JPH03114311A - Filter coefficient study system for adaptive filter - Google Patents

Abstract

PURPOSE:To execute the study of a filter coefficient of an adaptive filter properly at a high speed by applying processing while a filter coefficient assigned to a specific tap is sequentially updated and studied. CONSTITUTION:An integration means 24 applies processing to obtain power component while a filter coefficient Hj+1 (m) included in a signal inputted via a 1st arithmetic means 23 is emphasized. Then a 2nd arithmetic means 26 arranged at the output of the integration means 24 uses the power calculated by a power calculation means 25 to eliminate the power component included in an output of the integration means 24 thereby executing the extraction of the filter coefficient Hj+1 (m). Thus, the normalizing processing is executed through the processing in this way to execute the processing of the arithmetic system. Thus, the filter coefficient is studied properly at a high speed.

Description

[Detailed Description of the Invention] [Summary] Regarding a filter coefficient learning method for an adaptive filter that follows a learning identification method, the purpose of this invention is to reduce the learning identification method with the aim of making it possible to perform filter coefficient learning for an adaptive filter properly and quickly. By expressing it as a low-pass filter and configuring the low-pass filter coefficient of the low-pass filter to be a constant, the estimation accuracy of the filter coefficient of the adaptive filter becomes stable, and the input to the low-pass filter By configuring the disturbance to be small at the start of learning, the filter coefficients of the adaptive filter are configured to converge quickly. Even if the low-pass filter coefficients are not constants, the filter coefficients of the adaptive filter can converge quickly by configuring the disturbance input to the low-pass filter to be small at the start of learning. and,
When using the conventional learning identification method, by configuring the norm value of the normalization process to be large regardless of the number of taps of the adaptive filter, the estimation accuracy of the filter coefficients of the adaptive filter becomes stable. Configure it like this.

[Industrial application field]

The present invention relates to a filter coefficient learning method for an adaptive filter for learning filter coefficients of an adaptive filter, and in particular,
The present invention relates to a filter coefficient learning method for an adaptive filter that enables filter coefficient learning to be performed properly and quickly when learning filter coefficients according to a learning identification method.

There is an adaptive filter that performs a process of sending a signal to a system whose characteristics are unknown and estimating a transfer function (impulse response) given by the unknown system from the response. Such an adaptive filter is used, for example, in an echo canceller that shares one upstream and downstream communication path to achieve simultaneous bidirectional transmission. In order to make this adaptive filter practical, it is necessary to learn the filter coefficients properly and at high speed. From now on, we will introduce the widely used learning algorithm for filter coefficients.
Even in the case of the ``learning identification method,'' it is necessary to take measures to ensure that filter coefficient learning is performed properly and at high speed.

[Conventional technology]

According to the adaptive filter used in the echo canceller,
A method for learning filter coefficients of a conventional adaptive filter will be described.

In FIG. 26, 1 is a subscriber terminal, 2 is a hybrid transformer, 3 is a balance network, and 4 is a transversal type adaptive filter, which is output to the hybrid transformer 2. A pseudo echo is generated according to the output signal X=(j represents time). The adaptive filter 4 includes an output signal register 5, a filter coefficient register 6, a convolution integrating section 7, and a filter coefficient updating section 8.

This output signal register 5 holds a plurality of output signals XJ(i) (1 represents a tap number) output to the hybrid transformer 2 between the current time and a certain time in the past, and stores filter coefficients. Register 6 holds a filter coefficient H, (+) (i represents the tap number) set for each tap of the transversal filter,
The convolution integrator 7 calculates a pseudo echo ΣH= (i) from the output signal X, (i) held by the output signal register 5 and the filter coefficient Hj (+) held by the filter coefficient register 6.
The filter coefficient updating unit 8 generates the pseudo echo ΣH7(i)X* generated by the convolution integrating unit 7.
The filter coefficients in the filter coefficient register 6 are updated so that (i) is equal to the echo given by the actual echo path. Reference numeral 9 denotes a differential circuit, which receives an input signal Y from the hybrid transformer 2.

and the pseudo echo ΣHj (
i) The echo contained in the input signal Yj is erased by calculating the residual echo E, which is the difference from Xt(i).

FIG. 27 shows a detailed configuration of the output signal register 5, filter coefficient register 6, and convolution integrator 7 that constitute the adaptive filter 4. As shown in this figure, the output signal register 5 is configured to hold time-series data of the output signal by including a delay device 5a that delays the output signal XJ every sampling period. The inclusive integration section 7 calculates Hj(i
) A pseudo echo ΣHj (i) Configure it like this.

When learning filter coefficients according to the "learning identification method," the filter coefficient updating unit 8 updates the filter coefficient Hj (11) assigned to the tap m of the adaptive filter 4 obtained at time j and the coefficient correction constant. Then, from the residual echo E and the output signal X=(m) stored in association with the tap m of the adaptive filter 4, the filter coefficient Hj, , (-) of the adaptive filter 4 at time j+1 is calculated as ΣXJ ! (i) The filter coefficients in the filter coefficient register 6 are sequentially updated by calculating according to the equation (1). Here, ΣXi''(i) is a norm, and its cumulative addition number is selected as the number of taps I of the adaptive filter 4, so i=1
〜I. The division process by this norm is
The normalization process is performed to suppress the influence of amplitude fluctuations in the output signal Xj(m). Note that, for convenience of description, when "Σ" is written below unless otherwise specified, it represents that cumulative addition is performed for i-1 to ■, similarly to equation (1).

FIG. 28 shows the calculation system of the "learning identification method" that executes equation (1). In FIG. 0, 9 is the difference circuit explained in FIG. Pseudo echo ΣH calculated by section 7; (i) Calculating residual echo E by calculating the difference value from Xj(i), 80
81 is a multiplication circuit that calculates the multiplication value of the calculated residual echo E and the coefficient correction constant; 81 is a norm calculation circuit that calculates the norm value ΣX,''(i); 82 is a division circuit that divides the value calculated by the multiplication circuit 8 (12) by the norm value ΣXj''(i); 83 is a multiplication circuit that divides the value calculated by the division circuit 82 and the output signal X;
84 is an addition circuit that calculates the multiplication value with j(m), and the calculation value of the multiplication circuit 83 and the filter coefficient HJ(I
I), the filter coefficient H, and (intestine) at time j+1 are calculated.

In FIG. 29, 5 is an output signal register of the adaptive filter 4 explained in FIG. 26, and time-series data of the output signal those that store the number of taps, 10 and 1
1 is a multiplier for square calculation, which calculates the square value of the output signal Xp stored in the output signal register 5;
12 is a subtracter that calculates the difference between the value calculated by the square multiplier 1 (12) and the value calculated by the square multiplier 11; 13 is a cumulative addition circuit; The adder 14 adds the calculated value of the adder 14 to the output value of the delay device 15, and the delay device 15 processes the calculated value of the adder 14 by one sampling period.

In the norm calculation circuit 81 configured as described above, the output signal X1. When Xz,...X, is input,
The cumulative addition circuit 13 outputs the signal from the output signal register 5 to the square calculation multiplier 1 until (1=I (I is the number of taps)).
In response to the fact that the output signal X is not output for 1, T, -x, "+X," +...+Xq! However, it operates so as to output TQ such that q≦■. Next, q-
When the value becomes 1+1, the subtracter 12 calculates the difference between the square value of the output signal X1 output from the output signal register 5 and the square value of the output signal X I+ + input to the output signal register 5. Therefore, the cumulative addition circuit 13 has T1 and += (L, +"-X+")+T+-Xx"+X
s”+・・・+X+”10X1. Outputs +. Similarly, when q=I+2, the cumulative addition circuit 13 calculates T + −z −(X 1.z ”X z ”)
+ T + -+-X3"+X%+ ...+X1+
1" +X++z. Following similar processing, in q)!, the cumulative addition circuit 13 outputs To-(XQ"XQ-+")+Tq-1-X□1°,!+X□5.! + ... + As it continues to operate, it operates to output the norm value ΣXj''(i).

The convergence speed of the filter coefficient Hj-+ (■) of the adaptive filter 4 in the "learning identification method" is assumed to change depending on the coefficient correction constant K. In other words, the larger K is, the faster the convergence will be, but the estimation accuracy after convergence will be rougher.
Conversely, it is said that if K is set small, the estimation accuracy after convergence increases, although convergence will be delayed. On the other hand, it is known that the condition 0<K<2 is given as the value of the coefficient correction constant in order to provide a stable condition for estimation.

However, in reality, if the value of K exceeds 1, convergence will be significantly slowed down, so it is usually set in the range OAK≦1.

Conventionally, by utilizing the properties of the "learning identification method", the residual echo E is monitored and compared with a predetermined darkness value to determine whether the filter coefficients are in the process of convergence or in a stable state of convergence. When determining that it is in the convergence process, priority is given to the convergence speed, and the filter coefficient H7+ is
In order to speed up the convergence of I(m), the value of K is set to a value close to 1, and when it is determined that the convergence is in a stable state, priority is given to estimation accuracy, and the filter coefficient Hj-+(m) is set.
In order to improve the accuracy of convergence, the value of K was set to a value close to O.

[Problem to be solved by the invention]

However, when the filter coefficients of the adaptive filter 4 are learned according to such a "learning identification method", the learned filter coefficients vary greatly at the impulse input and at the beginning and end of words, and this causes the filter coefficients to be learned up to that point to change. There was a problem that the estimated operation would be invalidated all at once.For example, in speech, 200a+ before a voiceless consonant.
There are many sections in which there is no voice for about s, and the estimation of the filter coefficients is greatly disturbed during these sections, which becomes an extremely serious problem. Furthermore, since the voice is alternating current even if it is not at the beginning or end of a word, and its dynamic range is wide, there is a problem in that the filter coefficients vary greatly accordingly. This problem becomes more serious as the number of taps of the adaptive filter 4 decreases.
Correspondingly, there was a problem in that the value of the filter coefficient became unstable.

That is, the conventional norm-based normalization process has a problem in that it is not possible to suppress the influence of amplitude fluctuations of the output signal Xj (Ill).

Then, by changing the value of the coefficient correction constant depending on whether the filter coefficients are in the convergence process or in a stable state of convergence, the filter coefficients converge while maintaining harmony between the convergence speed and estimation accuracy of the filter coefficients. Even if techniques were used to speed up the processing, there was a problem in that there were limits to its application because the filter coefficients varied greatly depending on impulse input, word beginnings, and endings.

Next, we will discuss the reasons why such problems occur according to the analytical model of the "learning identification method" which was the starting point for devising the present invention.

As mentioned above, the residual echo E obtained by the difference circuit 9 is E, -Y, -ΣHj(i)Xj(i) (2
) expression. On the other hand, the input signal Y, is the received signal S
4. If the additional noise is Nj and the echo path gain is J(i), then Y, = (Sj+N,)+Σh j(i) X J(i)
This will be expressed as equation (3). Here, the filter coefficient updating unit 8 operates to estimate the echo path gain hj(i) by executing a learning process so that the filter coefficient Hi(i) becomes the echo path gain hr(i).

Substituting these equations (2) and (3) into equation (1) above,
To summarize, Hj,, (-)=HJ (so-called) ΣX,"(1) = Hj(m) ΣXI(t) ΣXj"(i) ΣXJ"(i) "=HJ(m) ΣXj"(i) However, Δj (+) = tz (i) - Hj (i) Equation (4) Furthermore, from the cumulative addition of Equation (4), take out the term im and separate it, HJ, + (m )−)1i(s) ΣXj”(i) ΣXj”(+) ΣXJ”(i) However, Σ, ; Accumulation (5) formula % formula excluding imm % ) +ΣlIh J(i) X 7(i) −Σ1lH1(i
)XJ(i)=(S,+NJ)+Σ,ΔJ(t)X
j(+), equation (5) becomes Ht, +(m)=Hj(m) + K E m 7 X; (m)/Σ Xi"(i
)+Khj(m)Xi”(+i)/ΣX=”(+)KH
jC-)X j"(m>/Σ X;"<4>-HJ(
11) [I K X , H”(m)/ΣX J”
(i)]+ K h J (w)X ; ” (e+
)/Σ XJ”(i)+KEm=X j(m)/Σ
j”(i)=HJ(m)[I KXj”(m)
]/Σ Xj”(i)] + K [h j(n) +
E mj/X j(m)] X J”(m)/ΣXj
! (i)=Hj(+s)[I KXJ'(m)/ΣX
J! (i)] 10K(hj(m)+Qj(*)]Xj”(
s)/ΣXJz (+) However, Q j (work) - Emj/X
J(11) “/” represents division.

(6) Formula It is organized as follows.

FIG. 30 shows a diagrammatic representation of this equation (6).

Here, α= [1-K Xj”(m)/ΣXJ”(i)]
(7) Formula % Formula %) This figure 30 shows that the "learning identification method" calculates the sum of the input value (multiplied by the gain correction coefficient) and the value corresponding to the past input value, as shown in figure 31. a delay device 17 that delays the calculated value of the adder 16 by one sampling period, and a low-pass filter coefficient α at the output of the delay device 17.
A multiplier 18 that calculates a value according to the past input value by multiplying it and outputs it to the adder 16, and a multiplier 19 that multiplies the input value by a gain correction coefficient expressed as (1-α). This means that it constitutes a low-pass filter.

