JP3147207B2 - Adaptive unknown system output estimation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for adaptively estimating the output of an unknown system in an acoustic echo canceller (echo canceller), an adaptive sound field controller, and the like, and more particularly to a method for estimating the output by a high-speed projection method.

[0002]

2. Description of the Related Art In this specification, time is described as discrete time. For example, the value of signal x at time k is represented as x (k). The present invention provides an acoustic echo canceller,
Although it can be applied to the estimation of the output of an unknown system adaptively in adaptive sound field control or the like, an explanation will be given here taking an acoustic echo canceller as an example. FIG. 4A shows a general configuration of the acoustic echo canceller. The input signal x (k) is supplied to a transfer function estimator 11, a convolutional pseudo echo path 12 and a loudspeaker 13, and the sound output from the loudspeaker 13 is collected by a sound receiver 14 via an unknown system (echo path) 15. The output signal of the sound receiver 14 is output from the convolution unit 12 by the subtractor 16 at the estimated output y ＾
(k) is subtracted and the estimation error e (k) is output.

Here, the characteristics of the unknown system are considered to include the characteristics of the loudspeaker 13 and the sound receiver 14. Therefore, FIG.
It is assumed that the characteristic of the loudspeaker 13 and the sound receiver 14 in A is 1, that is, the loudspeaker 13 and the sound receiver 14 each output the same signal as the input signal. When the sound output from the loudspeaker 13 is received by the sound receiver 14 via the unknown system 15, problems such as acoustic echo and howling occur. An acoustic echo canceller is a device for preventing such sound from wrapping around. Input signal to loudspeaker 13 (= input signal to unknown system) x (k) and output of unknown system (= receiver 1
4) y (k) and the estimated value y ＾ (k) of the output of the unknown system
, Noise (transmission signal) n (k), and estimation error e (k) are represented by the following equation.

E (k) = y (k) + n (k) −y ＾ (k) (1) The transfer function of the unknown system is estimated as an L-tap FIR filter, and the coefficient of the filter is h ＾ ₁ (k ), H ＾ ₂ (k),
..., h ＾ _L (k). For simplicity, the filter coefficients are represented as a vector defined by the following equation. h ＾ (k) = [h ＾ ₁ (k), h ＾ ₂ (k),..., h ＾ _L (k)] ^T (2) where ^T represents transpose. The transfer function estimator 11 estimates the transfer function of the unknown system 15 from the input signal x (k) and the estimation error e (k). The estimation result is expressed as a filter coefficient h ＾ (k), which is transferred to the convolution unit 12. The convolution unit 12 convolves the input signal x (k) with the filter coefficient h ＾ (k) to generate an estimated value y ＾ (k) of the response of the unknown system 15.

[0005] _{y ^ (k) = Σ i} = 1 L x (k-i + 1) h ^ i (k) (3) The transfer function of an unknown system 15 is often change with time. Therefore, h ＾ which is the estimated value of the transfer function of the unknown system 15
(k) also needs to detect changes in the unknown system 15 and adaptively correct them. The transfer function estimator 11 calculates a filter coefficient correction vector δh ＾ (k) from the estimation error e (k) and the input signal x (k) to the unknown system 15. The filter coefficient h ＾ (k) is updated by adding δh ＾ (k) to itself.

H ＾ (k + 1) = h ＾ (k) + αδh ＾ (k) (4) where α is a scalar quantity called a step size. A method of modifying the filter coefficient h ＾ (k) is known as an adaptive algorithm. LM as an adaptive algorithm
The S (Least Mean Squares) algorithm and the learning identification method are well-known, but here the document "Ozeki, Umeda" Adaptive filter algorithm using orthogonal projection to affine subspace and its properties ", IEICE Magazine (A), J
67-A pp. 126-132, (Showa 59-2) ". Although the projection algorithm requires more computation than the learning identification method,
It has excellent adaptive characteristics for audio input signals. In the p-th order projection algorithm, the filter coefficient correction vector δh ＾ (k) is determined so that the following p simultaneous equations are satisfied.

