JPH0946277A - Echo canceller - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an echo canceller capable of removing an echo quickly and high-accurately. SOLUTION: The echo canceller is provided with an adaptive filter 4 subtracting an echo estimate from a transmission signal and a nonlinear processor 7 removing a remaining echo included in the output signal from the adaptive filter. The correction coefficient of the coefficient update of the adaptive filter 4 is controlled by a coefficient controller 6 based on a transmitted signal inputted from the adaptive filter and the output signal of the adaptive filter. The correction coefficient of the coefficient update of the adaptive filter is controlled in compliance with the state of the adaptive filter so that a probable accuracy or a converging speed at the adaptive filter can be given preference in confomity with the state of the adaptive filter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an echo canceller used for telephone communication or the like, and more particularly to an echo canceller having improved echo suppression performance.

[0002]

2. Description of the Related Art In a telephone line composed of a transmission line having two subscriber lines and four relay lines, a hybrid is introduced at a connecting portion between the subscriber line and the relay line to achieve impedance matching. However, if this impedance matching is incomplete, a so-called echo occurs in which a part of the received signal leaks to the transmitting side through the 4-wire line. Echo canceller
The operation of suppressing such echo is performed.

In recent years, with the progress of digital signal processing technology, higher performance echo cancellers have been developed.

A conventional echo canceller will be described below.

FIG. 18 shows an example of use of a conventional echo canceller for telephone lines. The digital reception signal Rin (n) is input and the digital transmission signal Sout is input.
(N) The echo canceller 205, the DA converter 204 that converts the digital signal Rout (n) output from the echo canceller 205 into the analog signal ARout, and the hybrid 202 that achieves impedance matching between two lines and four lines
And the analog transmission signal ASin to the digital signal Sin
The AD converter 203 for converting into (n) is a device on the relay station side, and the 2-wire type telephone 200 is connected to the hybrid 202 via the 2-wire type transmission line 201.

The echo canceller 205 is included in the transmission signal Sout (n) based on the reception signal Rin (n) and the double talk determination unit 206 for detecting a state (double talk state) of the transmission signal from the telephone 200. An adaptive filter 207 that removes echo, and a non-linear processor (NLP) 20 that suppresses residual echo that cannot be removed by the adaptive filter 207.
8 and.

The echo canceller 205 outputs the digital reception signal Rin (n) as it is, and the digital reception signal Rout (n) is converted by the DA converter 204 into an analog reception signal ARout. The reception signal ARout is output to the two-wire type transmission line 201 by the hybrid 202, and the telephone 200 receives this signal. Further, the signal transmitted from the telephone 200 is output to the two-wire type transmission line 201, and the hybrid 202 outputs this signal as a transmission signal ASin.
Two-wire transmission line 201 between the telephone 200 and the hybrid 202
Transmits a bidirectional signal of a reception signal and a transmission signal. The hybrid 202 converts a 2-wire type signal and a 4-wire type signal. The output ASin of the hybrid 202 is converted into a digital transmission signal Sin (n) by the AD converter 203 and input to the echo canceller 205.

However, since the performance of the hybrid 202 is not sufficient, part of the received signal ARout leaks to the output ASin of the hybrid 202, and this becomes the echo signal Aecho.
The echo canceller 205 transmits this echo signal to the transmission signal Si.
acts to remove from n.

The echo canceller 205 estimates the impulse response of the echo path in a state where there is no signal transmitted from the telephone 200 and only the echo from the Rout signal is present in the Sin signal (single talk state). Double talk judge
206 detects a double talk state in which there is a transmission signal from telephone 200. In the double talk state, c (n) = 0 In the single talk state, the signal of c (n) = 1 is output to the adaptive filter 207, and the adaptive filter 207 outputs the signal of the echo path when c (n) = 1. Estimate the impulse response.