That is, from FIG. 30, the "learning identification method" suppresses the disturbance Q j (s) included in the input [hJ (■) + QJ] with a low-pass filter with a low-pass filter coefficient α, and Extract the component echo path gain hj(m),
It is clear that the function of converging the filter coefficient Hj(m) of the adaptive filter 4 to this echo path gain hj(m) is realized.

Thus, according to the analytical model of the "learning identification method" analyzed by the present inventor, it has become clear that the "learning identification method" exactly constitutes a low-pass filter.

By the way, in order to show in principle that the "learning identification method" can be constructed with a low-pass filter, FIG.
(to)) uses an expression that seems to cause some problems in practical use, such as dividing by X; (m). Therefore, this expression causes '
In order to avoid the situation of division by "Xj(m)=0", the following modifications are made. In other words, the second equation (6)
According to the term, the low-pass filter input is [hj (monk) + QJ]
From [hj(s+)+Q>1Xz(s) =h J (m) X J (-) + E m j-
h j (s)
i) Xj (i) − Σ, Hj (i)
N,) +Σh j(i) X J (i)−Σ, Ht(i)X
Transform to J(i)=-YJ-Σ1IHt(i)Xj(f). Y, -Σ, HJ (i) If the low-pass filter expression is rewritten, FIG. 32 is obtained. In other words, in the analytical model shown in Fig. 32, the third
KX, which is the gain correction coefficient of the low-pass filter shown in Fig.
Take out X t (m) from (s+)/ΣXj”(i) and get ■ h j(m)−4h j(m) X j(m)■
KXj”(m)/ΣxJ”(+)==>K Xj(+m
)/ΣX;” (+), the 30th
This is equivalent to the analytical model shown in the figure.

The stability condition for the low-pass filter is that the value of the low-pass filter coefficient α is in the range of -1〈α〈1.In other words, in order for the low-pass filter to be stable, from equation (7), 1 < I KXJ “(II)/ΣXj”(i)<1
(8) The relational expression 2 must hold true. By transforming this equation (8), we obtain 2 > K XJ''(11)/ΣXJ'' (i) > 0. In this equation, normally XJ"(1) <ΣXj"(+) holds, and it seems that a stable condition is ensured even with a fairly large K. However, for example, when an impulsive signal input or audio is used as an output signal, At the beginning and end of a word, if i≠m, then X j”(i) !=i 0 .

Since it is possible that Xj''(m)/ΣXj''(i)-1, the value of K needs to be 2>K>0. This stability condition formula is the same as the stability condition of the well-known "Learning Identification Method", and supports the validity of the analytical model of the low-pass filter expression of "Learning Identification Method J" analyzed by the present inventor. Of course, X, axis) = = (12) When the low-pass filter coefficient α
is l, and even if the condition of equation (8) is no longer satisfied, the filter input will also be 0, so the fact that the stability of the filter is maintained is no different from learning identification method J.

From the analytical model of the "learning identification method" shown in FIG. 30, problems with the "learning identification method" will be found. In other words, the low-pass filter coefficient α, where α=[I KXJ”(II)/ΣX j”(i)], is the output signal XJz(11)
The problem is that it is a function of the amplitude of the output signal XJ'' (m) and varies depending on the amplitude of the output signal XJ''(m).

To look at this problem specifically, if K=-1, the low-pass filter coefficient α is replaced by α-[1-Xj''(parrot)/ΣXj''(+)]. In this equation, since Xj"(-)≦ΣXノ"(i) holds true, 0<XJ"(seagull)/ΣXj"(+)≦1 holds true. Therefore, the low-pass filter coefficient α will fluctuate between 0≦α×1 according to changes in the instantaneous amplitude of the output signal Xj(m).

The reason why this is a problem is that the closer the low-pass filter coefficient α is to 1, the longer the time constant of the low-pass filter becomes and the slower the convergence (but the disturbance included in the input is more effectively suppressed and the output has the characteristic of being stable; conversely, the closer the low-pass filter coefficient α is to 0, the shorter the time constant and faster convergence, but the disturbances contained in human power are not suppressed enough and the output becomes unstable. Since the low-pass filter coefficient α fluctuates in this way, it means that the disturbance suppression performance fluctuates. The purpose of the low-pass filter is to extract the DC component hj (m), and changing the time constant means changing the fluctuation range of the extracted DC value. , the estimation accuracy of the filter coefficient Hj(m) of the adaptive filter 4 becomes unstable, which results in the residual echo increasing or decreasing depending on the amplitude of the output signal xj(11) in the echo canceller.

Also, in the low-pass filter coefficient α α=[1-KX,”(■)/ΣXj!(i)], the larger the value of K, the closer the low-pass filter coefficient α is to O; The smaller the value, the closer the low-pass filter coefficient α is to 1. On the other hand, as mentioned above, in a low-pass filter, the closer the low-pass filter coefficient α is to 1, the longer the time constant becomes; It has the characteristic that the time constant becomes shorter as the time constant becomes closer.This is exactly what is said in the "Learning Identification Method", and when the value of K is increased, the filter coefficient HJ-+(m) of the adaptive filter 4 converges. This means that the speed becomes faster, and conversely, the smaller the value of K, the slower the speed becomes.

As explained at the beginning of the [Problem to be Solved by the Invention] column, ``When learning the filter coefficients of the adaptive filter 4 according to the learning identification method J, the filter coefficients fluctuate significantly at the impulse input and at the beginning and end of words. This may cause the previous filter coefficient estimation operations to become invalid all at once, or the filter coefficients may become too thick (or fluctuate) even at the beginning or end of a word. As explained above, is exactly ``
The inventor's considerations have revealed that the "Learning Identification Method" operates as a low-pass filter, and the low-pass filter coefficient α is subject to fluctuations depending on the amplitude of the output signal. 9 For example, if we specifically explain impulsive signal input and word beginnings and endings, in such cases, ΣX j”(i) “q X J” (11) will occur, so equation (7) The low-pass filter coefficient α of is O
As a result, the time constant of the low-pass filter becomes extremely short and becomes unstable due to the strong influence of external disturbances, resulting in the estimated value of the filter coefficient of the adaptive filter 4 to fluctuate greatly. .

FIG. 33 shows the results of a simulation conducted to see how the low-pass filter coefficient α changes for the audio signal "ku".

Here, the simulation is performed with the number of taps I being 80. The simulation results revealed that the low-pass filter coefficient α becomes small for the consonant k, and that the low-pass filter coefficient α takes unstable values even at the beginning and end of words.

Next, we will discuss the meaning of the norm-based normalization process according to the low-pass filter expression of the adaptive filter 4. Conventionally, the normalization process using the norm is performed in the update process of the filter coefficient Hj, I (11) of the adaptive filter 4 expressed by the equation (1), ΣXj''(i), and the coefficient correction value of the second term is It is said that this is performed in order to cancel the amplitude fluctuation given by the included output signal However,
The analytical model of the "learning identification method" shown in FIG. 30 makes it clear that the normalization process using this norm does not cancel out the amplitude fluctuation given by the output signal X = (m). In fact, as shown in FIG. 33, this is confirmed from the fact that the low-pass filter coefficient α varies greatly in accordance with the variation in the amplitude of the output signal χj(m).

This is despite the fact that norm normalization only provides normalization on the signal section given by the output signal in the section cut out at time j, that is, normalization between the low-pass filter outputs. 0 This is interpreted as an error caused by equating it with normalization on the time axis.What is originally required is normalization of the amplitude fluctuation of the output signal that occurs between different signal vectors, that is, on the time axis. Of course, if there is no disturbance (Qj (m) = O) and the convergence value h (+*) is given as the filter output after convergence, the output of the low-pass filter in Fig. 30 for a new DC input is , h(-)KXj"(m)/ΣXノ"(i)+h60
1 KXj''(m)/ΣX7''(t)]=h(Otori), and it appears as if the amplitude fluctuation of the output signal Xj(m) has been eliminated by normalization using the norm. However, this result is not the effect of norm normalization. In fact, it is clear that the same result can be obtained by h(m)KXj''(m)+h(m)[1-KXJ!(Sho)l-h(禦), which omits the division by norm.

That is, the division process by the norm simply divides the output signal Xj(s+
) only has the effect of making the amplitude fluctuation relatively small. On the other hand, normalization using the norm makes the correction value of the filter coefficient dimensionless, and seems to have the effect of preventing the estimated filter coefficient from becoming a function of the output signal level. However, the low-pass filter coefficient α of the low-pass filter in FIG.
Even if the filter gain is rewritten as 1/(KXj"(w+)], the gain is corrected to 1 by the gain correction coefficient KX,"(m). Therefore, its effect is independent of normalization by norm.

As described above, according to the study of the analytical model of the "learning identification method" by the present inventor, it has become clear that division by the norm involves relative suppression of time constant fluctuations due to amplitude fluctuations of the output signal. That is, this normalization by division simply increases the denominator so that amplitude fluctuations in the output signal do not greatly affect the filter coefficients of the adaptive filter 4. If this is the case, there is no need to limit the norm value calculation process to the number of taps of the adaptive filter 4 as in the past, and by setting a large norm value, the amplitude fluctuation of the output signal can be reduced. It will be more absorbable.

The present invention has been made in view of the above circumstances, and it is an object of the present invention to enable learning of filter coefficients of an adaptive filter to be performed properly and at high speed when learning filter coefficients of an adaptive filter according to a learning identification method. The purpose of this paper is to provide a new filter coefficient learning method for an adaptive filter.

[Means to solve the problem]

Before entering into the explanation of the principle structure of the present invention, the inventor
We will explain the considerations advanced from the analytical model of the "learning identification method" shown in the figure.

As described above, in the "learning identification method", the low-pass filter coefficient α becomes a function of the output signal Xj (+*). This is the echo path gain hJ(
m) is not input as is to the low-pass filter shown in Figure 32, which is the analytical model of the ``Learning Identification Method,'' but is multiplied by the alternating current component ``h = (m) The learning identification method provides normalization within the low-pass filter in order to offset this with power and extract hj(a+) (7).
The low-pass filter coefficient α of Eq. However, as mentioned above, in the conventional "learning identification method", normalization is not actually performed accurately, and the low-pass filter coefficient α becomes a function of the output signal X j (m). .

The normalization calculation in the "learning identification method" does not necessarily need to be performed inside the low-pass filter; it is also possible and equivalent to perform it at the entrance or exit of the low-pass filter. If the normalization operation is performed outside the low-pass filter, it is no longer necessary to make the low-pass filter coefficient α a function of the output signal Xj.

This is because if normalization is performed outside the low-pass filter, the signal input to the low-pass filter can be controlled to be a DC component. In other words, if a configuration is adopted in which the input of the low-pass filter is multiplied by l/Xj(s+), it becomes possible to immediately exclude the output signal component from the low-pass filter coefficient α, realizing accurate normalization processing and reducing the The filter coefficient α can be set to a constant α゛.

Based on this consideration, the present inventor developed a calculation for learning filter coefficients of a new adaptive filter shown in FIG. 6 from the calculation system shown in FIG. 32 (which is equivalent to the "learning identification method"). devised a system.

However, in the calculation system shown in Fig. 6, Xj(■)−
The problem of division by zero arises. One method to solve this problem is to set the value of the low-pass filter coefficient α to 1n to stop the estimation calculation process of the low-pass filter when XJ(IR)=0. It is conceivable to take

Furthermore, as another solution, it is possible to adopt a configuration in which the normalization calculation is performed using the average power in a certain section. That is, by transforming the multiplication process of 1 / X j (m) in the arithmetic system of FIG. 6 into the multiplication process of X ; =
[h J(II)X j(intestine)+(S,+N,)+Σ
,hi(i)Xj(i)−Σ,Hj(i)Xj(+
)] X j (m) − [hj (■) Xj (@) + (Sj
+Nj)+Σ, Δj(i) X t (i)l X j
(+++)=h J(+w)X J”(II)+X
j(a) [(Sj+N])+Σ, Δj(i) X j(
+)] However, Δj(i) = h j(i) Hj(
i) Σ+s: Obtain the accumulation excluding i=m. This is cumulatively added for -K to interval j- at a certain time, Σ''[J(+m)Xj''(m)+X7(s+) ((S
7+Nj)+Σ, Δj(i) X j(i) l ];
Σ"hr r(m) X ="(m) ++Σ'Xr(m
) [(Sj+Nj)+Σ,ΔJ (i) X J(i)
] However, regarding the cumulative addition of Σ″:j= to −, the square value of the output signal Xj(m)
), the result of the normalization operation is h (s+), assuming that the echo path gain hJ (■) is a constant h (m). + Σ′χj(m) [(S J+ NJ)+Σ, Δj(i
) If a configuration is adopted in which normalization using this average power is performed on the input side of the low-pass filter of the calculation system shown in Figure 6, the second term in the above equation will be reduced as a disturbance by the low-pass filter, and the low-pass filter will If a configuration is adopted in which this is executed on the output side, the second term numerator in the above equation is treated as a disturbance and is reduced by a low-pass filter.