Y (k) = [x (k), x (k−1),..., X (k−L + 1)] (h ＾ (k) + δh ＾ (k)) y (k−1 ) = [X (k−1), x (k−2),..., X (kL)] (h ＾ (k) + δh ＾ (k)) (5) y (k−p + 1) = [ x (k−p + 1), x (kp),..., x (kp−L + 2)] (h ＾ (k) + δh) (k)) That is, the filter coefficient correction vector δh ＾ (k) is .., K−p + 1 at time k, k−1,..., K−p + 1
Is determined so that the error becomes zero for the input signal of
Thus, if the input signals are similar, it is expected that the error will be reduced at other times.

FIG. 4B shows a conventional projection method using a transfer function estimating unit 1.
1 shows a configuration of an acoustic echo canceller applied to Example 1. That is, the transfer function estimating unit 11 includes the autocorrelation calculating unit 21, the β calculating unit 22, and the filter coefficient updating unit 23. In the conventional projection method, the update of the filter coefficient is performed by the following equation. h ＾ (k + 1) = h ＾ (k) + α (β ₁ (k) X (k) + β ₂ (k) X (k−1) +... + β _p (k) X (k−p + 1) ) (6) X (k) = [x (k), x (k−1),..., X (k−L + 1)] ^T (7) where β ₁ (k), β ₂ (k ),..., Β _p (k) are pre-filter coefficients, which are obtained by the β calculation unit 22 as solutions of the following simultaneous equations.

Β (k) ^TR (k) = [e (k), (1−α) e (k−1),..., (1−α) ^p−1 e (k−p + 1)] (8) where β (k) = [β ₁ (k), β ₂ (k),..., Β _p ] ^T. R (k) is an autocorrelation matrix of the input signal x (k), and i + 1 row j + 1 column (0 ≦ i, j ≦ p−1) of R (k)
The element r _ij (k) of is defined by the following equation.

[0010] determine the _{r ij (k) = r ji} (k) = X (ki) T X (kj) (9) update of _{r ij (k) (r ij} (k) from r _ij (k + 1) Is performed by the autocorrelation calculation unit 21 as in the following equation. When i ≠ 0 and j ≠ 0, r _ij (k + 1) = _{ri -1, j-1} (k) In other cases, r _0j (k + 1) = r _0j (k) + x (k + 1 ) X (k + 1-j) -x (k-L + 1) x (kj-L + 1) (10) The processing procedure for updating the filter coefficient h ＾ (k) once by the conventional projection method is shown in FIG. As shown in FIG. 5, the autocorrelation r _{i, j} (k) of the input signal x (k) is calculated by the autocorrelation calculator 21 according to the equation (10) (S ₁ ). Filter coefficient h ＾
The convolution operation with (k) is performed according to equation (3).
To obtain the estimated value y ＾ (k) of the output of the unknown system (S
₂ ). Next, the subtracter 16 subtracts the output estimation value y ＾ (k) of the convolution unit 12 from the output of the receiver 14, that is, calculates Equation (1) to calculate the estimation error e (k) (S ₃ ). .Beta. Calculation unit 2 using the estimation error e (k) and the autocorrelation matrix R (k).
2, the pre-filter coefficient β (k) is calculated by the equation (8) (S ₄ ), and the pre-filter coefficient β (k) and the input signal x (k) are calculated.
Then, the filter coefficient updating unit 23 updates the filter coefficient hk (k) according to the equation (6) (S ₅ ). One update of the filter coefficient h ＾ (k) requires about (p + 1) L product-sum operations. An acoustic echo canceller using a high-speed projection method using an intermediate variable in order to reduce the product-sum operation amount is known. This is because, as shown in FIG. 4C, the pre-filter coefficient β (k) is smoothed by the s updating unit 30 and supplied to the intermediate variable updating unit 31, converted into the intermediate variable z (k) and supplied to the convolution unit 30. The multiplier 33 multiplies the autocorrelation coefficient r _0i (k) from the autocorrelation calculator 25 by the smoothing pre-filter coefficient s _i (k) from the s updater 30, and multiplies the result by the convolution unit 32. Are added by an adder 35 and output as an output estimated value y ＾ (k). This high-speed projection method is an algorithm in which the number of necessary product-sum operations is reduced to about 2L. In the fast projection method, an intermediate variable z (k + 1) given by the following equation is introduced instead of h ＾ (k + 1).

