observes a transition of the coefficient update amount with an output value. apparatus. 7. The coefficient of a non-recursive filter that outputs a response equivalent to the impulse response of the signal transmission system is estimated from the signal transmitted to the signal transmission system of unknown characteristics and its response by the individual normalized LMS method. In the filter coefficient estimation device, the width of the block lengthened to be less than the minimum coefficient update amount obtained in the coefficient update block up to the previous time,
An apparatus for estimating a filter coefficient, wherein a step gain used for updating a coefficient at the present time is calculated from a predetermined reference block width and a reference step gain. 8. A result obtained by sequentially adding coefficient update amounts obtained at a rate of one coefficient for each sampling period when calculating the step gain when the coefficient update is executed by the distributed update method in the individual normalized LMS method. And the minimum value of the coefficient update amount obtained by the previous coefficient update are sequentially compared,
The filter coefficient estimation according to claim 7, wherein the coefficient update is stopped when the result exceeds the minimum value, and the step gain used in the next coefficient update is calculated from the result and the minimum value. apparatus. 9. A result obtained by sequentially adding coefficient update amounts obtained at a rate of one coefficient for each sampling period when calculating the step gain when the coefficient update is executed by the distributed update method in the individual normalized LMS method. And the minimum value of the coefficient update amount obtained in the previous coefficient update are sequentially compared,
When the result exceeds the minimum value, the coefficient update is suspended, and the step gain is calculated from the number of coefficients updated up to the suspension point and the number of coefficients of the recursive filter. 7. The filter coefficient estimation device according to 7. 10. When calculating the step gain,
The result obtained by sequentially adding the coefficient update amounts obtained at the rate of one coefficient for each sampling period and the minimum value of the coefficient update amounts obtained by the coefficient update up to the previous time are sequentially compared. The filter coefficient estimation according to claim 8 or 9, characterized in that the coefficient update after the point when the minimum value is exceeded is continued with the step gain used in the next coefficient update calculated from the result and the minimum value. apparatus. 11. When calculating the step gain,
The result obtained by sequentially adding the coefficient update amounts obtained at the rate of one coefficient for each sampling period and the minimum value of the coefficient update amounts obtained by the coefficient update up to the previous time are sequentially compared. 9. The coefficient update after the time point when the minimum value is exceeded is continued with a step gain calculated from the number of coefficients updated up to the pause point and the number of coefficients of the recursive filter. 9. The filter coefficient estimation device according to item 9. 12. The filter coefficient estimating apparatus according to claim 8, wherein the length of the coefficient update block is calculated from the calculated step gain.