In this way, according to the method of stopping calculations and the method of normal processing using average power,
By solving the problem of division by j(m)=O,
The arithmetic system shown in FIG. 6 can now be realized.

Next, the principle structure of the present invention will be explained according to FIGS. 1 to 5. Here, FIG. 1 and FIG.
This shows the principle structure of an invention (hereinafter referred to as the first invention) in which filter coefficient learning is performed by solving the problem of division by XJ (m) = 0 according to the method of average power according to the calculation system shown in the figure. , and Fig. 3 shows the 6th
According to the calculation system shown in the figure, the invention (hereinafter referred to as the second invention) attempts to solve the problem of division by
Fig. 4 shows the principle configuration of the calculation system of Fig. 32 (i.e., the conventional "learning identification method").
FIG.
This figure shows the principle structure of an invention (hereinafter referred to as the fourth invention) which aims to improve the calculation system of FIG. 8 (that is, the calculation system of the conventional "learning identification method").

In FIG. 1, the same parts as shown in FIG. 26 are indicated by the same symbols. Here, the first invention shown in FIG. 1 is an invention in which the normalization process using the average power is implemented on the output side of the low-pass filter of the calculation system shown in FIG. Reference numeral 20 denotes a sum-of-products calculating means, which calculates the output value ΣH7(+)
) and the output signal XJ(m) of the output signal register 5 associated with this filter coefficient Hj(m) as inputs, filter coefficients of taps other than the specified tap,
Something that calculates Σ, Ht(i)Xj(+), which is the sum of the multiplication values with the output signals associated with these taps,
Reference numeral 21 denotes a correction means, which sets the value of the filter coefficient of the tap used by the product-sum calculation means 2 (12) to zero or a small value, or decreases the number of filter coefficients for a predetermined period of time to start learning the filter coefficients. 22 is a fourth calculation means that corrects the calculated value of the product sum calculation means 2 (12) by setting it to a smaller value. (12)
is the difference value from the calculated value]Yj-Σ, Hj(i)Xj(
i)], 23 is a first calculation means that calculates the multiplication value of the output signal X j (m) and the value calculated by the fourth calculation means 22, and 24 is a low-pass filter. The first calculation means 2 is an integration means constituted by the configuration, etc.
The calculated value of 3 is a constant integral coefficient α' (corresponding to the constantized low-pass filter coefficient α of the calculation system in Figure 6)
25 is a power calculation means that emphasizes the component value of the filter coefficient HJ 41 (m) included in the calculated value of the first calculation means 23 by performing integration processing according to, for example, the output signal X t ( The power value Pj(m) of the output signal Xj(m) is What to calculate, 26
is a second calculating means, which cancels out the power value P ) (m) of the output signal Xj (■) included in the output value of the integrating means 24 using the calculated value of the power calculating means 25. In the first invention with this principle configuration, the second calculation means 2
6, the updated value HJ4+(nl) of the filter coefficient is output.

In FIG. 2, the same components as those explained in FIG. 1 are indicated by the same symbols. Here, the first invention shown in FIG. 2 is an invention in which the normalization process using the average power is implemented on the input side of the low-pass filter of the calculation system shown in FIG. 27 is a cumulative addition means that cumulatively adds the value calculated by the first calculation means 23; 253 is a cumulative addition means included in the power calculation means 25 that cumulatively adds the value calculated by the square value calculation means 251; 28 is a third calculating means which calculates the power value of the output signal XJ(m) included in the calculated value of the cumulative addition means 27. 25 is an integrating means constituted by a low-pass filter configuration, etc., which integrates the calculated value of the 2nd day according to a constant integral coefficient α''. , the component value of the filter coefficient Hi(m) included in the calculated value of the third calculating means 28 is emphasized.In the first invention with this principle configuration, the updated value H4 of the filter coefficient is input from the integrating means 29.゜1
(m) will be output.

In the first invention shown in FIGS. 1 and 2, it is not necessarily necessary to provide the fourth calculation means 22, and the input signal Y is sent to the first calculation means 23. It is also possible to adopt a configuration in which direct input is performed. Then, the integrating means 24 and 29 process the value of the integral coefficient α° to increase from a value near zero to a steady value near 1 in accordance with the progress of filter coefficient learning. Also, the integrating means 24
゜29 emphasizes the filter coefficient component of the tap specified according to the arithmetic mean method when the product-sum calculating means 20 uses a filter coefficient with a zero value, and emphasizes the filter coefficient component of the tap specified by the arithmetic mean method, and when it uses a filter coefficient with a non-zero value, it emphasizes the filter coefficient component of the tap. Processing is performed to emphasize the filter coefficient components of the taps identified according to the filter.

Also, as the value of the integral coefficient α゛ of the integrating means 24.29,
It is configured to be set to a value greater than or equal to a value near the value determined by the number of taps corresponding to the reverberation time of the system.

Then, at the start of the filter coefficient learning process, for example, the output signal register 5 is cleared, so that the output signal X j (m) associated with the tap of the adaptive filter does not exist. The filter coefficients of the filter coefficient register 6 corresponding to the shortest delay time of the system are set to zero values.

In FIG. 3, the same components as those shown in FIG. 1 are indicated by the same symbols. 30 is a reciprocal calculation means,
1/X j(s
+), 31 is a first calculation means,
Reciprocal calculation means 3 (12) calculates the multiplication value of the calculated value and the calculated value of the third calculation means 22, 32 is an integration means constituted by a low-pass filter configuration, etc., and the first calculation means The component value of the filter coefficient Hj(ll) included in the calculated value of the first calculation means 31 is emphasized by integrating the calculated value of 31 according to the constant integral coefficient α';
Reference numeral 3 denotes a signal value judgment means, which judges whether the output signal
When it is determined that J(11) is at a predetermined minute value, the calculation processing of the first calculation means 31 or the integration means 32 is stopped. In the second invention having this principle configuration, the update value Hj-+(
■) will be output.

In this second invention, it is not necessarily necessary to provide the fourth calculation means 22, and it is also possible to adopt a configuration in which the input signal Yj is directly input to the first calculation means 31. be. Then, the integrating means 32 has an integral coefficient α
The value of ° is processed so as to increase from a value near zero to a steady value near 1 in accordance with the progress of filter coefficient learning. Further, when the sum-of-products calculating means 20 uses a filter coefficient with a zero value, the integrating means 32 emphasizes the filter coefficient component of the tap specified according to the arithmetic mean method, and uses a filter coefficient that is not a zero value. Sometimes, processing is performed to emphasize the filter coefficient components of the identified taps according to the low-pass filter. Also, the integrating means 32
The value of the integral coefficient α' is set to a value greater than or equal to a value near the value determined by the number of taps corresponding to the reverberation time of the system. Then, at the start of the filter coefficient learning process, for example, the output signal register 5 is cleared, so that the output signal Xj corresponding to the tap of the adaptive filter is
(■) is configured so that it does not exist.

The filter coefficients of the filter coefficient register 6 corresponding to the shortest delay time of the system are set to zero values.

In FIG. 4, the same parts as shown in FIG. 1 are indicated by the same symbols. 5a is filter coefficient register 6
5b is an output signal register 5 that stores the output signal Xj(i) associated with the filter coefficient of
an auxiliary output signal register 40 for storing the output signal Xj(i) overflowing from a;
41 for calculating ΣXj''(i) (=N), which is the sum of the squared values of the output signal Xz(+) stored in b.
is a power calculation means, and the square value X of the output signal Xj,
2, and an integrating means 412 that integrates the calculated value of the square value calculating means 411 to calculate the power value Pj of the output signal Xj. 42 is the power calculating means 41 is an arithmetic means that receives as input the power value P calculated by, for example, calculates the filter coefficient HJ(Ill) to be learned by executing a calculation of a delay operation.
43 is an integrating means, which calculates the power value P=(m) for the output signal X j (s+) corresponding to
! (However, f=P,,P.

(-), N) first multiplication means 4 that performs multiplication processing with
31, an addition means 432 whose one input is the value calculated by the first multiplication means 431, a delay means 433 which delays the value calculated by the addition means 432 by one sampling period, and a delay means 43.
with a delay value of 3 and [1-KX.

(m)/j2] (however, 1 = Pt, Pr(m),
a second multiplication means 434 that uses the multiplication value calculated by performing the multiplication process with N) as the other input of the addition means 432;
By providing an updated value HJ of the filter coefficient.

(m).

In this third invention, when the sum-of-products calculating means 20 uses filter coefficients with zero values, the integrating means 43 emphasizes the filter coefficient components of the taps specified according to the arithmetic mean method, and When a filter coefficient that is not a value is used, processing is performed to emphasize the filter coefficient component of the specified tap according to the low-pass filter. Further, as the value of the coefficient correction constant K of the integrating means 43 (this corresponds to the integral coefficient of the integrating means 24 etc. of the first invention),
It is configured to be set to a value less than or equal to a value near the value determined by the number of taps corresponding to the reverberation time of the system.

Then, at the start of the filter coefficient learning process, for example, the output signal register 5 is cleared, so that the output signal X (2) associated with the tap of the adaptive filter does not exist. The filter coefficients of the filter coefficient register 6 corresponding to the shortest delay time of the system are set to zero values.

In FIG. 5, the same parts as those explained in FIGS. 4 and 28 are indicated by the same symbols.

Reference numeral 21a denotes a correction means, which sets the value of the filter coefficient used by the convolution unit 7 to zero or a small value, or sets the number of filter coefficients to a small number, for a predetermined period of time when learning the filter coefficients starts. By doing so, the pseudo image echo ΣHj calculated by the convolution unit 7
(i)

In this fourth invention, the division circuit 82 calculates either ΣXj''(i) (=N) calculated by the norm calculation means 40 or the power value P calculated by the power calculation means 41.

or, by performing the multiplication process according to either one of the power values P7(m) calculated by the calculation means 42, according to the following formula, where l −P j + P j (s), N “Learning identification method” Process to execute. The coefficient correction constant used by the multiplication circuit 80 is configured to be set to a value equal to or smaller than a value determined by the number of taps corresponding to the reverberation time of the system. Then, at the start of the filter coefficient learning process, for example, the output signal register 5 is cleared, so that the output signal Xj(m) associated with the tap of the adaptive filter does not exist. The filter coefficients of the filter coefficient register 6 corresponding to the shortest delay time of the system are set to zero values.

[Effect]

In the first aspect of the present invention shown in FIGS. 1 and 2, X r (m)
The problem of division by = O is solved according to the method of average power, and the filter coefficients of the adaptive filter are learned.

That is, in the first invention shown in FIG. 1, the integrating means 24
Processing is performed to obtain the power component while emphasizing the filter coefficient Hj-+(■) included in the signal inputted via the first calculation means 23, and is arranged on the output side of the integration means 24. The second calculation means 26 uses the power value calculated by the power calculation means 25 to
By removing the power component included in the output value of 4 and extracting the filter coefficient H,,1(m), the normalization process explained in equation (9) is realized, and the 6th The processing of ``1/XJ(II)J'' of the calculation system shown in the figure is realized. By realizing the processing of '1/Xj(m)J, the integrating means 24 can execute the integration process according to the constant integral coefficient α°. Therefore, it is possible to prevent the inconvenience that the estimation accuracy of the updated filter coefficient value Hj-+(m) to be extracted changes depending on the value of the output signal Xj(...).

Further, in the first invention shown in FIG. 2, the third calculating means 28 disposed on the input side of the integrating means 29 has
By executing the division process between the calculated value and the calculated value of the power calculating means 25, the power component included in the output value of the first calculating means 23 is removed, and the integrating means 29 The filter coefficient HJ−+(+*) included in the output value of the calculation means 28 is emphasized to obtain the filter coefficient HJ.
-1(M) extraction, (9
) The normalization processing explained using the formula is realized to realize the processing of '1/XJ(Ill)J in the arithmetic system of FIG. this'
1 /
This makes it possible to prevent the inconvenience that the estimation accuracy of j, I (■) changes depending on the value of the output signal Xj (11).

As disclosed in the explanation up to the derivation of the arithmetic system in FIG. The constant becomes longer, the estimation accuracy of the filter coefficient improves, and “O
If the time constant is set to a value close to ", the time constant becomes small and the estimation accuracy of the filter coefficient deteriorates.

From now on, as the learning of the filter coefficients progresses, we will increase the value of the integral coefficient α from a value near zero to a steady value near 1, so that convergence will be fast at the beginning of learning. It is possible to quickly converge the filter coefficients to a certain point by prioritizing estimation accuracy, and when learning has progressed to a certain extent, prioritize estimation accuracy and perform filter coefficient learning. and can be executed with high precision.

1Σ1I that the fourth calculation means 22 adds to the input signal Y
Hj (i) "-Σ, Hj(i)Xj(+)" is applied to the disturbance "Σ" with this disturbance.

This is to cancel out "hj (i) On the other hand, this disturbance −Σ, Hw(i)X
If j(i)'' is large, the error in estimation accuracy will conversely increase, resulting in a delay in convergence of the filter coefficients.