Z (k + 1) = h ＾ (k + 1) −α [s ₁ (k) X (k) + s ₂ (k) X (k−1) +... + S _p−1 (k) X (k−p + 2)] (11) where s _i (k) = s _i−1 (k−1) + β _i (k), s ₀ (k) = 0 (i = 1,..., p) (12) This s _i (k) is called a smoothing pre-filter coefficient, and the calculation is performed by the s updating unit 30. Equation (6) for updating the filter coefficient by replacing h ＾ with z is the following regression equation for z.

Z (k + 1) = z (k) + αs _p (k) X (k−p + 1) (13) The time update of z is performed by the intermediate variable updating unit 31. The estimated value y ＾ (k) of the response of the unknown system is given by equations (3) and (11).
From the above, yk (k) = h ＾ (k) ^T X (k) = z (k) ^T X (k) + αΣs _i (k-1) X (ki) ^T X (k) (14) = 1 to p-1. If the autocorrelation value is used, y ＾ (k) = z (k) ^T X (k) + α {s _i (k−1) r _0i (k) (15)} is multiplied from i = 1 to p−1. The first term on the right side of Expression (15) is the output of the convolution unit 32, and the second and subsequent terms are the outputs of the multiplication unit 33.

FIG. 6 shows the procedure for updating the filter coefficients in the high-speed projection method. The autocorrelation calculator 21 calculates the autocorrelation r _ij (k) of the input signal x (k) according to equation (10) (S10). ₁ ), the convolution operation of the input signal x (k) and the intermediate variable z (k) is performed by the convolution unit 32, and the sum of the inner products of the smoothing pre-filter coefficient s _i (k) and the autocorrelation r _0i (k) To the multiplier 33
From these sums, the estimated value y ＾ (k) of the output of the unknown system is obtained, that is, the calculation of equation (15) is performed (S ₆ ). Next, the subtractor 16 calculates the estimation error e (k) by the equation (1) (S ₃ ), and the β calculator 22 calculates the pre-filter coefficient β (k) by the equation (8) (S ₄ ). , Its β (k) and the input signal x
(k), the s updating unit 30 updates the smoothing pre-filter coefficient s (k) according to the equation (12) (S ₇ ).
_The intermediate variable updating unit 31 updates the intermediate variable z (k) using _i (k) (S ₈ ).

As described above, in the high-speed projection method, the intermediate variable z (k) is updated without directly updating the transfer function (filter coefficient) h ＾ (k), and the estimated output y is calculated by the equation (15).
Calculate ＾ (k). In this case, the product-sum operation amount required for one update is about 2L when p << L. This is significantly reduced as compared with the case where the product-sum operation amount of the normal projection method is (p + 1) L.

Incidentally, in the acoustic echo canceller, there are two reasons why the estimation error e (k) greatly increases. One is a change in the unknown system 15, and the other is an increase in noise (transmitted voice) n (k). Although it is difficult to distinguish between the two, in the former case the filter coefficients must be updated, but in the latter case the estimation accuracy of the transfer function of the unknown system deteriorates if the filter coefficients are updated continuously. Therefore, it is necessary to stop updating the filter coefficients. As a method for satisfying this conflicting demand, FG / BG
(Fore Ground / Back Ground) method.

FIG. 7A is a diagram for explaining the FG / BG system. In FIG. 4A, a convolution unit 41 for the input signal x (k) is further provided, and the output estimation value of the convolution unit 41 and the sound receiver 14 are provided. Is subtracted by the subtractor 42, and the transfer determination unit 43 determines that the transfer is to be performed based on the estimated error e _f (k) resulting from the subtraction and the estimated residual e _b (k) from the subtractor 16. Turning on 44, the estimated value h ＾ (k) of the transfer function estimating unit 11
Is set in the convolution unit 41. This FG / BG method is 2
The folding section 12 is called a BG side folding section, and the folding section 41 is called an FG side folding section. The filter coefficient of the convolution unit 41 on the FG side is semi-fixed. on the other hand,
The filter coefficients of the convolution unit 12 on the BG side are sequentially updated based on the estimation result of the transfer function estimation unit 11. Transfer judgment unit 4
In No. 3, the powers of the two estimation errors e _f (k) and e _b (k) are compared, and the power of e _b (k) is smaller than the power of e _f (k), that is, the filter of the convolution unit 12. When it is determined that the coefficient represents the unknown system 15 better, the switch 4
4 is closed, and the filter coefficient of the convolution unit 12 on the BG side is set to F
The data is transferred to the convolution unit 41 on the G side.

