JPH0448005B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention eliminates echo signals caused by impedance mismatch between a two-wire line and a four-wire line in a telephone line, thereby improving speech quality on a long-distance telephone line. This relates to echo cancellers used in

Configuration of conventional examples and their problems Conventionally, as adaptive estimation algorithms for echo cancellers, those based on learning identification methods have been widely used.

The echo canceller based on the learning identification method can obtain good convergence characteristics when uncorrelated signals such as white noise signals are input, but it is difficult to obtain signals with strong correlation such as speech signals. There was a problem that it took a long time to converge when inputting it.

As an echo canceller for solving these problems, an echo canceller according to the method described in Japanese Patent Publication No. 58-31129 has been proposed.

Figure 1 shows its configuration. In FIG. 1, 1 is a receiving side input terminal, 2 is a receiving side output terminal, 3 is a transmitting side input terminal, 4 is a transmitting side output terminal, 5 is a prediction coefficient calculator, 6 is a first predictor, and 7 is a first subtractor, 8
is the first register, 9 is the second predictor, 10 is the second subtractor, 11 is the second register, 12 is the first
convolution operator, 13 is the third subtractor, 14
is a corrector, 15 is a third register, 16 is a second convolution operator, and 17 is a 44th subtractor.

The operation and problems of the echo canceller configured as described above will be explained below.

In the following explanation, the signal inside the echo canceller will be treated as a discrete time signal sampled with respect to time, but the sampler and hold circuit required for this are well known and will be omitted in the following explanation. .

Now, assume that the receiving side input signal at time j is x _j , the transmitting side input signal is y _j , and the transmitting side output signal is e _j . Also, the signal sequence X _j of the input signal on the receiving side is expressed by equation (1), and the estimated impulse response H^ of the echo path is expressed by equation (2).

X _j = (x _j , x _j-1 , ..., x _j-(N-1) )′ ...(1) H^ _j = (h^ _0j , h^ _1j , ..., h^ _N-1j )′...(2) Here, N represents the number of samples of the impulse response, and the symbol ′ represents the transposition of the vector.

When a transmitting side input signal of a predetermined time length is input, the prediction coefficient calculator 5 regards the receiving side input signal as an output signal of an M-order autoregressive model and calculates its linear prediction coefficient a _i (i= 1, 2,...,M). The algorithm for determining the linear prediction coefficient is, for example, Durbin's method. For more information on Durbin's method, see the following article (LRRabiner and RWSchafer, Digital
Processing of Speech Signals, Prentice−
Hall, Inc., 1978).

That is, assuming that the linear prediction coefficient in the interval with the number of samples L is constant, and defining the i-th short-time autocorrelation function R(i) using equation (3), Durbin's algorithm can be calculated from equation (4) as follows: 9) is given by Eq.

R(i)= _L-1-i 〓 ^m=0 x _l+n x _l+n+i (0iM) ……(3) E ^(o) = R(o) ……(4) k _i =( R(i)− _i-1 〓 ^m=1 a _n ^(i-1) R(i-m))/ E ^(i-1) (1iM) ...(5) a ⁽ⁱ⁾ _i =k _i ... …(6) a ⁽ⁱ⁾ _n = a _n ^(i-1) −k _i a ^(i-1) _in (1mi−1) …(7) E ⁽ⁱ⁾ = (1−k ² _i )E ^(i-1) ……(8) a _n =a ^(M) _n (1mM) ……(9) That is, after relating the short-term autocorrelation function up to the Mth order using equation (3), (4 ) by recursively calculating equations (5) to (8) as i = 1, 2, ..., M using equation (5) as an initial condition, and finally obtain the linear prediction coefficient from equation (9). be able to.