In order for the applied disturbance to speed up the convergence of the filter coefficients, Σ,hi(i)XJ(i) l > I Σ,Δj(i
)Xj(i)001. because,
As mentioned above, the output of the first calculation means 23 is (YJ-Σ, Hj(i)Xj(+)) Xj(n+)=
h ノ(++)X J”(m)+Xj(m)[(S
, +Nj)+Σ, ΔJ(i)
)Xj(i)'' will conversely increase the estimation error.

From now on, in the case where this fourth calculation means 22 is not omitted, the correction means 21 calculates the disturbance "-Σ, H, +
By processing so that the value of (i)

As for the structure of the integrating means 24 and 29, it is possible to adopt not only an integration process using a simple low-pass filter but also an integration process using cumulative addition. An arithmetic mean method is one of the integration processes based on this cumulative addition.

When a low-pass filter configuration is adopted, the dispersion at the output of the low-pass filter is σ!, if the dispersion at the input is expressed by σ2. (1-α
). However, α is a low-pass filter coefficient. on the other hand,
When the arithmetic mean method is configured, the variance in the output of the arithmetic mean method is σ”/q.

However, q is the number of samples. Therefore, in order to obtain the same variance between the output of the low-pass filter configuration and the output of the arithmetic averaging method, (1-α)=1/q 0For example, looking specifically at the example of q=256, , to obtain the same variance, α=
It is necessary to set it as 255/256. On the other hand, when integrating according to the low-pass filter coefficient α in a low-pass filter configuration, the value when up to q integration processes are executed is approximately (l - αQ). Considering the convergence time as a time constant that is 90% of the convergence value, 1-α''=0.
It becomes 9. In other words, α”=0.1. From this, the convergence time q in the low-pass filter configuration is α=255/25.
When set to 6, q=588. On the other hand, in the arithmetic mean method, q=256.

As can be seen from this specific example, integration processing using the arithmetic mean method can be executed faster than integration processing using a low-pass filter configuration. However, if the low-pass filter configuration is followed, the error that occurs at the beginning of the filter coefficient estimation decreases according to αq, whereas if the arithmetic mean method is followed, the error occurs according to 1/q. That is, in the integration process using the arithmetic mean method, there is a property that the influence of an error that occurs at the beginning of the filter coefficient learning process remains forever.

From now on, the integration means 24 and 29 are configured to implement two integration processes, for example, an integration process using the arithmetic mean method and an integration process using a low-pass filter configuration, and are provided with a switching means for switching between these processes to calculate the sum of products. Means 20
When is using a filter coefficient with a zero value, the filter coefficient component is emphasized according to the integration process by the arithmetic mean method selected by the switching means, and the sum of products calculation means 20 is using a filter coefficient with a non-zero value. When the low-pass filter configuration is selected by the selection means, the filter coefficient components are emphasized according to the integration process using the low-pass filter configuration, thereby making it possible to learn the filter coefficients at high speed and with high accuracy. .

Then, at the start of the filter coefficient learning process, for example, the output signal register 5 is cleared, so that the output signal XJ (Ill) corresponding to the tap of the adaptive filter is
If the filter coefficients of the filter coefficient register 6 corresponding to the shortest delay time of the system are configured to be set to zero, the disturbance “Σ , H>(t)Xj(i
)" does not contain any disturbance terms that are meaningless for learning the filter coefficients, so learning the filter coefficients can be performed at high speed and with high accuracy.

The setting of the value of the integral coefficient α' has a great influence on the convergence speed and estimation accuracy of the filter coefficients. According to the simulation results of the examples described later, it is clear that the estimation accuracy of the filter coefficients is almost saturated when the value of the integral coefficient α° takes the value of 00 taps given by the reverberation time. Ta. From now on, by setting the value of the integral coefficient α to a value that is greater than or equal to the value determined by this formula (the value determined by the number of taps corresponding to the reverberation time of the system), the learning of the filter coefficients will be made more accurate. become executable.

In the second aspect of the present invention, the filter coefficients of the adaptive filter are calculated by solving the problem of division by X7(s)=0 according to the calculation system shown in FIG. Perform learning. That is, in this second invention, the signal value determining means 33 determines that X t (m
) = When it is determined that the problem of division by O occurs, the calculation stop means 34 stops the calculation processing of the first calculation means 31 or the integration means 32, so that the sixth
The '1/XJ(II)J processing of the calculation system shown in the figure is realized.

By realizing the process of rI/X j (m) J, the integrating means 32 can perform the integral process according to the constant integral coefficient α°, so that the updated value Hj−+ of the extracted filter coefficient
This makes it possible to prevent the inconvenience that the estimation accuracy of (m) changes depending on the value of the output signal X J (+m).

In the integration process executed by the integration means 32, the integration means 24. ．． Similarly to 29, by increasing the value of the integral coefficient α from a value near zero to a steady value near 1 as learning of the filter coefficient progresses, at the beginning of learning, It is now possible to quickly converge the filter coefficients to a certain point by prioritizing faster convergence, and when learning has progressed to a certain extent, to prioritize estimation accuracy and carry out learning of the filter coefficients. Learning can be performed quickly and with high accuracy.

Similarly to the first invention, if speeding up is not required, the fourth calculation means 22 can be omitted, and if the fourth calculation means 22 is not omitted, it is possible to omit the fourth calculation means 22. However, according to the processing of the correction means 21, learning of filter coefficients can be executed at high speed.

Similarly to the first invention, the integration means 32 switches between the integration processing using the arithmetic mean method and the integration processing using the low-pass filter configuration as learning progresses, thereby learning the filter coefficients at high speed and with high efficiency. Be able to execute with precision.

Similarly to the first invention, there is a disturbance term that is meaningless with respect to filter coefficient learning in the disturbance °゛-Σ,Hr(t)Xj(i)'' calculated by the product-sum calculating means 20. Therefore, learning of filter coefficients can be performed quickly and with high accuracy.

Similarly to the first invention, by setting the value of the integral coefficient α to a value greater than or equal to the value determined by the number of taps corresponding to the reverberation time of the system, the filter coefficient σ can be learned with high accuracy. You will be able to run it with

In the third aspect of the present invention, according to the calculation system of FIG. 32 derived by the present inventor (that is, the calculation system in which the conventional "learning identification method" is expressed by a low-pass filter), and according to the processing of the correction means 21, Similar to the first and second inventions, learning of filter coefficients can be executed at high speed.

Then, similarly to the first and second inventions, the integration means 43 switches between the integration processing using the arithmetic mean method and the integration processing using the low-pass filter configuration as the learning progresses, thereby learning the filter coefficients. It can be executed at high speed and with high accuracy.

Similarly to the first and second inventions, the product sum calculation means 2
Since the disturbance term "Σ, Hj (i) It becomes like this.

According to the simulation results of the embodiment described later, it is clear that the coefficient correction constant K of the integrating means 43
It was found that the estimation accuracy of the filter coefficients is almost saturated when the value is determined by the number of taps corresponding to the reverberation time of the system. From now on, by setting the value of the coefficient correction constant to a value that is less than or equal to the value determined by the number of taps corresponding to the reverberation time of this system, learning of the filter coefficients can be executed with high precision.

In the fourth aspect of the present invention, N calculated by the norm calculation means 40, Pj calculated by the power calculation means 41, and the calculation are performed according to the calculation system shown in FIG. By using PJ(m) calculated by the means 42 as the norm value, the output signal Xt(a) is
It becomes possible to perform learning of filter coefficients without being affected much by fluctuations in .

Pseudo echo ΣHj that the difference circuit 9 adds to the input signal Y
(i) X j (i)" is the "-Σ
, HJ(i) The reason for applying the disturbance −ΣHJ(i)Xj(+)′, which should be
This is to cancel out "Xt(i)" and speed up the convergence of the filter coefficients.

From this, the disturbance "-Σa Hj (+) X ; (
In the same way that a large value of ``-ΣHj(t)'' added by the difference circuit 9 increases the error in estimation accuracy,
If Xi(i)'' is large, conversely, the error in the estimation accuracy increases, and the convergence of the filter coefficients is delayed.From now on, the correction means 21a will calculate the is processed so that the value of the pseudo echo "ΣHj(i) become.

And the first. As in the second and third inventions, at the start of the filter coefficient learning process, for example, the output signal register 5
is cleared, so that the output signal X is associated with the tap of the adaptive filter.

(-) does not exist, or if the filter coefficients of the filter coefficient register 6 corresponding to the shortest delay time of the system are configured to be set to zero, the convolution unit 7 calculates Pseudo echo “ΣH=(i)
Since a disturbance term that is meaningless with respect to learning of filter coefficients is not included in Xj(i), learning of filter coefficients can be executed at high speed and with high accuracy.

Similarly to the third invention, learning of the filter coefficients is enhanced by setting the value of the coefficient correction constant to a value less than or equal to the value determined by the number of taps corresponding to the reverberation time of this system. Be able to execute with precision.

As described above, according to the present invention, when learning the filter coefficients of an adaptive filter according to the learning identification method, it becomes possible to perform learning of the filter coefficients properly and quickly.

(Example〕 Hereinafter, the present invention will be explained in detail according to examples.

FIG. 7 shows the integration means 24°29. of the first and second inventions.
32 illustrates one embodiment of the invention. In this embodiment, a low-pass filter circuit 50 having the same configuration as that described in FIG.
Accordingly, it is disclosed that the integrating means 24, 29 and 32 are configured. As mentioned above, the first and second inventions include:
Since the learning of the filter coefficients of the adaptive filter is performed according to the arithmetic system of FIG. 6 derived by the present inventor, the low-pass filter circuit 500 multiplier 18 has a constant low-pass filter coefficient α The multiplication process is executed using , and the multiplier 19 of the low-pass filter circuit 500 executes the multiplication process using (1-α°), which is a constant.

The embodiment shown in FIG. 7(a) is an embodiment in which the integrating means 24, 29, 32 are configured according to the low-pass filter circuit 50 itself.

The embodiment shown in FIG. 7(b) has a low-pass filter circuit 5 (1
2) A low-pass filter coefficient setting circuit 51 that processes the value of the low-pass filter coefficient α' to increase from a value near "0" to a steady value near "1" as learning progresses. 24.2932 is an embodiment of an integrating means 24.2932 comprising: This low-pass filter coefficient setting circuit 51 may be configured, for example, based on the elapsed time from the start of learning, or based on the residual echo value E.
The value of the low-pass filter coefficient α° is set based on the magnitude of j. By providing this low-pass filter coefficient setting circuit 51, priority is given to speeding up the convergence of filter coefficients at the beginning of learning, and when learning has progressed to a certain extent, priority is given to estimation accuracy of filter coefficients. This makes it possible to learn filter coefficients at high speed and with high accuracy.

The embodiment shown in FIG. 7(c) includes, in addition to the low-pass filter circuit 50, an arithmetic mean circuit 52 that performs integration processing by the arithmetic mean method, and an arithmetic mean circuit 52 that performs an arithmetic mean integration process according to the control signal from the correction means 21. Integrating means 24, 29.3 comprising a switching circuit 53 for selecting one of the averaging circuits 52;
This is Example 2. In this embodiment, when the correction means 21 controls the sum-of-products calculation means 20 to use a filter coefficient of zero value, the correction means 21 executes the integration process according to the arithmetic averaging circuit 52, and the sum-of-products calculation means 20 When 20 is controlled to use a filter coefficient other than a zero value, a configuration is adopted in which integration processing is performed according to the low-pass filter circuit 50.

As specifically explained in the [Operation] section, although the integration process using the arithmetic mean method has the advantage of being faster than the integration process using the low-pass filter configuration, The nature of this is that the effects of errors will remain forever. From now on, in this embodiment, the arithmetic mean circuit 5
2 and selects the low-pass filter circuit 50 during learning which is controlled to use non-zero filter coefficients, thereby making it possible to perform filter coefficient learning at high speed and with high accuracy.

Next, it will be explained that the arithmetic averaging circuit 52 shown in FIG. 7(c) certainly implements the arithmetic averaging method. In this arithmetic averaging circuit 52, as shown in the figure, the multiplier 18 is (
The multiplication process is executed using n-1)/n, and the multiplier 19 processes to execute the multiplication process using n. here,
n represents the average number. The signal input to this arithmetic averaging circuit 52 is expressed as A(j), and the output for this A(j) is expressed as B(j), and the operation of this arithmetic averaging circuit 52 is observed over time. , B(1)=A(1) B(2)=A(1)/2+A(2)/2=[A(1)+
A(2)]/2 B(3)=B(2)X2/3+A(3)/3[A(1)
+A(2)]/2X2/3 +A(3)/3 −[A(1)+A(2)+A(3)l/3B(4) −
B(3)x3/4+A(4)/4=[A(1)+A(2
)+A(3)]/3X3/4+A(4)/4 = [A(1)+A(2)+A(3)+A(4)]/4, which certainly realizes the arithmetic mean method. This becomes clear.