[0017] Therefore, after the residual is increased by the change of the unknown system, e _b (k) is smaller than e _f (k) by the adaptive operation, the transfer occurs. On the other hand, if the noise (transmission voice) is large, the filter coefficient of the convolution portion 12 of the BG side is disturbed, e _b (k) is the transfer does not occur becomes larger than e _f (k).

[0018]

SUMMARY OF THE INVENTION Noise (transmitted voice) n
If the update of the filter coefficient is continued when (k) is larger than the response y (k) of the unknown system, the estimation accuracy of the transfer function of the unknown system is degraded. To avoid this, in such a case, it is necessary to stop the adaptive operation of the projection method. And time k _s
If you stop updating the filter coefficients later,
Estimated output in a subsequent _{y ^ (k s + m)} (m ≧ 0) is _{y ^ (k s + m)} = h ^ (k s) T X (k s + m) and made it desirable (16).

Now, in the case of the high-speed projection method, an intermediate variable is updated instead of updating the filter coefficient as described above. Therefore, as a configuration as an acoustic echo canceller having a filter coefficient update determination function, a configuration shown in FIG. 7B can be considered. That is, the update stop judging section 51 inputs the input signal x and the estimation error e, and sends an update stop signal to the intermediate variable updating section 31 when the input signal x is zero or x <e.

Time k_sHereinafter, the intermediate variable z (k)
When the update expression ((13)) is stopped, the time k_s+ M (m ≧
0), the estimated value y 出力 (k) of the output of the unknown system._s+ M) = h ＾
(K _s)^TX (k_s+ M) is obtained by the following equation from equation (15).
Can be y ＾ (k_s+ M) = z (k_s)^TX (k_s+ M) + αΣs_i(K_s-1) r_{0m + i}(K_s+ M) (17) Σ is from i = 1 to p−1 r in the above equation_{0m + i}(K_s+ M) is for i = 1,..., P-2
Then, it is updated by the following formula.

[0021] _{r 0m + 1 + i (k} s + m + 1) = r 0m + i + 1 (k s + m) + x (k s -i) x (k s + m + 1) -x (k _{s -iL) x (k s +} m-L + 1) (18) product-sum operation required for this update is 2 (p-2). i =
In _{p-1, r 0m + p} (k s + m) is not calculated because the value of m compared with the previous time is increased by one. Therefore, it is necessary to obtain the following equation according to the definition of equation (9).

R _{0m + p} (k _s + m + 1) = X (k _s + m + 1) ^T X (k _s −p + 1) (19) In the calculation of the above expression, L times of product-sum operation Is required. As described above, there is a problem that it is necessary to add about L + 2p new calculations in order to calculate Expressions (18) and (19). Further, equation (18), to calculate the (19), x (k _s -
1), ..., it is necessary to hold the _{x (k s -p-L +} 2), there is a problem that it requires L + p-2 pieces of extra storage space. FIG. 8 shows a procedure for updating the intermediate variables in the configuration shown in FIG. 7B. That is, first, the input signal x
The autocorrelation r _ij (k) of (k) is obtained (S ₁ ), and it is checked whether or not an intermediate variable update stop signal is generated from the update stop determination unit 51 (S ₉ ). Step S in which an estimated value y ＾ (k) of the output similar to step S _{6 in} FIG.
_Then, if there is an update stop signal, the autocorrelation calculator 21
In calculates the equation _{(18) r 0m + i (} k s + m) the calculated (S _10), further r _{0 m +} p by the equation _{(19) (k s + m} + 1)
The process proceeds to step S ₆ seeking (S _11). Thereafter, the estimation error e (k) is calculated in the same manner as in FIG. 6 (S ₃ ), the pre-filter coefficient β (k) is calculated (S ₄ ), and the smoothed pre-filter coefficient s _i (k) is updated. to (S _7). Then again checked for intermediate variable update stop signal (S _12), otherwise if intermediate variable z (k)
Returning to step S ₁ the update performed by equation (13) (S _8), if there is an intermediate variable updating stop signal intermediate variable z (k)
Returns to the step S ₁ without performing the update.