The first predictor 6 uses the linear prediction coefficients obtained as described above and the receiving side input signal to create a predicted value x ^p _j of the receiving side input signal according to equation (10).

x ^p _j = _M 〓 ⁱ⁼¹ a _i x _j-1 ...(10) The first subtractor 7 subtracts the predicted value of the receiving side input signal from the receiving side input signal as shown in equation (11). By doing this, a residual signal x〓 _j of the input signal on the receiving side is created.

x〓=x _j −x ^p _j ……(11) The first register 8 stores the signal sequence x〓 of the residual signal of the receiving side input signal obtained in the above manner as (12)
Store it in the form shown in the formula.

x〓=(x〓 _j , x〓 _j-1 , ...x〓 _j-(N-1) )' ...(12) In the second predictor 9, the linear Using the prediction coefficient and the input signal on the transmitter side, the predicted value of the input signal on the transmitter side is calculated according to equation (13).
y ^p _j , and the second subtractor 10 subtracts the predicted value of the transmitting side input signal from the transmitting side input signal as shown in equation (14). Create the residual signal y〓 _j .

y ^p _j = _M 〓 ⁱ⁼¹ a _i y _j-1 ...(13) y〓 _j =y _j −y ^p _j ...(14) Next, the estimated impulse response is expressed in the form shown in equation (2). The stored contents of the second register and the contents of the first register 8 are convolved as shown in equation (15) to obtain the estimated value y〓 _j of the residual signal of the input signal on the transmitting side.

y〓 _j = H^′ _j An estimated error signal e〓 _j of the residual signal of the input signal is created.

e〓=y〓 _j −y〓 _j ...(16) The corrector 14 corrects the contents of the second register 11, that is, the estimated impulse response, according to the algorithm of the learning identification method shown in equation (17). .

H^ _j+1 = H^ _j + aX〓je〓 _j ／‖X〓j‖ ² ...(17) Here, a is a constant of 0<a<2, and ‖X〓j‖
represents the Euclidean norm of X〓j.

The second convolution calculator 16 stores the contents of the second register 11 and the signal sequence X _j of the receiving side input signal in the form shown in equation (1), as shown in equation (18). The contents of the register 15 of 3 are convolved to create an estimated echo signal y^ _j .

y^ _{j =} H^′ _j , transmission side output terminal 4
Output to.

e _j =y _j −y^ _j (19) Note that the receiving side input signal is output as is to the receiving side output terminal 2.

The above is the operation of the echo canceller configured as shown in Figure 1. By reducing the correlation of the receiving side input signals based on the autoregressive model, it has good convergence characteristics even for receiving side input signals with strong correlation. can be obtained.

However, the echo canceller having the configuration shown in FIG. 1 has a problem in that the amount of calculation and memory increases compared to the echo canceller using the normal learning identification method.

The amount of calculation and memory amount will be specifically evaluated below. First, the amount of calculation is evaluated by the amount of multiplication per sampling period. In the echo canceller configured as shown in Fig. 1, it is assumed that there is no overlap in the interval in which the short-time autocorrelation function is calculated, and the number of samples in the interval is L.
is equal to N, the amount of multiplication per sample period is approximately (3N+4M+M ² /N). On the other hand, the amount of multiplication per sampling period of the echo canceller using the normal learning identification method is approximately 2N. Here, assuming N = 320 and M = 5, we calculate the ratio of the multiplication amount of the echo canceller with the configuration shown in Figure 1 to the echo canceler using the normal learning identification method.
It becomes 1.53. That is, in the echo canceller having the configuration shown in FIG. 1, by introducing the autoregressive model, it is necessary to improve the computational power by about 50% compared to the case where the model is not introduced.

Next, we will specifically evaluate the amount of memory. 1st
In the echo canceler having the configuration shown in the figure, three registers of length N occupy most of the required memory. On the other hand, an echo canceller using the normal learning identification method requires two registers of length N. Therefore, in the echo canceller with the configuration shown in Figure 1, by introducing an autoregressive model,
It is necessary to increase the amount of memory by approximately 50% compared to the case where the above model is not introduced.

As described above, the echo canceller with the configuration shown in Figure 1 achieves good convergence characteristics even for highly correlated receiving side input signals by reducing the correlation of the receiving side input signals based on the autoregressive model. be able to. However, there was a problem in that the amount of calculation and memory had to be increased by introducing the autoregressive model.