The embodiment shown in FIG. 7(d) does not include the switching circuit 53 as shown in FIG. (Also included in the arithmetic mean circuit 52 in FIG. 7(C)), an integrating means is provided with an n value fixing circuit 53b that stops updating the n value and fixes it to the n value at that time in accordance with an instruction from the correction means 21. This is an example of 24.29.32. In this embodiment, when the correction means 21 controls the sum-of-products calculation means 20 to use a zero-value filter coefficient, the n-value update circuit 53a updates the n-value so that the calculated average When the correction means 21 executes the integration process according to the circuit 52 and controls the sum-of-products calculation means 20 to use a filter coefficient that is not a zero value, the n value fixing circuit 53
b operates to convert the calculation averaging circuit 52 to the low-pass filter circuit 50.
By processing to set it as , the same operation process as the embodiment of FIG. 7(C) is executed.

FIG. 8 illustrates an embodiment of the integrating means 43 of the third invention. The embodiment shown in FIG. 8(a) is the same as that shown in the principle configuration of FIG. 4.

The integration circuit 43a of this embodiment basically has the same configuration as the low-pass filter of the calculation system shown in FIG.
As explained in the figure, the first calculation means 431 calculates the multiplication coefficient KX, (rin)/l (where l-P j+ P > (s
)+N), and the second multiplication means 434 executes the multiplication process using the multiplication coefficient [I KXj(m)/11.

The embodiment shown in FIG. 8(b) has an arithmetic mean circuit 54 which performs integration processing by the arithmetic mean method in addition to the integration circuit 43a.
and the integrating circuit 43 according to the control signal from the correction means 21.
This is an embodiment of the integrating means 43 including a switching circuit 55 for selecting either one of the arithmetic averaging circuit 54 and the arithmetic averaging circuit 54. Here, the value of the coefficient correction constant is assumed to be 1''. In this embodiment, when the correction means 21 controls the sum-of-products calculation means 20 to use a filter coefficient of zero value, Integration processing is performed according to the averaging circuit 54, and the correction means 21
When the sum-of-products calculation means 20 is controlled to use a filter coefficient that is not a zero value, a configuration is adopted in which the integration process is executed according to the integration circuit 43a. That is, as explained in FIG. 7(c), at the beginning of learning when the product-sum calculation means 20 is controlled to use filter coefficients with zero values, the arithmetic mean circuit 54 is selected, and the filter coefficients with non-zero values are selected. By selecting the integrating circuit 43a during the progress of learning controlled to use coefficients, the configuration is such that learning of filter coefficients can be performed at high speed and with high accuracy.

Next, it will be explained that the arithmetic averaging circuit 54 of FIG. 8(b) does indeed implement the arithmetic averaging method. In this arithmetic averaging circuit 54, as shown in the figure, the first multiplication means 435 is X, (m) Σ The second calculation means 438 executes the process, and the second calculation means 438 calculates Σ
k"(m) Σ

For example, when n=1, [ht(m)Xt(m)+Z+
], the arithmetic averaging circuit 54 inputs Hz(m)−
[ht(-)Xt(m)+LIX+(m)/X+"(糟)-ht(a)+Z+X+(m)/X+"(m) a+)Xz(m)
When there are inputs of 10 Z, l, the arithmetic mean circuit 54 calculates H, (
■) - [hz(m)Xz(m)+Zz]X[Xt(■)
/(X +”axis)+L”(麟)))+[ht(m)+Z
+X+(s)/X+”(m)]X [X+”(m)/
(X +”(m)+X z”(m))]−[ht(
m) Xi(s) Xi(s+)+ZzXz(m)]X[1
/(Xt”(m)+Xz”(m))]+[ht(m)X
t”(m)+Z+X+(m)]x[1/(X+must)+
Xt"(鳳))=[hz(+g)Xz"(m)+h
t(m)Xt"(m)]X[1/(Xt"(m)+L"
axis)))+[Z3Xi(m)+Z+X+(m)]X[
1/(Xtnecessary−)+Xz”(m))], but here, if the echo path is stable and ht(s
) = hg(m) - h(m) holds, then it becomes clear that the arithmetic mean method is indeed realized, as [Xt"(m) + Xz"(m)] .

In the embodiment of the integrating means 43 shown in FIG. 8(b), as disclosed in FIG. By configuring to include an n value fixing circuit that stops updating the n value and fixes it to the n value at that point in accordance with the instruction, it is possible to adopt a configuration in which the integrating circuit 43a and the arithmetic averaging circuit 54 are used in common. It is.

FIG. 9 illustrates an embodiment of the power calculation means 25 of the first invention. The embodiment shown in FIG. 9(a) is a square calculation output circuit that receives an output signal X j (n) associated with tap m of an adaptive filter and calculates the square value of this output signal X j (m). 56 and a low-pass filter circuit 50 that calculates the power value PJ (II) of the output signal X7 (s+) by integrating the value calculated by the square calculation output circuit 56. be. Further, in the embodiment shown in FIG. 9(b), the output signal Xj (=Xj(1)) input to the output signal register 5 is input,
A square calculation output circuit 5 that calculates the square value of this output signal Xj
6, a low-pass filter circuit 50 that calculates the power value P of the output signal Xj by integrating the value calculated by the square calculation output circuit 56, and a power value P calculated by the low-pass filter circuit 5 (12). , by one sampling period to output the power value Pj(m) of the output signal XJ(ml). If the embodiment of FIG. 9(a) is followed, it is necessary to prepare a low-pass filter circuit 50 and a square clearing circuit 56 for the number of taps, whereas if the embodiment of FIG. 9(b) is followed, In this case, it is sufficient to prepare one set of the low-pass filter circuit 50 and the square calculation circuit 56.

In the embodiment of the integrating means 24 in FIG. 8, the low-pass filter circuit 5
0, the power calculation means 25 in FIGS. 9(a) and (b)
Since the low-pass filter circuit 5° is also used in the embodiment, as will be described later, these low-pass filter circuits 5 (
12) Multiplier 19 can be omitted.

The embodiment of the power calculating means 25 shown in FIG. 9(c) utilizes the output signal register 5 shown at 270 to
This is an embodiment in which the delay circuit 57 required in the embodiment of FIG. 3(b) can be omitted.

In this embodiment, a switch circuit 58 sequentially takes out the output signal X (x) from the output signal register 5, and a square calculation circuit that calculates the square value of the output signal X j (++) selected by the switch circuit 58. 56a, the power value PJ obtained by the low-pass filter circuit 50 (however, the multiplier 19 is omitted) and the square calculation circuit 56b, or the power value stored in the register circuit 59. a subtraction circuit 61 that calculates the difference between the power value output from the switching circuit 60 and the value calculated by the square calculation circuit 56a; and a subtraction circuit 61 that multiplies the calculated value of the subtraction circuit 61 by 1/α゛. The multiplication circuit 62 is also provided, and the register circuit 59 is configured to hold the value calculated by the multiplication circuit 62.

The power calculation means 2 of FIG. 9(c) configured in this way
In the embodiment No. 5, first, p,=x,"+α'xj-,"10α+ 2
+α”3XJ-ko2+α”XJ-4”+... is output, and in the subtraction circuit 61, PJ(-
Pj(1)) and the calculated value XJz(
The difference value from Xj=Xj(1)) is calculated, and in the multiplier circuit 62, (Pz
J-a"+ (X14Xj-5" + river-i.e. P
j−+ (=P j (2)) is calculated and held in the register circuit 59. Subsequently, this PJ(2) is outputted to the outside via the switching circuit 60, and at the same time, the subtraction circuit 61 calculates Pt+ (-PJ(2)) and the next calculated value XJ- of the square calculation circuit 56a. 1” (XJ-1-X
j(2)) is calculated, and in the multiplier circuit 62, (P j−+ Xj−+”) /α゛=XJ−1”+
(r'Xj-2"+Ct"Xj-a+α"Xj-5χ
+αI 4 X J-, 2+... In other words, P,
−□(−Pj(3)) is calculated and held in the register circuit 59. Thereafter, according to the same process, the power values P, (■) are outputted to the outside via the switching circuit 60.

In the embodiment of FIG. 9(c), the square calculation output circuit 5
6b is changed to XJ(1), which is the output of the final stage of the output signal register 5, the switching order of the switch circuit 58 is reversed, and the multiplication coefficient used by the multiplication circuit 62 is changed to α°. By changing the subtraction circuit 61 to an adder configuration, it becomes possible to output the power value Pj(m) in the same way.

As shown in FIG. 1, when the power value Pz(m) calculated by the power calculating means 25 is used for division processing with the value calculated by the integrating means 24, the power calculating means 25 is The low-pass filter coefficient α° of the low-pass filter circuit 5 (12) constituting the integrating means 24 and the low-pass filter coefficient α° of the low-pass filter circuit 5 (12) constituting the integrating means 24 need to be set to the same value. , in a transient state at the start of filter coefficient learning, it is not necessarily necessary to set them to match.

Although FIG. 9 has been disclosed as an embodiment of the power calculation means 25 of the first invention, these embodiments are also examples of the power calculation means 41 and the calculation means 42 of the third and fourth inventions. It is. That is, since the embodiment of FIG. 9(a) outputs the power value Pj(m), it can be applied as an embodiment of the power calculation means 41 and the calculation means 42 of the third and fourth inventions. Yes, since the square value calculation circuit 56 and the low-pass filter circuit 50 in the embodiment of FIG. As an example, since the delay circuit 57 inputs the power value P1 and outputs the power value Pj(m), the third and fourth
The square value calculation circuit 56b and the low-pass filter circuit 50 in the embodiment of FIG.
j, it can be applied as an embodiment of the power calculation means 41 of the third and fourth inventions, and the other switch circuits 58 etc. input the power value P and calculate the power value PJ(m). Since it outputs , it can be applied as an embodiment of the calculation means 42 of the third and fourth inventions.

FIG. 10 illustrates an embodiment of the product-sum calculation means 2 (12) of the first, second, and third inventions. The sum-of-products calculation means 20 of this embodiment calculates the multiplication value of the filter coefficient Hj(a) assigned to the specified tap and the output signal Xt(m) associated with the filter coefficient Hj(m). The multiplication circuit 63 and the calculated value HJ(II
) X J (m) and the pseudo echo ΣHj (i) Σ, Hj(i), which is the sum of the multiplication values of the filter coefficients and the output signals associated with these taps;
Processing will be performed to calculate XJ(i). Calculated by the sum-of-products calculation means 20 Σ, Hj(i)Xj(i)
is applied according to the format in which it is subtracted from the input signal Yj according to the fourth calculation means 22, but these Σ, H
= (i) Since XJ(+) is originally a disturbance, it is not essential to finding the updated value Hj,, (■) of the filter coefficient and can be omitted. It is. However, as mentioned above, there is a characteristic that learning of filter coefficients can be executed at a higher speed when this disturbance is applied.

It should be noted that the pseudo echo ΣH calculated by the convolution integrator 7
j (i) It is also possible to omit it. However, when this pseudo echo is applied, learning of filter coefficients can be performed faster.

FIG. 11 shows correction means 21 of the first, second and third inventions.
A filter coefficient correction circuit 65 which is an embodiment of the correction means 21a of the fourth invention is illustrated.

The filter coefficient correction circuit 65 shown in FIG. 11(a) corrects the values of the filter coefficients used by the convolution integrating section 7 and the product sum calculating means 20 to zero values for a predetermined period of time when learning of filter coefficients is started. This is an example of a filter coefficient correction circuit 65 that executes the process. This filter coefficient correction circuit 65 sets the magnitude of the clock count value or residual echo value Ej representing the passage of time from the start of learning to a predetermined reference value Pl.
By providing a comparison circuit 66 that compares the value with , and a selection circuit 67 that selects either the filter coefficient Hj(i) or the zero value according to the comparison result of the comparison circuit 66, the predetermined period for starting learning can be adjusted. operates to substantially set all or part of the filter coefficients Hj(i) of the filter coefficient register 6 to zero values.

According to this operation, the disturbance ゛−Σ,H7(i)Xj(i)′ added to the input signal Yj by the fourth calculation means 22
The value of can be made small, and the value of the disturbance "-ΣH=(i) Therefore, as detailed in the [Operation] section, learning of the filter coefficients can be speeded up.

The filter coefficient correction circuit 65 shown in FIG. 11(b) reduces the value of the filter coefficient used by the convolution integrator 7 and the sum-of-products calculation means 20 for a predetermined period for starting filter coefficient learning. This is an example of a filter coefficient correction circuit 65 that executes a process of correction. The filter coefficient correction circuit 65 determines which zone of a plurality of predetermined reference values Pt the magnitudes of the clock count value and the residual echo value E representing the passage of time from the start of learning fall into. A judgment circuit 68 and a multiplication circuit 69 which determines the multiplication coefficient β length according to the judgment result of the judgment circuit on the 6th day and executes a multiplication process of the determined multiplication coefficient β positive and the filter coefficient Hj(i). By providing this, the filter coefficient Hi (i) of the filter coefficient register 6 is operated to increase sequentially from a small value such as a zero value as learning progresses. According to this operation, the disturbance “−Σ−Hj” applied to the input signal Yj by the fourth calculation means 22
(i) The value of Xj (i)" can be made small, and the value of the disturbance "-ΣH, (i) can be made small, so that the learning speed of the filter coefficients can be increased as described in detail in the [Operation] section.