As described above, according to the high-speed projection method, the calculation amount can be significantly reduced as compared with the conventional projection method, but the intermediate variable update stop signal is output so as not to be affected by noise. In that case, the amount of calculation at that time is large,
In addition, the present invention requires extra storage area.
_{0m + i} (k _s + m) without calculating y ＾ (k _s + m)
(M ≧ 0) is obtained, and it is possible to provide an output estimation method by a high-speed projection method that is not affected by noise in a small amount of calculation and in a storage area.

[0024]

SUMMARY OF THE INVENTION In the unknown system output estimate method with fast projection algorithm in the present invention, after the time k _s that issued the coefficient update stop signal, i.e. the time k _s after stopping the adaptation resume adaptation Prefilter coefficient β
_i a _{(k), β i (k} ) = 0, i = 1, ..., and p (20). In this way, the output of the adder 35 in FIG. 4C without adding new operations, estimated output shown in equation _{(16) y ^ (k s} + m) = h ^ (k s) T X ( k
_s + m) (m ≧ 0). This will be described below using the time k _s +1 as an example.

By using equation (20), at time k _s , equation (12) becomes s _p (k _s ) = s _p −1 (k _s −1).
When this is substituted into equation _{(13), z (k s} +1) = z (k s) + αs p-1 (k s -1) x (k s -p + 1) becomes (21). On the other hand, the estimated value y ＾ (k _s +1) of the output of the unknown system at the time k _s +1 is obtained by Expression (15), and by Expression (14) equivalent thereto, y ＾ (k _s +1) = z ^{_{(k s +1) T X (}} k s +1) + αΣs i (k s) X (k s + 1-i) T X (k s +1) (22) Σ is written as i = 1 to p-1. Furthermore, from equations (12) and (20), s
_{Since i} (k _s ) = s _i−1 (k _s −1), the equation (22)
Right ^{_{side, z (k s +1) T}} X (k s +1) + αΣs i-1 (k s -1) X (k s + 1-i) T X (k s +1) = z (k s +1) T X _{(k s +1) + αΣs i} -1 (k s -1) X (k s + 1-i) T X (k s +1), (∵s 0 = 0) (p-1 from the Σ of the right side i = 2 _{up) = z (k s +1)} T X (k s +1) + αΣs i (k s -1) X (k s -i) T X (k s +1) (23) Σ is p-2 from i = 1 Up to. Also, z (k _s +
1) with the formula and the formula (21) to the k +1 was k _s in equation _{(11), z (k s} +1) = h ^ (k s) -αΣs i (k s -1) X (k s - i) (24) Σ is from i = 1 to p−2. Substituting equation (24) into equation (23), equation (22) becomes y ＾ (k _s +1) = h ＾ (k) ^T X (k _s +1) (25). In other words, in the case of the condition of Expression (20), when (k _s +1) is calculated by Expression (15), the estimated value y ＾ (k _s +1) of the output of the unknown system becomes a new calculation r _{01 + 1 + i} ( k _s +1
+1) (equation (18)), r _{01 + p} (k _s + 1 + 1) (equation (1
Required without 9)). Note that r _{0i in} equation (15)
(K + 1) is always calculated sequentially by the autocorrelation calculator 21. From time k _s +2 onward, the output of the adder 35 is given by h ＾ (k _s ) ^T X (k _s +
m), m ≧ 1. In this way, by setting β _i = 0 from the time k _s at which the adaptation is stopped until the adaptation is started, the new operation described in the section of [Problems to be Solved by the Invention] (Equation (18)) , (19)) and the storage area, it is possible to obtain the estimated value y ＋ (k _s + m) of the output of the unknown system.