As shown in FIG. 19, the adaptive filter 207 subtracts the echo estimation value y (n) from the signal d (n) and outputs a signal e (n), and a signal e (n). A multiplier 229 that multiplies the correction coefficient μ (n) with the input signal x
Delay units 219 to 221 for sequentially delaying (n) and a multiplier 229
Output and the past input signal x output from the delay devices 219 to 221
(N), x (n-1), x (n-N + 2), x (n-N)
+1) and multipliers 212, 214, 218 and a multiplier 21
Estimated value hi (n) (0 ≦ i <N) of the echo response of the echo path updated using the outputs of 2, 214, and 218 and the delay device
Multipliers 211, 213, 215 and 217 for multiplying the past input signals x (n-i) output from 219 to 221 and the respective multipliers 211 and 21.
An adder 210 for adding the outputs of 3, 215 and 217 to output an echo estimated value y (n) and a square of the outputs of the delay devices 219 to 221 to obtain the power of the input signal x (n) 2 Multiplier calculator
222 to 225, an adder 226 that adds the outputs of the square calculators 222 to 225, and a divider 22 that divides the constant α by the output of the adder 226.
7 and a multiplier for multiplying the output of the divider 227 by the output c (n) of the double talk determiner 206 and outputting a correction coefficient μ (n)
228 and. Note that n represents the number of signal samples, and N represents the number of taps of the response filter 207.

In this adaptive filter 207, the estimated value y (n) of the echo is equal to the estimated value hi (n) of the impulse response and x.
It is calculated as follows by the convolution operation with (n−i). y (n) = Σ hi (n) × x (n−i) (where Σ adds i from 0 to N−1) The subtractor 209 calculates the echo estimated value y (n) from the signal d (n).
Is subtracted, and the echo component is canceled by the following equation. e (n) = d (n) -y (n) Further, the estimated value hi (n) of the impulse response is updated as follows.

Hi (n + 1) = hi (n) + μ (n)
× e (n) × x (n−i) (0 ≦ i <N) The correction coefficient μ (n) is calculated by the following equation.

Μ (n) = {α × c (n)} / {Σ x
(N−i) × x (n−i)} (where Σ adds i from 0 to N−1) where α is a constant and 0 <α <2. Further, in the double talk state, c (n) = 0 and the estimated value hi (n) of the impulse response is held.

Hi (n + 1) = hi (n) (0 ≦ i <N).

Thus, the adaptive filter 207 removes the estimated value y (n) of the echo from the input signal d (n), and outputs the signal e (n) in which the echo component is suppressed. However, in reality, residual echo remains in e (n) for the following reason. (1) The convergence of the estimated value of the impulse response is not sufficient. (2) The transfer function of the echo path is not linear. (3) There is quantization noise in the quantizer.

The NLP 208 functions to remove the residual echo component of this e (n). The conventional NLP208 is
As shown in FIG. 20, it has an input / output characteristic of a center clip function. This NLP208 has a low level signal a
Is input, the output is clipped and the output Sout (n) becomes 0. On the other hand, when the signal b having a high level is input, it is output in the form of the signal c. Normally, the level of residual echo is very small compared to the level of the signal transmitted from the telephone, and
208 sets the output to 0 when this minute input signal,
The residual echo is removed.

FIG. 21 shows a model of the echo canceller. The echo obtained by multiplying the reception signal by the transfer function 230 of the echo path enters the transmission side, and the adder 231 adds the transmission signal and the echo. The adder 232 further adds noise n (n) to the output of the adder 231, and outputs a signal s (n).

The echo canceller 233 receives the received signal x
(N) is multiplied by the estimated value 234 of the transfer function (impulse response) of the echo path to obtain the estimated value y (n) of the echo signal, and the subtractor 235 subtracts y (n) from the signal s (n). To output e (n), and the NLP 236 removes the residual echo of e (n). The echo canceller 233 can estimate the echo path only in a single-talk state in which s (n) is an echo signal only when there is no transmission signal.

In this way, the echo canceller prevents the echo of the received signal from flowing back.

[0020]

However, the conventional echo canceller has the following problems.

(1) The operation of the adaptive filter shown in FIG. 19 changes as follows depending on how to select the value of the constant α. When α is large, the estimated impulse response value h
The convergence of i (n) is fast, but the estimation accuracy is poor. Therefore, the residual echo remains in e (n) even in the converged state. Conversely, when α is small, the estimation accuracy of the impulse response estimation value hi (n) is improved, but the convergence is delayed.
In the conventional adaptive filter 207, the estimation accuracy and the convergence speed of the adaptive filter are determined by the selection of the constant α, and there is a problem that the characteristics of the adaptive filter cannot be changed according to the situation.