In the filter coefficient correction circuit 65 of FIG. 11,
It is also possible to use the value of the filter coefficient Hj(+) itself as a parameter for determining the progress of learning. That is, as learning progresses, the filter coefficient Hj
Since (i) reaches a steady value while being updated sequentially from a small value such as a zero value, the value of this filter coefficient H, + (+) is detected by means such as the low-pass filter circuit 50. The progress of learning is determined based on the progress of learning.

As explained in Fig. 1, in the first invention, the problem of division by Become. FIG. 12 illustrates an embodiment in which normalization processing using this average power is executed on the output side of the integrating means 24, and FIG. FIG.

The embodiment of the first invention shown in FIG. 12 uses the integration means 24 shown in FIG. 7(a) and the power calculation means 25 shown in FIG. 9(b). An embodiment is disclosed. This power calculation means 25
From is the power value P.

Since (-) is to be output, a division circuit shown at 26a in the figure is used as the second calculation means 26. From now on, the low-pass filter circuit 5 (12) multiplication circuit 19 that constitutes the integration means 24 and the low-pass filter circuit 5 (12) multiplication circuit 19 that constitutes the power calculation means 25 are canceled out according to this division circuit 26a. Therefore, it is possible to omit it as shown in the figure. Here, the low-pass filter circuit 5 of the power calculation means 25
If a reciprocal circuit for calculating the reciprocal is provided between 0 and the delay circuit 57, l/P t (m) will be output from the power calculation means 25, so the second A configuration can be adopted in which a multiplication circuit is used as the calculation means 26. With this configuration, the division circuit 2
Since it is not necessary to have 6a corresponding to the number of taps of the adaptive filter, the number of gates is reduced, which is convenient for implementation.

In the embodiment shown in FIG. 12, the filter coefficient Hj(
Although only the configuration of filter coefficients HJ(t) is shown, in detail, as shown in FIG.
, an echo path gain calculation block 70 consisting of first and fourth calculation means 23, 22 and integration means 24 and a division circuit 26a are provided, and the power value Pj is calculated for these echo path gain calculation blocks 70. (i)
The configuration is such that one power calculation means 25 is provided.

The embodiment of the first invention shown in FIG. 14 is an embodiment in which the normalization process using the average power is executed on the input side of the integrating means 29. In this embodiment, a first cumulative addition circuit 71 that performs cumulative addition of N input signals Y with respect to time j is provided at the subsequent stage of the first calculation means 23, and an output signal X ) (m) , and a second cumulative addition circuit 73 that performs N cumulative additions of the calculated value of the square value calculation circuit 72 with respect to time j, The first cumulative addition circuit 7 according to the division circuit 26b provided on the input side of the
1 and the calculated value of the second cumulative addition circuit 73 (corresponding to the power value), and input the divided value to the integrating means 29 to obtain the updated value H of the filter coefficient.
, +, 1(+*). That is, the signal value manually input to the first cumulative addition circuit 71 is as described above: (YJ-1, HJ(i)Xj(+)) Xj(m)=
(hj(m)Xj(m)+Qj(m)XJ(m))Xj
(m)−J(s+)Xj”(m)+Qj(m)X7”(
m), so if the first cumulative addition circuit 71 calculates the cumulative addition of +1 to +N to j=, the calculated value of the first cumulative addition circuit 71 is [h, (amount) 10Qj (■)]] ΣX7-a(m), on the other hand, the second cumulative addition cycle! The calculated value of 73 is ΣXJ,, 1z(ll). Therefore, the division circuit 26b calculates (hj(w)+
Since Qt(m)) is calculated, normalization processing can be realized on the input side of the integrating means 24.

If a configuration is adopted in which the normalization process is executed on the input side of the integrating means 29 as in the embodiment shown in FIG.
Since the sampling period is converted to a slow one according to the first order of magnitude, it is convenient for the implementation of the dividing circuit 26b and the integrating means 29.

While the first invention attempts to solve the problem of division by X7(m)=0, which is a problem when using the arithmetic system shown in FIG. As explained in FIG. 3, the solution is to stop the calculation.

FIG. 15 illustrates an embodiment of the second invention.

In the figure, the same parts as those explained in FIG. 3 are indicated by the same symbols. In this embodiment, as the integrating means 32, the one shown in FIG. It is. 74
75 is a first coefficient setting circuit that sets the low-pass filter coefficient α' of the multiplication coefficient for the multiplier 18 constituting the integrating means 32, and 75 sets the multiplication coefficient (1 -α°). In the embodiment configured as described above, when the signal value determining means 33 determines that the output signal
Then, the second coefficient setting circuits 74 and 75 are forced to set "Bi" as the value of the low-pass filter coefficient α. When "1" is set in this way, the multiplier 19
Since the signal value is no longer inputted, the delay device 17 that holds the integrated processing value of the integrating means 32 operates to continue holding the value up to that point. That is, the integration process of the integrating means 32 is stopped, and the problem of division by Xt(s)=0 can be solved.

In the first and second inventions, when learning filter coefficients by the "learning identification method", updating I of the filter coefficients is performed.
an integrating means 24 for extracting a HJ−+ (s);
A feature is that an integral coefficient of 29.32 (which becomes a low-pass filter coefficient α when configured by the low-pass filter circuit 50) can be set to a constant. By being able to set it as a constant in this way, it becomes possible to eliminate the drawback of the conventional "learning identification method" that the estimation accuracy of the filter coefficient changes depending on the size of the output signal X = (m). It is.

On the other hand, the third invention and the fourth invention are basically:
This invention is based on the conventional "learning identification method."

That is, in the third invention, in the process of deriving the arithmetic system shown in Fig. 6, it became clear that the conventional "learning identification method" constitutes a low-pass filter as shown in Fig. 32. In response to this, the present invention is configured to include a method for speeding up convergence of filter coefficients that can be applied to the first and second inventions. As shown in FIG. 11, for a predetermined period at the start of learning, all or part of the filter coefficient H7(i) of the filter coefficient register 6 is set to a substantially zero value or to a substantially small value. The filter coefficient correction circuit 65 is configured to provide a fourth
The disturbance added to the input signal Y by the calculation means 22 of
-X, Hj (i) The configuration is configured as in the embodiment described above, and at the beginning of learning, priority is given to speeding up convergence by selecting the arithmetic averaging circuit 54, and when learning has progressed to a certain extent, by selecting the integrating circuit 43a. Priority is given to estimation accuracy, and the learning speed and accuracy of filter coefficients is increased.

The fourth invention is that in the process of deriving the arithmetic system shown in Figure 6, the normalization process using the norm in the conventional "learning identification method" actually reduces the time constant fluctuation due to the amplitude fluctuation of the output signal. In response to the fact that it has become clear that only relative suppression is being performed, the number of output signals Xj(i) stored in the output signal register 5 is changed to the number of output signals filter coefficient Hj(+)
By setting the value regardless of the number of , the norm value is increased, and the negative influence of the time constant fluctuation that the conventional "learning identification method" has is reduced. In view of the fact that the normalization process using the norm merely reduces the output signal component in the low-pass filter coefficient α, the power calculation configured by the embodiment disclosed in FIG. The power value P J+ P 7 (m) obtained by means 41 and calculation means 42 is configured to be used as the norm value, thereby effectively reducing the adverse effects of time constant fluctuations that the conventional "learning identification method" has. It is structured as follows.

Furthermore, the fourth invention is configured to include a method for speeding up convergence of filter coefficients that can be applied to the first and second inventions. That is, as shown in FIG. 11, for a predetermined period at the start of learning, all or part of the filter coefficients HJ(i) in the filter coefficient register 6 are set to a substantially zero value or to a substantially small value. By reducing the value of the pseudo echo −ΣH,(i)Xj(i)” added to the input signal Y by the difference circuit 9, the filter The purpose is to speed up learning of coefficients.

As is clear from the low-pass filter expression of the "learning identification method" shown in FIG. 32, the disturbances to be eliminated in the "learning identification method" are: (1) S, +N.

■ Σ, hj (i) Xj (i) ■ Σ - Hj (i) It can be ignored. On the other hand, the disturbance (2) is a disturbance that is specially applied within the echo canceller, and is presumed to have the effect of canceling out the disturbance (2) and accelerating the convergence. The expected effect of applying this disturbance ■ is (
As explained in formula 10), 1Σ, hz(i)Xj(i) l > lΣ, Δj(i
) X = (t) However, Δj(i) −h *(i)
-H, when (i) holds true. In order to satisfy the condition of the GO) formula, (a) stop applying the disturbance (2) at the beginning of learning until the 00 formula is satisfied.

(b) Application is applied sequentially in accordance with the progress of learning while shifting the application start timing for each filter coefficient HJ (+).

(C) Apply a small filter coefficient H7(+) at the beginning of learning, and gradually increase it as learning progresses.

It is possible to take this approach. The filter coefficient correction circuit 65 shown in FIG. 11 implements this method.

If method (a) is adopted, since the disturbance (2) is not applied at the beginning of learning, the estimation of the filter coefficients will be performed stably without divergence, no matter how large the disturbances (2) and (2) are. From now on, this method (a
), in order to speed up the convergence, as mentioned above, when the disturbance ■ is not applied, the integration means 24, 29, 32, 43 configured by the arithmetic mean method are used.
When applying the disturbance (2), it is preferable to use the integrating means 24.29, 32.43 constituted by a low-pass filter.

Method (b) is a method that attempts to satisfy Equation 00 by reducing the number of filter coefficients applied as disturbance (2) at the beginning of learning when the estimation error is large. FIG. 16 shows the results of a simulation conducted to see the difference in convergence speed due to the difference in the number speed of applied filter coefficients. Here, this simulation
The calculation was performed using the arithmetic system of the first invention shown in FIG. 2, the output signal X was white noise, the reverberation time of the system was 64 m5, the sampling period was 8 kHz, and the number of taps of the adaptive filter was 512.

Further, the learning process was also performed for the filter coefficients Hj(i) to which no application was applied (that is, the method (a) was used in combination). Note that the reason why the integration process using the arithmetic mean method was not incorporated is that learning of filter coefficients that have not been applied will be executed due to the disturbance (3) caused by the filter coefficients that have already been applied.

FIG. 16(a) shows the filter coefficient H during the disturbance (2).

This is a simulation result when the number of items (i) is increased by one every sampling period, where the horizontal axis represents time and the vertical axis represents echo cancellation residual. 16th
Figure (b) is the simulation result when the number is increased by one every two sampling periods, and Figure 16 (C) is,
Figure 16(d) shows the simulation result when the number is increased by one every three sampling cycles, and Figure 16(d) shows the simulation result when the number is increased by one every four sampling cycles. e) is the simulation result when increasing by 1 every 6 sampling periods, the first
FIG. 6(f) shows the simulation results when the number is increased by one every 10 sampling periods.

This simulation result confirmed the effectiveness of sequentially applying disturbances while waiting for the estimation error of filter coefficients to decrease in order to speed up convergence. It became clear that this would delay convergence. That is, 00
It became clear that once the formula was satisfied, it was necessary to immediately apply the disturbance (2).

As interpreted from FIG. 30 showing the calculation system of the "learning identification method", the estimated value of the filter coefficient is given as the response of the low-pass filter. Therefore, in the "learning identification method",
It can be interpreted that the estimated value of the filter coefficient constituting the disturbance (2) is small at the beginning of learning and gradually increases as learning progresses. From now on, "Learning identification method"
This can be said to be using method (C), and from now on, method (C) will be actively implemented for the third and fourth inventions based on the "learning identification method". As a result, it can be expected that the convergence of the filter coefficients can be made faster. Furthermore, compared to the "learning identification method", a method in which all filter coefficients have almost the same weight and do not constitute disturbance ■ (
By using method a) or method (b), it is possible to expect faster convergence of filter coefficients.

Figure 17 shows method (a) for this “learning identification method”.
The results of a simulation conducted to confirm the effectiveness of applying method (b) are illustrated. Figure 17(a) shows the simulation results of the "learning identification method" when no operation is performed for the disturbance ■.
b) is the simulation result of the ``learning identification method'' when the number of filter coefficients in the disturbance ■ is increased by one every four sampling periods, similar to Fig. 16(d). . From the results of this simulation, it was confirmed that if method (a) or method (method master) is also applied to the "learning identification method", the convergence of the filter coefficients can be made faster. That is, if method (a) and method (b) are applied to the third and fourth inventions,
It was confirmed that the convergence of the filter coefficients could be made faster.

Method (C) is a method in which the value of the filter coefficient applied as disturbance (2) is reduced to ensure that the 00 formula holds true at the beginning of learning when the estimation error is large. Figure 18 shows the effectiveness of this method (C). This figure shows the results of a simulation conducted to confirm the results.