Further, for equation (12), each k _s +1, k
Substituting i = 1,..., p for _s +2,... and considering the relationship between s ₀ (k) and β _i (k) = 0, i = 1,.
After the time point p after updating of the filter coefficient is stopped, s _i (k _s + m) = 0, i = 1,..., P, m ≧ p (26), so that the output of the multiplier 33 in FIG. Is always 0. From these _{things, z (k s + m)} T X (k s + m) = h ^ (k s) T X (k s + m) (27) i.e., z (k + m) = h ^ (k), m ≧ p (28) and the intermediate variable z is equal to h ＾.

[0027]

In FIG. 7B, when the update stop signal is issued, the update of the intermediate variable z (k) is stopped during that time. According to the present invention, while the update stop signal is issued, β = 0 is set. Update unit 3
0 by s _i (k) and z by the intermediate variable updating unit 31
The updating of (k) is performed, and as a result, the operations of Expressions (18) and (19) become unnecessary.

[0028]

FIG. 1 shows an embodiment of an acoustic echo canceller to which the present invention is applied, and portions corresponding to those in FIG. 7B are denoted by the same reference numerals. In this embodiment, a β reset unit 61 is inserted in series between the β calculation unit 22 and the s update unit 30, and β is set to 0 in the β reset unit 61 by an update stop signal from the update stop determination unit 51. Unlike FIG. 7B, y ＾ (k) can be generated even after updating of the intermediate variables is stopped. The processing procedure in this case is, as shown in FIG. 2, the input signal x (k)
The autocorrelation r _{i, j} (k) of is calculated by equation (10) (S ₁ ),
Convolution of the input signal x (k) with the intermediate variable z (k);
From the sum of the inner product of the smoothing pre-filter coefficient s _i (k) and the autocorrelation r _0i (k), the output estimation value y 未知 of the unknown system is obtained by equation (15).
(K) is obtained (S ₆ ), and an estimation error e (k) is calculated from the difference between the estimated value y ＾ (k) and the output of the sound receiver by equation (1) (S ₆ ).
₃ ), the pre-filter coefficient β (k) is calculated by equation (8) (S ₄ ), and it is checked whether there is a variable update stop signal (S ₉ ). If not, the smoothed pre-filter coefficient s _i ( k) is updated according to equation (12) (S ₇ ), and the intermediate variable z (k) is further updated.
Is updated by equation (13) (S ₈ ). That is, the process is the same as the process in FIG. 8 when there is no variable update stop signal, that is, the process in FIG.

However, if there is a variable update stop signal in step S ₉ , the β reset unit 61 calculates β _i (k) according to equation (20).
= 0 as the flow proceeds to step S ₇ (S _13). That is, in this embodiment, while the variable update stop signal is present, the same processing is performed as in the case where there is no variable update stop signal except that β _i (k) is set to 0. There is no increase in the number of product-sum operations, and a variable update stop signal is generated 1
The result of convolution of the filter coefficient h ＾ before the time and the input x (k) at the current time becomes the estimated value y ＾ (k) of the response of the unknown system.

Next, FIG. 3 shows an embodiment in which the present invention is applied to an acoustic echo canceller which operates the high-speed projection method shown in FIG. 4C and the FG / BG method shown in FIG. 7A.
A, parts corresponding to those in FIG. 1 are denoted by the same reference numerals. The filter coefficient of the convolution unit 32 is z (k), but the filter coefficient of the convolution unit 41 is h ＾ (k), and the difference between the output estimated value y ＾ (k) and the output of the sound receiver 14 is obtained. in a BG side estimation error e _b (k) is smaller than a difference between the output estimated value and the sound receiving unit 14 of the convolution portion 41 FG side estimation error e _f (k), determines the transfer in the transfer determination unit 43 Then, the judgment is given to the β reset unit 61 and the switch 44, and β (k)
Is set to 0 for at least the time p, that is, after the intermediate variable z (k) from the intermediate variable updating unit 31 becomes equal to h ＾ (k), the switch 44 is turned on, and the filter coefficient h at that time is set. ＾ (k) is transferred to the convolution unit 41. The operation when the update stop signal is generated from the update stop determination unit 51 is shown in FIG.
Is the same as

[0031]

As described above, according to the present invention, when the noise (transmitted voice) is large, the adaptive operation of the high-speed projection method can be stopped, and the deterioration of the estimation accuracy of the filter coefficient can be prevented. It is possible to convert an intermediate variable into a filter coefficient without increasing the amount of calculation and the storage area.