(2) NLP20 of conventional echo canceller
As is clear from the example of FIG. 20, No. 8 has a problem that the transmission signal transmitted from the telephone is distorted because the input / output characteristic is non-linear.

(3) In the conventional echo canceller, the estimation value of the impulse response of the adaptive filter is significantly different from the impulse response of the echo path when the determination error of the double talk determiner 206 or the change of the echo path occurs. It may end up. In such a case, the echo estimated value y (n) of the adaptive filter may be significantly deviated from the echo signal, and noise may be added to e (n).

The present invention solves the above-mentioned problems of the prior art. Echoes can be removed quickly and with high accuracy, and the deterioration of the quality of the transmitted signal can be prevented.
Moreover, it aims at providing the echo canceller which can respond appropriately also at the time of abnormality.

[0025]

Therefore, in the present invention, in an echo canceller device comprising an adaptive filter for subtracting an echo estimation value from a transmission signal and a non-linear processor for removing residual echo contained in an output signal of the adaptive filter, The correction coefficient for updating the coefficient of the adaptive filter is controlled based on the transmission signal input to the adaptive filter and the output signal of the adaptive filter.

When the SN ratio of the transmission signal is small,
The correction coefficient is controlled to be small. When the SN ratio is large, the correction coefficient is controlled to be large when the output signal tends to increase, and the correction coefficient is controlled when the output signal tends to decrease. It is controlled to be small.

The non-linear processor is composed of a multiplier for inputting the output signal of the adaptive filter and a gain control means for determining the gain of the multiplier by fuzzy inference.

Further, a control means for controlling the subtraction amount of the echo estimated value from the transmission signal is provided, and the subtraction amount is suppressed when an abnormal state occurs in the adaptive filter.

Further, the control means is configured to control the subtraction amount based on the magnitude of the power of the transmission signal and the power of the signal obtained by subtracting the echo estimation value from the transmission signal. In addition, a plurality of data tables that hold the control amount as a function of the difference between these respective powers or the ratio of each power and the control amount to be applied from the control amount read from this data table It is set by means of selecting.

[0030]

By controlling the correction coefficient for updating the coefficient of the adaptive filter according to the state of the adaptive filter, it is possible to give priority to either the estimation accuracy or the convergence speed of the applied filter in accordance with the state of the adaptive filter.

When the SN ratio of the transmission signal is poor, the estimation will include many errors. Therefore, the correction coefficient is made small to suppress the update amount of the estimated value of the impulse response. In the state where the SN ratio of the transmission signal is good, when the output signal tends to increase, that is, when the estimation does not work well, the correction coefficient is increased to increase the update amount of the estimated value of the impulse response. When the output signal tends to decrease, the correction coefficient is reduced to reduce the convergence speed and increase the estimation accuracy.

Further, in the apparatus in which the non-linear processor is composed of the multiplier and the gain control means using fuzzy inference, when there is no transmission signal from the telephone, the gain of the multiplier is reduced to suppress the residual echo, and When there is a transmission signal from the telephone, the gain of the multiplier is set to 1 and the transmission signal is transmitted as it is without being distorted.

Further, in the device provided with the control means for controlling the amount of subtraction of the echo estimated value from the transmission signal, when the echo component cannot be suppressed due to the deviation of the impulse response estimated value, the estimated value of the echo in the adaptive filter is changed. The subtraction is suppressed and the addition of noise to the output signal is avoided.

[0034]

【Example】

(First Embodiment) In the echo canceller of the first embodiment, the correction coefficient μ (n) is increased and h is increased in a state where the estimated value hi (n) of the impulse response needs to be quickly converged.
If i (n) does not need to converge quickly, μ
Μ (n) is controlled so that (n) is reduced.

The adaptive filter of this echo canceller is
As shown in FIG. 4, a multiplier 38 for multiplying the output of the divider 37 by a coefficient k is provided. Other configurations are the same as those of the adaptive filter of the conventional echo canceller (FIG. 19). This coefficient k is variable according to the state of the echo canceller. As a result, the correction coefficient μ (n) dynamically changes.