Purpose of the Invention In view of the problems of the conventional example described above, the present invention provides good convergence characteristics even for receiving side input signals with strong correlations such as audio signals, without increasing the amount of calculation and memory. This provides an echo canceller that shows the

Composition of the Invention The present invention provides a prediction coefficient calculator that calculates a linear prediction coefficient for a receiving side input signal having a predetermined time length by regarding the receiving side input signal as an output signal of an autoregressive model; a first predictor that creates a predicted value of a receiving side input signal using an input signal and the linear prediction coefficient; and a first predictor that generates a predicted value of a receiving side input signal by using an input signal and the linear prediction coefficient; a first subtracter that creates a residual signal of the receiving side input signal; a first register that stores the residual signal of the receiving side input signal; a second predictor that creates a predicted value; and a second predictor that creates a residual signal of the transmitter input signal by subtracting the predicted value of the transmitter input signal from the transmitter input signal.
a second register for storing an estimated impulse response of the echo path; and a convolution operation for convolving the contents of the first register and the second register to create an estimate of the residual signal of the echo signal. a third subtractor that creates a residual signal of the transmitter output signal by subtracting the estimated value of the residual signal of the echo signal from the residual signal of the transmitter input signal; and the first register. a corrector that sequentially corrects the estimated impulse response by correcting the contents of the second register using the contents of the second register and the residual signal of the transmitter output signal;
a third predictor for creating a predicted value of the transmitting side output signal using the transmitting side output signal and the linear prediction coefficient, and adding a residual signal of the transmitting side output signal and a predicted value of the transmitting side output signal. an adder that creates a transmitting side output signal by combining the receiving side input signal and a residual signal of the receiving side input signal, or in place of the prediction coefficient calculator and the first predictor in the above configuration, the receiving side input signal and the residual signal of the receiving side input signal It is composed of an adaptive predictor that sequentially calculates linear prediction coefficients so as to minimize the residual signal of the input signal on the receiving side and creates a predicted value of the input signal on the receiving side. It is an object of the present invention to provide an echo canceller that exhibits good convergence characteristics even for receiving-side input signals having strong correlations such as audio signals without substantially increasing the amount of the echo canceller.

DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 2 shows a configuration diagram of the first embodiment of the present invention. In FIG. 2, 18 is a receiving side input terminal, 19 is a receiving side output terminal, 20 is a transmitting side input terminal, 21 is a transmitting side output terminal, 22 is a prediction coefficient calculator, 23 is a first predictor, and 24 is a 1st subtractor, 25 is the first register, 26 is the second predictor, 27 is the second subtractor, 28 is the second register, 29 is the convolution operator, 30 is the third subtractor , 31 is a corrector, 32 is a third predictor, and 33 is an adder.

The operation of the echo canceller configured as described above will be explained below.

When a reception-side input signal of a predetermined time length is input, the prediction coefficient calculator 22 regards the reception-side input signal as an output signal of an M-th order autoregressive model and calculates its linear prediction coefficient a _i =(i =1, 2,...,M). The algorithm to find the linear prediction coefficient is
For example, according to the method of Durbin mentioned above. After calculating the linear prediction coefficient, the linear prediction coefficient calculator uses the linear prediction coefficient to determine the stability of the regression filter, and if it is determined to be unstable, calculates the value of the linear prediction coefficient. Switch to a stable value. In order to operate the recursive filter stably, all the poles must exist within the unit circle on the Z plane. Therefore, stable operation of the regression filter is ensured by limiting the values of the linear prediction coefficients so that all poles are within the unit circle. For example, in the case of M=2, if the values of a ₁ and a ₂ exist within the triangular area shown in FIG. 3, the regression filter operates stably. On the other hand, if the values of a ₁ and a ₂ exist outside the above range, stable operation is ensured by forcibly switching the values of a ₁ and a ₂ to values inside the above range.

FIG. 4 is an internal configuration diagram of the prediction coefficient calculator 22. In FIG. 4, 34 is a receiving side input terminal, 35 is a prediction coefficient output terminal, 36 is a microprocessor, 3
7 is a multiplier, 38 is a bus line, 39 is a program memory, and 40 is a data memory. FIG. 4 shows a configuration centered around a normal microprocessor, and data and addresses are exchanged via a bus line 38. The program memory 39 is usually a read-only memory and stores a program for controlling the operation of the microprocessor 36. Further, the data memory 40 is usually constituted by a random access memory, and is mainly used for storing input signals on the receiving side. When a predetermined number of receiving side input signals are input through the receiving side input terminal 34, the microprocessor 36 uses the built-in arithmetic and logic unit and the external multiplier 37 to calculate (3) from equation (3). 9) Execute the calculation shown in formula to calculate the linear prediction coefficient. Next, the stability of the obtained linear prediction coefficients is determined to ensure stable operation.
The linear prediction coefficients obtained in this manner are sent to the first predictor 23, second predictor 26, and third predictor 32 via the prediction coefficient output terminal 35.