FIG. 18(a) is the same as the simulation data in FIG. 17(a), and is the simulation result of the "learning identification method" when no operation is performed for the disturbance ■. Figure (b) shows the filter coefficient H
The correction coefficient β that is multiplied to reduce the value of j(i) is β=Qj−m+1≦0 β=j−/j O<j−m+1<2Jβ−1j−m
11≧2J However, Jm=j m+1 The simulation result of the first invention in the case where J is the time determined by the number of taps (however, the estimation operation is not performed for filter coefficients not included in the disturbance ■ where β=0) ), and the first
FIG. 8(C) shows the simulation results of the first invention under the same simulation conditions as FIG. 18(b), when the estimation operation is also executed for filter coefficients not included in the disturbance ■. .

From this simulation result, it was confirmed that by applying method (C), the convergence of the filter coefficients can be achieved faster than in the conventional "learning identification method J, which does not perform any operations on the disturbance (2).

Method used to reduce disturbance ■ (C)
There are various ways to set the correction coefficient β.For example, β-Oj -m+ l≦0 β-ja/m O<j-m+1<mβ= 1
3−m+12m, β” j, / j j < Jβ= j a
/J j≧J β= 1 3-m+ 1≧J, β='Oj-m+l≦O β-j, /j o<j-m+1<Jβ= 1
It is conceivable to adopt a setting method such as j −m+ 1≧J.

So far, we have disclosed a method for effectively canceling out the disturbance ■ and the disturbance ■, but in order to speed up the convergence of filter coefficient learning, there is another method to reduce the disturbance ■. There is a way to go. One method is to make the transmission start time of the output signal XJ(+) coincide with the filter coefficient learning start time. In other words, if such a method is adopted, the output signal Xj(i) constituting the disturbance ■ becomes "On" when i < j, so that the value of the disturbance ■ becomes small and the learning of the filter coefficients converges. This results in faster speeds.

FIG. 19 shows the results of a simulation conducted to confirm this point. Figure 19(a) is similar to Figure 18(a).
Under the simulation conditions b), the output signal Xj(
Figure 19 (b) is a simulation result of the "learning identification method" executed under similar conditions. This is the result. From this simulation result, it was confirmed that by making the transmission start time of the output signal Xz(i) coincide with the filter coefficient learning start time (by doing so, the convergence of filter coefficient learning can be accelerated. The effects of the configuration are commonly effective for the first, second, third, and fourth inventions.In addition, in the embodiments of FIGS. 12, 14, and 15,
Specifically, in order to match the transmission start time of the output signal Xj(+) and the filter coefficient learning start time, the output signal register 5 is reset at the start of filter coefficient learning, or In order to clear the filter coefficients, the output signal Xj having a value of 0'' is sent to the output signal register 5 before starting learning of the filter coefficients.

In addition, as another method for reducing disturbance ■, the value of the filter coefficient assigned to the tap of the adaptive filter that corresponds to the shortest echo path delay of the system to be processed (the system that generates the echo to be canceled) is There is a method of taking advantage of the fact that it is turned on and setting it as ``0'' in advance. By using this method, the disturbance ■ will be reduced, so the convergence of filter coefficient learning will be faster. The effects of this configuration are the first, second,
This is commonly effective for the third and fourth inventions.

In order to make the first and second aspects of the present invention more effective, the value of the low-pass filter coefficient α° of the low-pass filter circuit 5 (12) constituting the integrating means 24, 29, 32 is appropriate. In addition, in order to make the third aspect of the present invention more effective, it is necessary to appropriately set the value of the coefficient correction constant of the integrating means 43, and In order to make the fourth aspect of the present invention more effective, the value of the coefficient correction constant of the multiplication circuit 8 (12) needs to be appropriately set. A simulation performed to find the optimum value for the coefficient correction constant and the coefficient correction constant will be explained. Here, similar to the above simulation, this simulation was performed using white noise as the output signal xj, the reverberation time of the system as 64g1s, the sampling period as 8kl(z, and the number of taps of the adaptive filter as 512. The number of taps of the filter, 512, is a value corresponding to the reverberation time (64 ms x 13 kHz).

First, simulation results regarding the optimal value of the low-pass filter coefficient α' will be explained.

FIG. 20 shows simulation results of echo cancellation residuals when the low-pass filter coefficient α' is varied. Here, the simulation is performed on the first
The disturbance “−Σ,Hj(i)X” applied by the product sum calculation means 20 at this time is
j(i)'' is a type in which the number of filter coefficients Hj(i) during disturbance increases by one every four sampling periods. In Fig. 20(a), lr' = 1023 Ha024(
=0.9990) (D simulation result, second
In Figure 0 (b), α”=5111512 (=0.9980
) simulation result, Fig. 20(c), shows α'-
This is a simulation result of 255/256 (=0.9961). As can be seen from this simulation result, when the value of the low-pass filter coefficient α is large, the time constant becomes long but the echo cancellation residual becomes small;

It is clear that when the value of is small, the time constant becomes short but the echo cancellation residual tends to become large.

This result supports the validity of the analytical model that the "learning identification method" derived by the present inventor constitutes a low-pass filter.

The optimal value of the low-pass filter coefficient α' lies at the value of α' at which the echo cancellation residual becomes saturated. This is because even if the value of α'' is increased further, it will not be possible to improve echo cancellation and will only lead to an increase in the time constant.According to the simulation results shown in Figure 20, The value of the optimal low-pass filter coefficient α is α”=5111512 (
= 0.9980). In order to clarify this point, a simulation was performed with the number of taps of the adaptive filter set to 128 in order to increase the amount of echo residual and check whether the estimation was correct. 21st
The simulation results are illustrated in the figure. In Fig. 21(a), a' = 1023/1024 (=0.999
The simulation result of 0), FIG. 21(b), shows α'
The simulation result of =5111512 (=0.9980), FIG. 21(c), is α'-127/12B (=
0.9922) is the simulation result. As can be seen from this simulation result, the value of α° at which the echo cancellation residual becomes saturated is α”−5111512
(=0.9980), that is, around the value determined by the number of taps corresponding to the reverberation time of the system expressed by the above-mentioned formula 00.

From now on, in the first and second inventions, the integrating means 24
, 29.32, it is concluded that it is preferable to use values around the value determined by the number of taps corresponding to the reverberation time of the system. However, the presence of additional noise N (the presence of S in the received signal can be ignored using a general method that detects double talk and stops learning the filter coefficients), so in reality, , it is necessary to use a value larger than the value determined by the number of taps corresponding to the reverberation time of the system as the value of the integral coefficient of the integrating means 24, 29, 32.

Next, simulation results regarding the optimal value of the coefficient correction constant will be explained.

FIG. 22 illustrates simulation results of echo cancellation residuals when the coefficient correction constant K of the "learning identification method" is varied. Here, the simulation is performed using the disturbance "-Σ" applied by the product-sum calculating means 20, similar to the simulation in FIG.

HJ(i)X,(i)'' is the disturbance filter coefficient HJ
The number of items (i) was increased by one every four sampling periods. Also, the number of taps of the adaptive filter is 1
I went at 28. Figure 22(a) shows the simulation results for K=0.125, and Figure 22(b) shows the simulation results for K-0,25.
The simulation result of FIG. 22(e) is the simulation result of K=1. Here, from the formula (7) α = [I KXJ'' (+1) / ΣXt' (i) 1, the average value α of the low-pass filter coefficient α can be regarded as α1 = 1 Number of filter taps 021, so FIG. 22(a) shows the first and second inventions (hereinafter sometimes referred to as inventions of the constant coefficient method) in which the low-pass filter coefficient α' has a value such as "1023/1024 (=0.9990)" Corresponding to the simulation results of
In FIG. 22(b), the low-pass filter coefficient α is “51115
12(=0.9980)", and FIG. 22(C) shows that the low-pass filter coefficient α' is "127/12B(=0.9980)".
This corresponds to the simulation result of the invention of the constant coefficient method with the value ``9922)''.As can be seen from this simulation result, the value of K at which the echo cancellation residual becomes saturated is Before and after, that is,
Similar to the first and second inventions, which are inventions of the constant coefficient method, 0
It is clear that the value is around the value determined by the number of taps corresponding to the reverberation time of the system determined by Equation 0.

From now on, in the third invention as well, as in the first and second inventions which are constant coefficient method inventions, the value of the coefficient correction constant corresponding to the integral coefficient of the integrating means 43 corresponds to the reverberation time of the system. It can be concluded that it is preferable to use values around the value determined by the number of taps. and,
It can be concluded that even in the fourth invention, which is equivalent to the third invention, it is preferable to use values around the value determined by the number of taps corresponding to the reverberation time of the system as the value for the coefficient correction constant. become. However, in any case, due to the presence of additional noise N, it is actually necessary to use a value smaller than this.

Finally, we will discuss the convergence characteristics of the invention of the constant coefficient method, which makes it possible to set the integral coefficient of the integrating means to a constant, and the conventional "learning identification method".
We will explain the simulations we conducted to compare the convergence characteristics of . Here, in this simulation, the number of taps of the adaptive filter is 512, which corresponds to the reverberation time.
I set it to .

FIG. 23 illustrates the simulation results.

FIG. 23(a) is a simulation result of the convergence characteristic of the invention of the constant coefficient method in which the value of the low-pass filter coefficient α° is “1023/1024”, which is also shown in FIG. 20(a). Figure 23(b) shows α' according to equation 02).
This is the convergence characteristic of the conventional "learning identification method" in which the value of the coefficient correction constant associated with = 1023/1024 is "0.5". Moreover, in FIG. 23(c), the value of the low-pass filter coefficient α is '2(14)7/2(14)
23(d) is a simulation result of the convergence characteristic of the invention of the constant coefficient method with a value of 8", and FIG. This is the convergence characteristic of the conventional "learning identification method" in which the coefficient correction constant has a value of "0.25".

From this simulation result, if the disturbance suppression ability is the same, that is, if the value of the low-pass filter coefficient α′ and the value of the coefficient correction constant are equivalently equal, the invention of the constant coefficient method is significantly better. It is clear that learning of filter coefficients can be performed with fast convergence characteristics.

FIG. 24 shows the simulation results of the convergence characteristics of the invention of the constant coefficient method and the simulation results of the convergence characteristics of the conventional "learning identification method" plotted on the same graph. Here, the horizontal axis shows time in seconds, and the vertical axis shows d
The amount of echo suppression represented by B is shown. In this simulation, the number of taps of the adaptive filter was 256. Therefore, the invention of the constant coefficient method of α"-"2(14)7/2(14)B" and the "learning identification method" of =1/8 should be compared because they have the same disturbance suppression ability. It is a thing,
Furthermore, the invention of the constant coefficient method of α='4095/4096'' should be compared with the ``learning identification method'' of 1/16 because they have the same disturbance suppression ability. As you can see from this figure, α”-2(14)7/2(14)8”
Using the invention of the constant coefficient method, convergence is approximately 0.4 seconds faster than the conventional r learning identification method, and α° = 4095/4
096'' invention of the constant coefficient method, convergence will be approximately 1 second faster than the conventional [Learning Identification Method J].

In the conventional learning identification method, in order to speed up the convergence of filter coefficient learning, at the beginning of learning, the coefficient correction constant is set to a large value to speed up convergence, and at the same time as learning progresses. In this case, a method was adopted in which the coefficient correction constant was set to a small value to improve estimation accuracy. However, according to the simulation results shown in Figure 25 conducted with K = 1.75, the convergence characteristics of the conventional "learning identification method" with the value of K = 0.5 shown in Figure 23 (b) It becomes clear that the convergence is not accelerated in any way compared to the convergence characteristics of the conventional learning identification method J, which has a value of 0.25 as shown in Figure 23(d).More detailed simulations According to the results, it was confirmed that even if K-1 or more is used, the convergence characteristics are not improved over the convergence characteristics when K=1.

On the other hand, according to the invention based on the constant coefficient method, the second
As shown in FIG. 0 (C), it has been confirmed that the convergence characteristics are improved even when the value of the low-pass filter coefficient α° is “255/256”. This α” = 255/256 is
Since the number of taps at this time is 512, this corresponds to =2 according to equation (12).

As can be seen from this, the invention based on the constant coefficient method, which aims to improve convergence characteristics by controlling the value of the low-pass filter coefficient α, is better than the invention based on the constant coefficient method, which aims to improve the convergence characteristics by controlling the value of This makes it possible to improve convergence characteristics more effectively than the conventional ``learning identification method'' that attempts to improve convergence characteristics.

Although the illustrated embodiment has been described, the present invention is not limited thereto. For example, in an arithmetic circuit with a division configuration, a multiplication configuration is adopted in which the reciprocal is calculated first and then the multiplication process is executed. It may be configured as such. Furthermore, the addition circuit that performs addition processing on the inverted signal value may be configured with a subtraction circuit.

In addition, although the example was disclosed as a hardware image,
The present invention is not limited to this, and may be realized by a software function or in combination with a software function. and,
The application of the present invention is not limited to echo cancellers as described in the embodiments.

〔Effect of the invention〕

As explained above, according to the present invention, when learning the filter coefficients of an adaptive filter according to the learning identification method, it is possible to properly learn the filter coefficients without causing fluctuations in estimation accuracy. As a result, learning of filter coefficients can be performed faster than before. From now on, it will be possible to realize adaptive filters with effective filter functions.