As shown in FIG. 1, this echo canceller has a signal sin containing echo and noise as a whole.
A power calculator 1 for calculating the power SP of the signal, a double talk determiner 3 for determining double talk, an adaptive filter 4 having the configuration shown in FIG. 4, and an output e of the adaptive filter 4.
A computer 2 for calculating the power EP of (n), a noise power estimator 5 for calculating the power NP of the noise component included in e (n), and a coefficient k given to the adaptive filter 4 using SP, EP and NP. A coefficient controller 6 for controlling
NLP7 excluding the residual echo included in (n).

The power calculators 1 and 2 have an absolute value calculator 8 for calculating the absolute value of the input signal, for example, as shown in FIG.
And a multiplier 9 for multiplying the absolute value of the input signal by a fixed value a
A delay device 12 for delaying the output signal, a multiplier 10 for multiplying the delayed output signal by a fixed value 1-a, a multiplier 9,
A known computer including an adder 11 that adds the outputs of 10 and outputs a signal representing the power of the input signal can be used.

The noise power estimator 5 estimates the noise power NP of FIG. 21 from the value of EP in a state other than the double talk state. Therefore, as shown in FIG. 3, the noise power estimator 5 includes a multiplier 1 in the power calculator.
A filter coefficient controller 18 for controlling the coefficients g of 3 and 14, and 1
A filter coefficient controller provided with a subtractor 16 for outputting -g
The output c (n) of the double talk determiner 3 is input to 18.
Then, when c (n) = 1, that is, in the single-talk state, the filter coefficient controller 18 outputs a value close to 1 which is 1 or less as g, and when c (n) = 0, that is, double. In the talk state, a value of 0 or more, which is close to 0, is output as g.

The coefficient controller 6, as shown in FIG.
A divider 41 that divides (n) by NP (n) to calculate the SN ratio (SNR) in the input signal of the echo canceller;
A table reader 43 that reads a value corresponding to the SNR from the data table 42, a delay device 44 that delays the output EP of the power calculator 2, and a subtractor that subtracts the delayed EP from the input EP to calculate a change value of the EP. 45, a data converter 46 for converting data into a value according to the change value, and a coefficient controller 6
47 and a table reader
The minimum value calculator 48 which outputs the smaller one of the output of 43 and the output of the adder 47, and the maximum which outputs the larger one of the output of the minimum value calculator 48 or 0.05 as a coefficient k (n) It has a value calculator 49 and a delay device 50 for delaying the output of the maximum value calculator 49 and inputting it to the adder 47.

In the coefficient controller 6, the divider 41 outputs the signal d
The SNR of (n) is estimated by the following equation. SNR = power of echo component / power of noise component The data table 42 has a shape as shown in FIG.
When the value of is small, the table reader 43 outputs a small value. Therefore, the output k (n) of the coefficient controller 6 is limited to a small value. As a result, when the SNR value is small and the impulse response of the echo path cannot be estimated accurately, the correction coefficient μ (n) of the adaptive filter becomes small, and as a result, the estimated value hi of the impulse response is hi.
The update amount of (n) becomes small.

On the other hand, the delay unit 44 and the subtractor 45 calculate the change value of EP, and the data converter 46, as shown in FIG. 7, shows the positive data when the change value is positive and the negative value when the change value is negative. Output negative data.

When the impulse response is properly estimated, the EP should be gradually reduced in the single talk state because the echo estimated value y (n) is subtracted. When this EP is increasing, it means that the estimation is not working well.

With the EP increasing, the data converter 46
From the adder 47 outputs large positive data
Data for changing (n) to a large value is output. k
When (n) becomes large, the impulse response estimation value hi
The convergence speed of (n) becomes faster.

In the state where EP decreases, the data converter 46 outputs negative data, and the adder 47 outputs k.
Data for changing (n) to a small value is output. k
As (n) becomes smaller, the impulse response estimated value hi
The convergence speed of (n) becomes slower, but the estimation accuracy becomes better.