The first predictor 23 uses the linear prediction coefficients obtained as described above and the receiving side input signal to create a predicted value x ^p _j of the receiving side input signal according to equation (20).

x ^p _j = _M 〓 ⁱ⁼¹ a _i x _ji ...(20) The first subtractor 24 subtracts the predicted value of the receiving side input signal from the receiving side input signal as shown in equation (21). Create a residual signal x〓 _j of the receiving side input signal.

x〓 _j = x _j −x ^p _j ...(21) The first register 25 stores the signal sequence X〓j of the residual signal of the receiving side input signal obtained as above.
is stored in the form shown in equation (12).

The second predictor 26 uses the linear prediction coefficient calculated by the prediction coefficient calculator 22 and the transmission side input signal to create a predicted value y ^p _j of the transmission side input signal according to equation (22), Moreover, in the second subtractor 27,
As shown in equation (23), a residual signal y〓 _j of the transmitting side input signal is created by subtracting the predicted value of the transmitting side input signal from the transmitting side input signal.

y ^p _j = _M 〓 ⁱ⁼¹ a _j y _ji ...(22) y〓 _j = y _j −y ^p _j ...(23) Next, the contents of the first register 25 and the estimated impulse of the echo path The contents of the second register 28, which stores the response in the form shown in equation (2), are convolved using the convolution calculator 29 as shown in equation (24), and the residual signal of the echo signal is calculated. Create an estimated value y〓.

y〓 _j = H^′ _j A residual signal e〓 _j of the transmitting side output signal is created by subtracting the estimated value e〓 _j = y〓 _j −y〓 _j ...(25) The corrector 31 uses the learning identification method shown in equation (26). According to the algorithm, the first register 25
The estimated impulse response is successively corrected by correcting the contents of the second register 28 using the contents of the second register 28 and the residual signal of the transmitting side output signal.

H^ _j+1 = H^ _j + ae〓 _j e〓 _j /‖X〓j‖ ² ... (26) where a is a constant of 0<a<2.

The third prediction coefficient calculator 32 uses the linear prediction coefficient obtained by the prediction coefficient calculator 22 and the transmission side output signal to calculate the predicted value e ^p _j of the transmission side output signal as shown in equation (27). Further, the adder 33 creates a transmitting side output signal e _j by adding the residual signal of the transmitting side output signal and the predicted value of the transmitting side output signal as shown in equation (28).

e ^p _j = _M 〓 ⁱ⁼¹ a _i e _ji ... (27) e _j = e ^p _j + e〓 ... (28) Fig. 5 shows an internal configuration diagram of the predictor. In FIG. 5, 41 is a signal input terminal, 42 is a prediction coefficient input terminal, 43 is a signal prediction value output terminal, 44 is a register for signal storage, 45 is a register for prediction coefficient storage, and 46 is a convolution calculator. be. Next, its operation will be explained. The signal of the past M samples input from the signal input terminal 41 is stored in the register 44 for signal storage, and the prediction coefficient input terminal 4
The M linear prediction coefficients input from 2 are stored in a register 45 for storing linear prediction coefficients. Next, convolve the contents of the above two registers into the convolution operator 4.
6 to create a predicted value of the signal,
It is output from the signal predicted value output terminal 43.

The first predictor 23, the first subtractor 24, and the second
The predictor 26 and the second subtractor 27 each constitute a non-regressive filter, and the third predictor 32
and adder 33 constitute a regression filter.
Further, the non-recursive filter and the recursive filter have an inverse filter relationship with each other. Therefore, as explained in the operation of the prediction coefficient calculator 22, it is necessary to determine the stability of the linear prediction coefficients.

In general, when a filter is inserted, a signal delay occurs, so depending on the insertion position of the filter, a delay may occur in the adaptive control and the operation may become unstable. 26) Since there are no filters in the loop necessary to calculate the equation, and all filters exist outside the loop, unstable operation due to adaptive control delays does not occur.