[Brief explanation of drawings]

Figures 1 and 2 are the principle configuration diagrams of the first invention of the present invention, Figure 3 is the principle configuration diagram of the second invention of the present invention, and Figure 4 is the principle configuration diagram of the third invention of the present invention. Figure 5 shows the fourth embodiment of the present invention.
Figure 6 is an explanatory diagram of the arithmetic system for learning the filter coefficients of the adaptive filter, which is the basis of the present invention, and Figure 7 is an integral diagram applicable to the first and second inventions. An embodiment of the means; FIG. 8 is an embodiment of the integrating means applicable to the third invention; FIG. 9 is an embodiment of the power calculation means; FIG. 10 is an embodiment of the sum-of-products calculation means; FIG. 11 shows an embodiment of the correction means, FIG. 12 shows an embodiment of the first invention in which normalization processing is executed on the output side of the integrating means, and FIG. 13 shows the overall configuration of the embodiment of FIG. 12. 14 shows an embodiment of the first invention in which the normalization process is executed on the input side of the integrating means, FIG. 15 shows an embodiment of the second invention, and FIG. 16 shows the method of applying the disturbance ■. Figure 17 is an explanatory diagram of a simulation conducted to see the difference in convergence speed due to the difference. Figure 17 is an explanatory diagram of a simulation conducted to examine the effectiveness of controlling disturbance ■ for the learning identification method. Figure 18 Figure 19 is an explanatory diagram of a simulation conducted to examine the effectiveness of controlling disturbance ■; Figure 19 is an explanatory diagram of a simulation conducted to examine the effectiveness of reducing disturbance ■; Figure 20 is an explanatory diagram of simulation conducted to examine the effectiveness of reducing disturbance ■. Figure 21 is an explanatory diagram of a simulation conducted to examine the details of the simulation results in Figure 20. Figure 22 is an explanatory diagram of a simulation conducted to examine the optimal value of α''. Figure 23 is an explanatory diagram of a simulation conducted to examine the value of . An explanatory diagram of the convergence characteristics of the learning identification method and the convergence characteristics of the constant coefficient method. Figure 25 is an explanatory diagram of the simulation conducted to investigate the convergence characteristics when the value of K is increased. Figures 26 and 27 are A configuration diagram of the echo canceller, Figure 28 is an explanatory diagram of the calculation system of the learning identification method, Figure 29 is a diagram of the conventional norm calculation circuit, Figure 30 is an explanatory diagram of the analytical model of the learning identification method, and Figure 31 is an illustration of the calculation system of the learning identification method. Fig. 32 is an explanatory diagram of the analytical model of the learning identification method, and Fig. 33 is an explanatory diagram of the simini-region representing fluctuations in the low-pass filter coefficient. In the figure, ■ indicates the subscriber terminal. , 4 is an adaptive filter, 5 is an output signal register, 6 is a filter coefficient register, 7 is a convolution integrating section, 8 is a filter coefficient updating section, 9 is a difference circuit, 16 is a summerizer, 17 is a delay device, 18 and 19 is a multiplier, 20 is a product sum calculation means, 21 is a correction means, 22 is a fourth calculation means, 23 is a first calculation means, 24 is an integration means, 25 is a power calculation means, and 26 is a second calculation means. means, 27 is cumulative addition means, 2nd is third calculation means, 29 is integration means, 30 is reciprocal calculation means, 31 is first calculation means, 32 is integration means,
33 is a signal value judgment means, 34 is a calculation stop means, 40 is a norm calculation means, 42 is a calculation means, and 43 is an integration means.

Claims (17)

[Claims] (1) Send an output signal to a system with an unknown transfer function,
Filter coefficient learning of an adaptive filter that processes to learn filter coefficients of an adaptive filter representing a transfer function equivalent to a transfer function given by the system from the output signal and an input signal that is a response to the output signal. In the method, a first calculation means (
23), an integrating means (24) that emphasizes the filter coefficient component of the tap included in the calculated value of the first calculation means (23) according to a constant integral coefficient, and a power of the output signal associated with the tap. a power calculation means (25) for calculating a value, and the above-mentioned integration means (24)
The power value included in the output value of the power calculation means (2
5) second calculation means (26) for offsetting using the calculated value;
A filter coefficient learning method for an adaptive filter, characterized in that the filter coefficients assigned to identified taps are learned while being sequentially updated. (2) Send an output signal to a system with an unknown transfer function,
Filter coefficient learning of an adaptive filter that processes to learn filter coefficients of an adaptive filter representing a transfer function equivalent to a transfer function given by the system from the output signal and an input signal that is a response to the output signal. In the method, a first calculation means (
23); cumulative addition means (27) for cumulatively adding the values calculated by the first calculation means (23); and power calculation means (25) for calculating the power value of the output signal associated with the tap. The cumulative addition means (2
a third calculating means (2) for offsetting the power value included in the calculated value of 7) using the calculated value of the power calculating means (25);
8), and an integrating means (29) that emphasizes the filter coefficient component of the tap included in the calculated value of the third calculating means (28) according to a constant integral coefficient, so that the filter coefficient component can be assigned to the specified tap. A filter coefficient learning method for an adaptive filter, which is characterized in that the filter coefficients are learned while sequentially updating them. (3) Send an output signal to a system with an unknown transfer function,
Filter coefficient learning of an adaptive filter that processes to learn filter coefficients of an adaptive filter representing a transfer function equivalent to a transfer function given by the system from the output signal and an input signal that is a response to the output signal. In the method, the first calculation means (31) calculates a value corresponding to the multiplication value of the input signal and the reciprocal of the output signal associated with the tap of the identified adaptive filter;
) an integrating means (32) for emphasizing the filter coefficient component of the tap included in the calculated value according to a constant integral coefficient; and a signal for determining whether or not the output signal associated with the tap is at a predetermined minimum value. a value determining means (33), and when the signal value determining means (33) determines that the value is a minute value, stopping the arithmetic processing of the first calculating means (31) or the integrating means (32); Calculation stop means (
34) A filter coefficient learning method for an adaptive filter characterized by performing learning while sequentially updating filter coefficients assigned to identified taps. (4) In the filter coefficient learning method for an adaptive filter according to claims (1), (2), and (3), the value of the integral coefficient of the integrating means (24, 29, 32) determines the progress of filter coefficient learning. Additionally, a filter coefficient learning method for an adaptive filter is characterized in that the adaptive filter is controlled to increase from a value near zero to a steady value near one. (5) In the filter coefficient learning method for an adaptive filter according to claims (1), (2), and (3), the filter coefficient of a tap other than the specified tap is multiplied by the output signal associated with the tap. A first calculating means (22) comprising: a sum-of-products calculation means (20) for calculating the sum of the values; and a fourth calculation means (22) for calculating a difference between the input signal and the value calculated by the sum-of-products calculation means (20). A filter coefficient learning method for an adaptive filter, characterized in that the calculation means (23, 31) are configured to use the calculated value of the fourth calculation means (22) instead of the input signal. (6) In the filter coefficient learning method for the adaptive filter according to claim (5), for the predetermined period for starting filter coefficient learning, the values of the filter coefficients of the taps used by the sum of products calculating means (20) are set to zero or A correction means (21) is provided for correcting the value calculated by the sum-of-products calculation means (20) to a substantially small value by setting a small value or by setting a small number of filter coefficients. , A filter coefficient learning method for adaptive filters. (7) Send an output signal to a system with an unknown transfer function,
Filter coefficient learning of an adaptive filter that processes to learn filter coefficients of an adaptive filter representing a transfer function equivalent to a transfer function given by the system from the output signal and an input signal that is a response to the output signal. In the method, a sum-of-products calculation means (20) calculates the sum of the multiplication values of the filter coefficients of the taps other than the specified tap and the output signal associated with the taps, and Fourth calculation means (22
), and an integrating means (
43), and for a predetermined period of time to start learning the filter coefficients, the values of the filter coefficients of the taps used by the sum-of-products calculation means (20) are set to zero or a small value, or the number of filter coefficients is reduced. By providing a correction means (21) that corrects the value calculated by the sum-of-products calculation means (20) to a substantially small value by setting the filter coefficients to be assigned to the identified taps, A filter coefficient learning method for adaptive filters that is characterized by processing that learns while updating. (8) In the adaptive filter filter coefficient learning method according to claims (6) and (7), the integrating means (24, 29, 32, 43) comprises the correcting means (2).
According to 1), when the product-sum calculating means (20) uses a filter coefficient with a zero value, it emphasizes the filter coefficient component of the tap specified according to the arithmetic mean method, and when it uses a filter coefficient with a non-zero value, A filter coefficient learning method for an adaptive filter, characterized in that the filter coefficient component of a tap specified according to a low-pass filter is emphasized. (9) In the filter coefficient learning method for an adaptive filter according to claims (1), (2), and (3), the value of the integral coefficient of the integrating means (24, 29, 32,) corresponds to the reverberation time of the system. 1. A filter coefficient learning method for an adaptive filter, characterized in that the coefficient is set to a value greater than or equal to a value adjacent to a value determined by the number of taps. (10) In the filter coefficient learning method for an adaptive filter according to claim (7), the value of the coefficient correction constant corresponding to the integral coefficient of the integrating means (43) is a tap corresponding to the reverberation time of the system. A filter coefficient learning method for an adaptive filter characterized in that the coefficient is set to a value less than or equal to a value near a value determined by a number. (11) Send an output signal to a system with an unknown transfer function, calculate a norm value from the time series data of the output signal, and combine the time series data of the output signal with the filter coefficients of the adaptive filter. A pseudo echo is generated according to the integral integral, a residual echo is obtained from the pseudo echo and an input signal that is a response to the output signal, and a multiplication value of the residual echo, the output signal, and a coefficient correction constant is calculated. By normalizing with the norm value to find the updated value of the filter coefficient and updating the filter coefficient, the filter coefficient of the adaptive filter that represents the transfer function equivalent to the transfer function given by the system is learned. In the filter coefficient learning method of the adaptive filter to be processed, for a predetermined period of time to start learning the filter coefficients, the value of the filter coefficients used for generating pseudo echoes is set to zero or a small value, or the number of filter coefficients is A correction means that corrects the value of the pseudo echo so that it becomes substantially small by setting it to a smaller value (
21a) A filter coefficient learning method for an adaptive filter. (12) In the filter coefficient learning method for the adaptive filter according to claim (11), the value of the coefficient correction constant is configured to be set to a value equal to or less than a value near a value determined by the number of taps corresponding to the reverberation time of the system. A filter coefficient learning method for an adaptive filter is characterized by: (13) Claims (1), (2), (3), (7), (1
1) In the adaptive filter filter coefficient learning method described above, the adaptive filter is configured such that there is no output signal associated with the tap of the adaptive filter at the start of the learning process of the filter coefficients of the adaptive filter. Filter coefficient learning method for filters. (14) Claims (1), (2), (3), (7), (1
1) In the adaptive filter filter coefficient learning method described above, the filter coefficient of the adaptive filter corresponding to the shortest delay time of the system is configured to be set to zero value. Coefficient learning method. (15) Send an output signal to a system with an unknown transfer function, calculate a norm value from the time series data of the output signal, and combine the time series data of the output signal with the filter coefficients of the adaptive filter. A pseudo echo is generated according to the integral integral, a residual echo is obtained from the pseudo echo and an input signal that is a response to the output signal, and a multiplication value of the residual echo, the output signal, and a coefficient correction constant is calculated. By normalizing with the norm value to find the updated value of the filter coefficient and updating the filter coefficient, the filter coefficient of the adaptive filter that represents the transfer function equivalent to the transfer function given by the system is learned. In the filter coefficient learning method of the adaptive filter to be processed, the number of time-series data of the output signal stored for norm value calculation is configured to be larger than the number of filter coefficients of the adaptive filter, and the stored output A filter coefficient learning method for an adaptive filter characterized by calculating a norm value according to time-series data of a signal and performing normalization processing. (16) Send an output signal to a system with an unknown transfer function, calculate a norm value from the time series data of the output signal, and combine the time series data of the output signal with the filter coefficients of the adaptive filter. A pseudo echo is generated according to the integral integral, a residual echo is obtained from the pseudo echo and an input signal that is a response to the output signal, and a multiplication value of the residual echo, the output signal, and a coefficient correction constant is calculated. By normalizing with the norm value to find the updated value of the filter coefficient and updating the filter coefficient, the filter coefficient of the adaptive filter that represents the transfer function equivalent to the transfer function given by the system is learned. In a filter coefficient learning method of an adaptive filter to be processed, a power calculation means (41) is provided for calculating a power value for an output signal, and normalization processing is performed using the calculated value of the power calculation means (41) as a norm value. A filter coefficient learning method for an adaptive filter characterized by processing to (17) In the adaptive filter filter coefficient learning method according to claim (16), the power value of the output signal associated with the filter coefficient to be learned is calculated using the output value of the power calculation means (41) as input. A filter coefficient learning method for an adaptive filter, characterized in that a calculation means (42) is provided, and a value calculated by the calculation means (42) is used as a norm value to perform normalization processing.