The minimum value calculator 48 outputs either the output of the table reader 43 or the output of the adder 47, whichever is smaller. Therefore, when the SNR of d (n) is extremely poor,
Regardless of what data is output from the adder 47, the coefficient k (n) can be kept small. On the other hand, when the SNR is good, the output of the adder 47 is output as the coefficient k (n), and as a result, the convergence speed of the impulse response estimation value hi (n) becomes faster and EP increases as EP increases. In the decreasing state, the convergence speed of the impulse response estimation value hi (n) becomes slow.

The maximum value calculator 49 functions to keep the value of k (n) at least 0.05 or more, and k (n) is 0 or less.
This prevents the situation where the estimation does not work.

As described above, the echo canceller of the first embodiment feeds back the state of the echo canceller and controls the value of the correction coefficient μ (n) of the adaptive filter according to this state.

(Second Embodiment) The echo canceller of the second embodiment fuzzy controls a non-linear processor (NLP) in order to remove residual echo.

This echo canceller, as shown in FIG. 8, is a power calculator 51 for calculating the power SP of the signal sin.
, The double talk determiner 53, the adaptive filter 56 having the conventional configuration, and the power E of the output e (n) of the adaptive filter 56.
A computer 52 that calculates P, a noise power estimator 54 that calculates the power NP of the noise component included in e (n), and an E that estimates the echo suppression amount (ERLE) of the adaptive filter 56.
It comprises an RLE estimator 55 and an NLP 57 which performs processing for removing residual echo using EP, NP and ERLE.

The configurations of the power calculators 51 and 52 are as shown in FIG. Further, the configuration of the noise estimator 54 is similar to that of FIG.

The ERLE estimator 55, as shown in FIG.
A divider 58 that divides EP by SP is provided.
Is output to the multiplier 59. Other configurations are the same as those of the noise estimator (FIG. 3). The ERLE estimator 55 estimates the echo suppression amount (ERLE) in the single talk state by the following equation.

ERLE = EP / SP Next, the configuration of the NLP 57 will be described. As shown in FIG. 10, the NLP 57 includes a multiplier 78 that multiplies the output e (n) of the adaptive filter by a coefficient (gain) gsw, and a mechanism that determines the gain gsw by fuzzy inference.

This gain determining mechanism includes a divider 65 for dividing the desired value (TLOSS) of the echo suppression amount by ERLE,
A constant (maximum noise power) EN divided by EP, a divider 66, a divider 67 divided by EP, and a divider 65.
Or a maximum value calculator 68 that outputs the larger one of the outputs of the divider 66, a minimum value calculator 69 that outputs the smaller one of the outputs of the maximum value calculator 68 and 1.0, and a membership function , A table reader 71 for reading the state coincidence degree α ₁ from the data table 70 according to the value of the EP to be input, and an EP for inputting.
Table read unit 73 that reads the state coincidence degree α ₂ from the data table 72 in accordance with the value of, and the state coincidence degree α ₃ that is read from the data table 74 in accordance with the EP / NP value output from the divider 67 The table reader 75, the maximum value calculator 76 that outputs the greater of the coincidence α ₂ output from the table reader 73 or the coincidence α ₃ output from the table reader 75, and the maximum value calculator 76 from each data by the centroid method. And a defuzzifier 77 for calculating the gain gsw.

The NLP 57 suppresses the residual echo by bringing the gain gsw of the multiplier 78 close to 1 in the double-talk state and bringing the gain gsw of the multiplier 78 close to 0 in the single-talk state. This gain gsw is EP, N
It is determined by fuzzy inference using the values of P and ERLE. The minimum value calculator 69 has a gain g in the single talk state.
output on, and the gain goff in the double talk state is 1.
Set to 0. Further, the degree of coincidence αon indicating how much the current state matches the single-talk state is output from the maximum value calculator 76, and the degree of coincidence αoff between the current state and the double-talk state is output from the table reader 71. To be done.
The defuzzifier 77 uses these gon, goff, αon, αo.
Calculate gsw using ff.

The method of determining the gain gsw will be described in detail below.

(1) Gain gon in single talk state
The gain gon of the multiplier 78 in the single talk state is calculated as follows.