The above is the explanation of the operation of the first embodiment.

Next, the amount of calculation in the first embodiment is evaluated under the same conditions as when evaluating the amount of calculation in the conventional example. 1st
The amount of multiplication per sample period in the embodiment is approximately (2N+5M+M ² /N). Therefore, N=
320, M=5, and the ratio of the multiplication amount of the echo canceller of the first embodiment to the echo canceller of the normal learning identification method is 1.04.
That is, the echo canceller of the first embodiment is
It can be realized with approximately the same amount of calculation as an echo canceller using a normal learning identification method.

Next, the amount of memory required to implement the first embodiment will be evaluated. In the first embodiment, the length N is 2
A total of 3N memories, ie, 3N registers and a memory for storing the N receiver input signals in the prediction coefficient calculator 22, constitutes the majority of the memory required. Therefore, the first embodiment requires approximately the same amount of memory as the conventional example shown in FIG.

As described above, according to the echo canceller of the first embodiment, the amount of calculation is hardly increased, and
By providing a means to reduce the correlation of transmitter input signals based on an autoregressive model, we provide an echo canceller that exhibits good convergence characteristics even for receiver input signals with strong correlations such as audio signals. I can do that.

A second embodiment of the present invention will be described below with reference to the drawings. FIG. 6 shows a configuration diagram of an echo canceller according to a second embodiment of the present invention.
In FIG. 6, 47 is a receiving side input terminal, 48 is a receiving side output terminal, 49 is a transmitting side input terminal, 50 is a transmitting side output terminal, 51 is a first subtracter, 52 is an adaptive predictor, and 53 is a first subtracter. 1 register, 54 is the first predictor, 55 is the second subtractor, 56 is the second register, 57 is the convolution operator, 58 is the third subtractor, 59 is the corrector, 60 is the Predictor of 2, 61
is an adder.

The structure of the second embodiment shown in FIG. 6 is the same as the structure of the first embodiment shown in FIG. 2, except for the following points. That is, in place of the prediction coefficient calculator 22 and first predictor 23 of the first embodiment, the adaptive predictor 52 is used in the second embodiment.

Comparing the operating principles of the first embodiment and the second embodiment, it is found that both are the same in that they attempt to decorrelate the receiving side input signal based on the linear prediction method, but the calculation of the linear prediction coefficient Laws are different. That is,
In the first embodiment, a receiving side input signal of a predetermined time length is regarded as an output signal of an autoregressive model with a constant linear prediction coefficient, and the linear prediction coefficient is calculated by the prediction system calculator 22. In the second embodiment, the adaptive predictor 52 is used to obtain linear prediction coefficients by sequentially performing estimation for each sample period so as to minimize the residual signal of the receiving side input signal, which is the prediction error signal. They are different in some respects.

Hereinafter, a description will be given of the difference between the operation of the second embodiment and the first embodiment, that is, the operation of the adaptive predictor 52.

The adaptive predictor 52 performs linear prediction using the receiving side input signal and the residual signal of the receiving side input signal created by the first subtractor 51 so as to minimize the residual signal of the receiving side input signal. The coefficients are sequentially calculated and a predicted value of the input signal on the receiving side is created.

FIG. 7 shows an internal configuration diagram of the adaptive predictor. 7th
In the figure, 62 is a signal input terminal, 63 is a signal prediction value output terminal, 64 is a prediction coefficient output terminal, 65 is an error signal input terminal, 66 is a register for signal storage, and 67
is a register for storing prediction coefficients, 68 is a convolution operator, and 69 is a corrector. In FIG. 7, the receiving side input signal input from the signal input terminal 62 is stored in a signal storage register 66 that stores signals of past M samples. The convolution calculator 68 convolves the contents of the prediction coefficient storage register 67 that stores M linear prediction coefficients with the contents of the signal storage register 66 as shown in equation (29), Create the predicted value x ^p _j of the receiving side input signal,
It is output to the signal predicted value output terminal 63.

x ^p _j = _M 〓 ⁱ⁼¹ a _i x _ji ...(29) In the corrector 69, as shown in equation (30), the contents of the register 66 for signal storage and the contents of the register 67 for storing prediction coefficients are and a residual signal of the receiving side input signal which is an error signal of the predicted value inputted from the error signal input terminal 65 to minimize the residual signal of the receiving side input signal. The contents of 67 are corrected and the values of the linear prediction coefficients are successively corrected.