When the high-speed projection method and the FG / BG method are used, β is set to be equal to or longer than the p time at the time of transfer, so that the filter coefficient h ＾ (k) is set in the convolution section on the FG side. And a relatively simple configuration can be achieved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a functional configuration example of an acoustic echo canceller to which an embodiment of the invention of claim 1 is applied.

FIG. 2 is a flowchart showing an example of a processing procedure in FIG. 1;

FIG. 3 is a block diagram showing a functional configuration example of an acoustic echo canceller to which an embodiment of the invention of claim 2 is applied.

4A is a block diagram illustrating a general functional configuration of an acoustic echo canceller, FIG. 4B is a block diagram illustrating a functional configuration of an acoustic echo canceller using a conventional projection method, and FIG. FIG. 3 is a block diagram illustrating a functional configuration of an echo canceller.

FIG. 5 is a flowchart showing a filter coefficient updating procedure in an acoustic echo canceller using a conventional projection method.

FIG. 6 is a flowchart showing an intermediate variable updating procedure in a conventional acoustic echo canceller using a high-speed projection method.

FIG. 7A is a block diagram illustrating a functional configuration of a conventional acoustic echo canceller using the FG / BG system, and FIG. 7B illustrates a configuration in which an update stop determination function is added to a conventional acoustic echo canceller using the high-speed projection method; It is a block diagram.

FIG. 8 is a flowchart showing an intermediate variable updating procedure of the echo canceller shown in FIG. 7B.

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-3-226026 (JP, A) JP-A-6-13940 (JP, A) 1990 National Institute of Electronics, Information and Communication Engineers Spring Annual Conference Proceedings [Section 3] (1990-3-18-21) p. 3-322 IEIEC Transactions on Fundamentals of Electronics, Communications and Computer Sciences E78-A [10] (1995-10) p. 1355-1361 (58) Fields investigated (Int. Cl. ⁷ , DB name) H03H 21/00 H04B 3/23 JICST file (JOIS) Practical file (PATOLIS) Patent file (PATOLIS)

Claims (1)

(57) [Claims] 1. An unknown input signal x (k) and an output signal y
A method from the (k) to estimate the output of the unknown system, the autocorrelation r _ij of the input signal x to the unknown system (k) (k) (0 ≦
i, j ≦ p−1 ), convolving the intermediate variable z (k) with the input signal x (k) and the autocorrelation value r
_From the sum of _ij (k) and the inner product of the smoothing prefilter coefficients, an estimated signal y ＾ (k) of the output of the unknown system is obtained, and the output signal y (k) of the unknown system and the estimated output y ＾ ( k), a pre-filter coefficient β (k) is calculated using the difference signal e (k) and the autocorrelation value r _ij (k), and the pre-filter coefficient β with (k), I determined the smoothing pre-filter coefficients s _i (k), output by the high speed projection method for updating the intermediate variable z (k) using the smoothing pre-filter coefficients s _i (k) In the estimation method, a coefficient update stop signal is generated based on the input signal x (k) and the estimation error e (k), and the pre-filter coefficient β
(k) is set to 0, the input signal x (k) and the filter coefficient h ＾ (k) are convolved by the second convolution means, and the difference signal between the output of the second convolution means and the output signal of the unknown system e _f (k) is calculated, and the difference signal e _f (k) is compared with the difference signal e (k).
If (k) is small, the pre-filter coefficient β is set to 0 at time p or more, and then the intermediate variable z (k) is set in the second convolution means as the filter coefficient h ＾ (k). <br/> suitable応的unknown system output estimate how to.