Gon = (TLOSS / ERLE∨EN / EP) ∧1.0 (Equation 1) Here, TLOSS and EN are constants, ∨ represents max operation, and ∧ represents min operation.

TLOSS represents the desired value of the echo suppression amount, and TLOSS / ERLE represents the echo suppression amount by TLSS.
Represents the gain of the multiplier 78 for

EN represents the maximum value of noise power, and EN
/ EP represents the gain of the multiplier 78 for setting the output power of the NLP 57 to EN.

Gon takes the larger of the above two gains. In addition, the minimum value calculator 69 so that the gain does not exceed 1.
Limited by

The divider 65 executes the division of TLOSS / ERLE and the divider 66 executes the division of EN / EP.
The maximum value calculator 68 and the minimum value calculator 69 calculate gon in Expression 1.

(2) Gain goff in double talk state
The gain goff of the multiplier 78 in the double talk state is 1.
It is a constant of 0.

(3) The gain gsw of the multiplier 78 is calculated according to the following inference rules.

If EP is LARGE then gsw = goff Rule 1 If EP is SMALL then gsw = gon Rule 2 If EP <NP then gsw = gon Rule 3 Rule 1 is a case where the output of the adaptive filter is large, so Rule 2 is a state of being transmitted, and Rule 2 is a state of single talk because the output of the adaptive filter is small. Rule 3 is a single-talk state in which no voice is transmitted, because the noise power is larger than the output of the adaptive filter.

Rule 1 is represented by a membership function as shown in FIGS. Rule 1 antecedent section "EP
"is LARGE" is represented by the membership function of FIG. Also, "g = goff" in the consequent part is shown in FIG.
It is represented by the membership function of (b).

Rules 2 and 3 are also shown in FIGS.
It is represented by the membership function of FIGS. The membership functions of FIGS. 11A, 11C, and 11E are respectively represented by the data table 70 of FIG.
It is stored in the data table 72 and the data table 74.

Next, the table readers 71, 73 and 75 operate as EP
Or, when the EP / NP is input, the data tables 70, 7
EP that input the membership function from 2, 74
Alternatively, based on EP / NP, the degrees of coincidence α ₁ , α ₂ , α ₃ of the antecedent part of each rule are obtained. This state is shown in FIG.
11 (a), 11 (c) and 11 (e).

By the degree of coincidence α ₁ , α ₂ , α ₃ ,
Membership functions of the consequent part, FIG. 11 (b),
11 (d) and FIG. 11 (f), the head cutting operation is performed, resulting in FIGS. 12 (a), 12 (b) and 12 (c).

When these three membership functions are combined by the ∪ operation, the membership function shown in FIG. 12D is obtained. Here, a _{αoff = α 1 αon = α 2} ∨α 3. The degree of coincidence read by the table reader 71 is this α of
f, and the degree of coincidence output by the maximum value calculator 76 is this αon.

The defuzzifier 77 calculates the gain gsw by the following equation using the centroid method.

Gsw = (αoff × goff + αon × gon) /
(Αoff + αon) By controlling the multiplier 78 with the gain gsw thus obtained, the residual echo contained in e (n) is removed, and the distortion of the audio signal can be avoided. This NL
In P, since the gain gsw is determined by fuzzy inference, it is possible to respond to changes in the situation very smoothly, and the receiver does not feel discomfort. In addition, since the calculation of the gain does not require complicated calculation, it can be calculated easily and at high speed.

(Third Embodiment) The echo canceller of the third embodiment prevents a situation in which noise is added to the output due to an abnormal operation of the adaptive filter.

The adaptive filter of this echo canceller is
As shown in FIG. 14, the input signal d to the echo canceller
The signal f by subtracting the echo estimated value y (n) from (n)
A subtractor 86 that outputs (n) and a multiplier 87 that multiplies the echo estimated value y (n) by a coefficient gp (n) are provided. Other configurations are the same as the conventional one (FIG. 19).

The output e (n) of this adaptive filter is as follows.