In equation (30), a _ij is the linear prediction coefficient a _i at time j
, and β is a constant of 0<β<2.

Next, the linear prediction coefficients calculated as described above are output to the prediction coefficient output terminal 64.

Here, the learning identification method is used as an algorithm for sequentially estimating linear prediction coefficients, but other algorithms such as steepest descent method may also be used. 6th
As is clear from a comparison between Figure 7 and Figure 7, as well as equations (26) and (30), the internal configuration of the adaptive predictor is for estimating the impulse response of the echo canceller itself and estimating the echo signal accordingly. Therefore, the second embodiment has the advantage that common hardware and software can be used for both. Furthermore, in the first embodiment, a memory for storing a predetermined number of receiving side input signals is required inside the prediction coefficient calculator 22, whereas in the second embodiment using an adaptive predictor, sequential estimation is required. Since such memory is used, such memory is no longer necessary, and the amount of memory can be saved.

Next, the calculation amount and memory amount of the second example will be evaluated under the same conditions as the first example. The amount of multiplication per sample period in the second embodiment is approximately (2N+4M)
Therefore, the echo canceller according to the second embodiment can be realized with approximately the same amount of calculation as the echo canceller using the normal learning identification method. Furthermore, 2 of length N occupies most of the memory required to realize the second embodiment.
Therefore, the memory capacity can be realized with approximately the same amount of memory as an echo canceler using a normal learning identification method.

FIG. 8 shows an example of experimental results regarding the convergence characteristics of a prototype echo canceler for a pseudo voice signal specified by CCITT's FEC G.227. The echo canceller used in the experiment had the configuration according to the second embodiment of the present invention, and by changing the value of M, a comparison between the echo canceler according to the present invention and the echo canceler using the ordinary learning identification method was made. I did it. The main parameters of the echo canceller used in the experiment are sampling frequency =
8KHz, N=320, a=1. In FIG. 8, the solid line indicates the echo canceller (M=2) according to the present invention.
The broken line indicates the convergence characteristic of the echo canceller (M=0) using the normal learning identification method. It can be seen from FIG. 8 that the echo canceller according to the present invention exhibits good convergence characteristics for pseudo-speech signals having autocorrelation.

As described above, according to the second embodiment, by providing a means for reducing the correlation of receiving side input signals based on adaptive prediction without substantially increasing the amount of calculation and memory, strong It is possible to provide an echo canceller that exhibits good convergence characteristics even for correlated receiving side input signals.

In the first and second embodiments, the learning identification method was used as the impulse response sequential estimation algorithm, but other algorithms (for example, the steepest descent method) may be used.

Effects of the Invention As described above, according to the present invention, by providing a means for reducing the correlation caused by the autoregressive model without substantially increasing the amount of calculation, the receiving side input signal having a strong correlation such as an audio signal can be It is possible to provide an echo canceller that exhibits good convergence characteristics even for

Furthermore, the method based on adaptive prediction can be realized with almost no increase in memory capacity.

In those based on an adaptive predictor, it is also possible to make the configuration and operating algorithm of the adaptive predictor the same as the configuration and algorithm for sequential estimation of the impulse response of the echo canceller itself, and in such a case, It has the advantage that common hardware and software can be used.