E (n) = d (n) -gp (n) × y (n) This coefficient gp (n) is controlled to 0 at the time of abnormality. As a result, the contribution of the echo estimation value y (n) in the subtractor 85 is suppressed, and the output e (n) of the adaptive filter becomes equal to the input d (n). On the other hand, the subtractor 86 generates an output signal f (n) that is used for controlling the coefficient gp (n) and that does not suppress the contribution of the echo estimation value y (n).

As shown in FIG. 13, the echo canceller has a power S of the input signal d (n) as a whole.
A power calculator 79 that calculates P and a double talk determiner 81
14, an adaptive filter 82 shown in FIG. 14, a power calculator 80 that calculates the power EP of the signal f (n) output from the adaptive filter 82, and a coefficient gp (n) of the adaptive filter using SP and EP. Controlling protection coefficient controller 83 and NLP84
And

The configurations of the power calculators 79 and 80 are the same as those in FIG.

As shown in FIG. 15, the protection coefficient controller 83 includes a subtracter 108 for subtracting SP from EP and SP for EP.
The divider 109 for dividing by, the data tables 110, 112, 114 for storing the correspondence relationship between the input signal and the data, and the data table 11 for the data corresponding to the input EP-SP value.
Table readers 111 and 115 for reading from 0 or 114, and EP
/ SP reading out the data corresponding to the value of / SP from the data table 112, the table reader 111 or the maximum value calculator 116 which outputs the larger one of the outputs of the table reader 113, and the maximum value calculator 116. Alternatively, there is provided a minimum value calculator 117 which outputs the smaller one of the outputs of the table reader 115 as the coefficient gp.

The data tables 110, 112 and 114 store the data shown in FIGS. 16A, 16B and 16C, respectively. FIG. 17 shows the relationship between EP and SP input to the protection coefficient controller 83 and gp output at this time.

In FIG. 17, a region below the dotted line, that is, a region of EP ≦ SP is a state where echo is attenuated in single talk, that is, a state where the echo canceller is normally operating. At this time, gp = 1 and the normal operation is continued.

The region of EP> SP is a state where an abnormality has occurred. However, in practice, normal echo attenuation operation may be performed even in a region slightly above the dotted line, so gp = 1 in this region.

In the shaded area, g within the range of 0 <gp <1
p gradually changes. On the upper side of the shaded area, gp = 0, and the input of y (n) to the subtractor 85 is suppressed. In this case, the echo cannot be removed, but it is possible to prevent extra noise from being added to the output of the echo canceller.

As described above, the echo canceller of the third embodiment can prevent a situation in which correction is made in a bad direction even when an abnormal state occurs in which the estimated value of the impulse response is deviated. Also, with this echo canceller,
Since the coefficient change area is defined using the data table and the maximum value / minimum value calculator, a fine area can be defined with a simple configuration.

[0084]

As is apparent from the above description of the embodiments, the echo canceller of the present invention is provided with a coefficient controller for controlling the modification coefficient of the adaptive filter, so that the modification coefficient μ (n) of the adaptive filter is adjusted. The value can be controlled to an appropriate value, and thus the convergence speed and the estimation accuracy can be appropriately controlled.

By configuring the NLP of the echo canceller with a multiplier and a mechanism for controlling the gain by fuzzy inference, residual echo can be suppressed without causing distortion in the transmission signal.

Further, when the operation state of the echo canceller is not normal, it is possible to prevent the adaptive filter from adding noise to the transmission signal.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an echo canceller according to a first embodiment of the present invention,

FIG. 2 is a configuration diagram of a power calculator in the first embodiment,

FIG. 3 is a configuration diagram of a noise power estimator according to the first embodiment,

FIG. 4 is a configuration diagram of an adaptive filter according to the first embodiment,

FIG. 5 is a configuration diagram of a coefficient controller according to the first embodiment,

FIG. 6 is a data table diagram of a coefficient controller according to the first embodiment,

FIG. 7 is a diagram of a data converter of the coefficient controller in the first embodiment,

FIG. 8 is a configuration diagram of an echo canceller according to a second embodiment of the present invention,

FIG. 9 is a configuration diagram of an ERLE estimator according to the second embodiment,

FIG. 10 is a configuration diagram of an NLP in the second embodiment,

FIG. 11 is a diagram of a membership function for explaining the operation of the NLP in the second embodiment,

FIG. 12 is a diagram of a membership function that has undergone a head cut operation in the second embodiment;

FIG. 13 is a configuration diagram of an echo canceller according to a third embodiment of the present invention,

FIG. 14 is a configuration diagram of an adaptive filter according to a third embodiment,

FIG. 15 is a configuration diagram of a protection coefficient controller according to a third embodiment,

FIG. 16 is a diagram of a data table of a protection coefficient controller in the third embodiment,

FIG. 17 is a diagram showing characteristics of a protection coefficient controller in the third embodiment,

FIG. 18 is a configuration diagram of a conventional echo canceller,

FIG. 19 is a configuration diagram of an adaptive filter of a conventional echo canceller,

FIG. 20 is a diagram for explaining the operation of the NLP of the conventional echo canceller;

FIG. 21 is an explanatory diagram of a model of an echo canceller.

[Explanation of symbols]

1, 2, 51, 52, 79, 80 Power calculator 3, 53, 81, 206 Double-talk decision device 4, 56, 82, 207 Adaptive filter 5 Noise power estimator 6 Coefficient controller 7, 57, 84, 208 , 236 Non-Linear Processor (NL
P) 8 Absolute value calculator 9, 10, 13, 14, 21 to 28, 38 to 40, 59, 60, 78, 87, 89
~ 96, 106, 107, 211 ~ 218, 228, 229 Multiplier 11, 15, 20, 36, 61, 88, 104, 210, 226, 231, 232
Adder 12, 17, 29 to 31, 44, 50, 63, 97 to 99, 219 to 221 Delay device 16, 19, 45, 47, 62, 85, 86, 209, 235 Subtractor 18, 64 Filter coefficient control 32 to 35, 100 to 103, 222 to 225 Square calculator 37, 41, 58, 65 to 67, 105, 227 Divider 42, 70, 72, 74 Data table 43, 71, 73, 75 Table reader 46 data converter 48, 68 minimum value calculator 49, 69, 76 maximum value calculator 54 noise power estimator 55 ERLE estimator 77 defuzzifier 83 protection factor controller 200 telephone 201 two-wire transmission line 202 hybrid 203 AD converter 204 DA converter 205, 233 Echo canceller 230 Echo path 234 Echo path estimated value

Claims (6)

[Claims] 1. An echo canceller device comprising an adaptive filter for subtracting an echo estimation value from a transmission signal and a non-linear processor for removing residual echo contained in an output signal of the adaptive filter, wherein a correction coefficient for updating the coefficient of the adaptive filter is An echo canceller device, which is controlled based on a transmission signal input to the adaptive filter and an output signal of the adaptive filter. 2. When the SN ratio of the transmission signal is small,
When the correction coefficient is controlled to be small, and when the SN ratio is large, the correction coefficient is controlled to be large when the output signal is increasing, and when the output signal is decreasing. The echo canceller device according to claim 1, wherein the correction coefficient is controlled to be small. 3. An echo canceller device comprising an adaptive filter for subtracting an echo estimation value from a transmission signal and a non-linear processor for removing residual echo contained in the output signal of the adaptive filter, wherein the non-linear processor inputs the output signal. An echo canceller device comprising: a multiplier for controlling the gain of the multiplier and gain control means for determining the gain of the multiplier by fuzzy inference. 4. An echo canceller device comprising an adaptive filter for subtracting an echo estimation value from a transmission signal and a non-linear processor for removing residual echo contained in an output signal of the adaptive filter, wherein the echo estimation value is subtracted from the transmission signal. An echo canceller device, characterized in that control means for controlling the amount is provided, and the subtraction amount is suppressed when an abnormal state occurs in the adaptive filter. 5. The control means controls the subtraction amount based on the magnitudes of the power of the transmission signal and the power of a signal obtained by subtracting an echo estimation value from the transmission signal. The echo canceller device described in 1. 6. The magnitude relationship should be applied from a plurality of data tables holding a controlled variable as a function of a difference between the respective powers or a ratio of the respective powers, and a controlled variable read from the data table. The echo canceller device according to claim 5, wherein the echo canceller device is set by means for selecting a control amount.