Furthermore, the echo canceller according to the present invention is also effective in other cases where transmission system estimation is performed using highly correlated signals such as audio signals.
For example, howling prevention in loudspeaker telephones,
It can also be applied to prevention of singing in forwarding telephone systems.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a conventional echo canceller, FIG. 2 is a block diagram of an echo canceller in a first embodiment of the present invention, and FIG. 3 is a diagram showing a stable region of a second-order linear prediction coefficient. FIG. 4 is an internal configuration diagram of the prediction coefficient calculator of the first embodiment, FIG. 5 is an internal configuration diagram of the predicted value, and FIG. 6 is a configuration diagram of the echo canceller in the second embodiment. FIG. 7 is an internal configuration diagram of the adaptive predictor, and FIG. 8 is a diagram showing experimental results regarding the convergence characteristics of the prototype echo canceller for pseudo voice signals. 18... Receiving side input terminal, 19... Receiving side output terminal, 20... Transmitting side input terminal, 21... Transmitting side output terminal, 22... Prediction coefficient calculator, 23... First predictor, 24 ...first subtractor, 25...first register, 26...second predictor, 27...
Second subtractor, 28... Second register, 29...
...Convolution operator, 30...Third subtractor, 3
1... Corrector, 32... Third predictor, 33...
Adder, 34... Receiving side input terminal, 35... Prediction coefficient output terminal, 36... Microprocessor, 3
7... Multiplier, 38... Bus line, 39... Program memory, 40... Data memory, 41...
...Signal input terminal, 42...Prediction coefficient input terminal, 4
3... Signal prediction value output terminal, 44... Register for signal storage, 45... Register for prediction coefficient storage, 46... Convolution operator, 47... Receiving side input terminal, 48... Receiving side output Terminal, 49...Sending side input terminal, 50...Sending side output terminal, 51...
...first subtractor, 52...adaptive predictor, 53...
First register, 54...First predictor, 55...
...Second subtractor, 56...Second register, 57
...Convolution operator, 58...Third subtractor,
59... corrector, 60... second predictor, 61...
... Adder, 62 ... Signal input terminal, 63 ... Signal predicted value output terminal, 64 ... Prediction coefficient output terminal, 6
5...Error signal input terminal, 66...Register for signal storage, 67...Register for prediction coefficient storage,
68...Convolution operator, 69...Corrector.

Claims (1)

[Scope of Claims] 1. A prediction coefficient calculator that calculates a linear prediction coefficient for a receiving side input signal having a predetermined time length by regarding the receiving side input signal as an output signal of an autoregressive model; a first predictor that creates a predicted value of a receiving side input signal using an input signal and the linear prediction coefficient; and a first predictor that generates a predicted value of a receiving side input signal by using an input signal and the linear prediction coefficient; a first subtracter that creates a residual signal of the receiving side input signal; a first register that stores the residual signal of the receiving side input signal; a second predictor for creating a predicted value; a second subtractor for creating a residual signal of the transmitter input signal by subtracting the predicted value of the transmitter input signal from the transmitter input signal; a second register that stores an estimated impulse response; a convolution calculator that convolves the contents of the first register and the second register to create an estimated value of the residual signal of the echo signal; and the transmitting side input. a third subtractor that creates a residual signal of the transmitter output signal by subtracting an estimated value of the residual signal of the echo signal from the residual signal of the signal; a corrector that sequentially corrects the estimated impulse response by correcting the content of the second register using the content and the residual signal of the transmitting side output signal; and the transmitting side output signal and the linear prediction coefficient. a third predictor for generating a predicted value of the transmitter output signal;
An echo canceller comprising: an adder that creates a transmission side output signal by adding a residual signal of the transmission side output signal and a predicted value of the transmission side output signal. 2. A first subtractor that creates a residual signal of a receiving side input signal by subtracting a predicted value of the receiving side input signal from the receiving side input signal, and a residual signal of the receiving side input signal and the receiving side input signal. an adaptive predictor that successively calculates linear prediction coefficients so as to minimize the residual signal of the input signal on the reception side and creates a predicted value of the input signal on the reception side; a first register that stores a residual signal; and a first register that creates a predicted value of the transmitting side input signal using the transmitting side input signal and the linear prediction coefficient.
a second subtractor for creating a residual signal of the transmitter input signal by subtracting the predicted value of the transmitter input signal from the transmitter input signal; and a second subtractor for storing the estimated impulse response of the echo path. a convolution calculator that convolves the contents of the first register and the second register to create an estimated value of the residual signal of the echo signal; a third subtractor for creating a residual signal of the transmitting side output signal by subtracting an estimated value of the residual signal of the echo signal; and the contents of the first register and the second register and the transmitting side output signal. a corrector that sequentially corrects the estimated impulse response by correcting the contents of the second register using the residual signal; and predicting the transmitting side output signal using the transmitting side output signal and the linear prediction coefficient. A second predictor that creates a value, and an adder that creates a transmitter output signal by adding the residual signal of the transmitter output signal and the predicted value of the transmitter output signal. Echo canceller featuring.