JPH1168518A - Coefficient updating circuit - Google Patents

Abstract

PROBLEM TO BE SOLVED: To keep estimation error constant even at the time of disturbance variations, such as in double talk by controlling a block length in an addition normalization LMS method, so that an estimation error estimating circuit executes coefficient updating, only when estimation value is equal to or less than a desired estimation error. SOLUTION: A power calculation circuit 120 in an estimation error stably holding circuit 100 calculates the magnitude of a power PN of a disturbance Nj superimposed on a response signal gj from the differential signal Ej between a response signal gj of a signal transmission system 200 and an output signal Gj of a non-recursive filter 210 as an approximation disturbance power Pen , and estimates an estimation error Cn caused by updating a coefficient at the current time, based on the power Pen and the power Pn . A block length in the addition normalization LMS method is controlled, so that an estimation error estimating circuit 110 executes coefficient updating, only when the estimation value Cn is equal to or less than the desired estimation error Co .

Description

DETAILED DESCRIPTION OF THE INVENTION

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a coefficient updating circuit, and more particularly to a non-cyclic circuit for simulating a response of a signal transmission system from a known signal (reference signal) sent to a signal transmission system having unknown characteristics and its response signal. The present invention relates to a coefficient updating circuit for estimating coefficients of a type (FIR: Finite Impulse Response) filter.

Error (residual error) remaining as a result of coefficient update
There is a demand for a circuit that can maintain a constant value with respect to disturbance superimposed on the response signal and power fluctuation of the reference signal.

FIG. 6 shows a general system including a non-recursive filter and its coefficient updating circuit. In the figure, when a known reference signal _Xj is sent to the signal transmission system 200, a response expressed by the following equation is obtained at the output of the signal transmission system 200.

(Equation 1)

Here, j represents time (block length), and the impulse response h _j of the signal transmission system 200 has the length of the I sampling period, and is given by the following equation together with the reference signal X _j. Shall be.

(Equation 2)

(Equation 3)

At the same time, the non-recursive filter 210 is given by equations (2) and (3)
A pseudo response represented by the following equation is synthesized.

(Equation 4)

The output of the subtractor 230 is given by the following equation.

(Equation 5)

Here, N _j is an independent disturbance superimposed on the response g _j , and the adder 231 is a pseudo one that is not configured as an actual circuit.

[0007] The coefficient updating circuit 220 updates the coefficient of the non-recursive filter 210 so that the output E _j is minimized as follows.

(Equation 6)

In such a system, the input signal y
_When the power ratio between the response g _j and the disturbance N _j included in _j decreases, it becomes difficult to estimate the coefficients of the acyclic filter 210,
The estimation error increases.

FIG. 7 shows a specific example of this system.
Considering the acoustic echo canceller shown in (1), an increase in the estimation error of the coefficient of the acyclic filter 210 means that the degree of cancellation of the echo g _j by the synthesized pseudo echo G _j decreases.

A reduction in the degree of cancellation causes an increase in the reverberation sensation. This is not a favorable phenomenon for the acoustic echo canceller.

Therefore, it is necessary to provide a coefficient updating circuit of the non-recursive filter 210 which keeps the estimation error constant even if the power ratio rises or falls.

[0012]

2. Description of the Related Art A learning identification method NL is used as an algorithm for constructing such a coefficient updating circuit of the acyclic filter 210.
The MS method (Normalized Least Mean Square Algorithm) is well known. The coefficient updating equation can be expressed as the following equation.

(Equation 7)

Here, μ is a constant called a step gain and selected in the range of the following equation.

(Equation 8)

This learning identification method is characterized by a small amount of computation, and is therefore one of the algorithms most often studied for application to practical devices.

The error (estimation error) of the coefficient of the non-recursive filter 210 estimated by this learning and identification method includes the disturbance power P _N , the reference signal power P _X , and the step gain μ.
Is known as a function (coupling gain) of the following equation.

(Equation 9)

That is, when the power P _{N of} the disturbance increases or the power P _X of the reference signal X _j decreases, the estimation error increases.

In an acoustic echo canceller as shown in FIG. 7, the increase in disturbance power P _N is also caused by the utterance of a near-end speaker, which is called double talk, and is usually called double talk.
At the time of the increase, a measure is taken to stop the coefficient update.

Further, treatment to stop the coefficient update because the estimated error C increases as is apparent from equation (9) when the power P _N of the disturbance power P _X of the reference signal be constant was reduced Be taken.

In the acoustic echo canceller, the reference signal is a voice, and the power of the reference signal fluctuates greatly. Therefore, a measure for stopping the coefficient update is frequently performed, and the measure for stopping the coefficient update is one of the factors that delay the convergence of the coefficient.

The present inventors have already dealt with the latter method of stopping the updating of the coefficient when the reference signal power is reduced by using the addition-normalized LMS method [the adaptive filter coefficient in an unvoiced section useful for an acoustic echo canceller]. Renewal Continuation Law, Theory of Science
(A), vol. J78-A, no. 11, pp. 1403-1409 (1995-11)].

That is, when the coefficient update has converged (when the residual is minimized), the estimation error becomes constant. Therefore, a constant estimation error C is required to keep the system stable irrespective of the power fluctuation of the reference signal. _In order to maintain _{0 in} the presence of the disturbance power Q ₀ when not in the double talk state, the coefficient updating equation in the addition normalized LMS method shown in the following equation

(Equation 10) However,

[Equation 11]

(Equation 12) The coefficient updating in, when the reference signal power P _n went summed, and suggested may be configured to only execute when a threshold value P ₀ or more, previously calculated by the following equation.

(Equation 13)

Here, I is the number of taps of the non-recursive filter for synthesizing the pseudo response G _j . The threshold value P ₀ is obtained from Expression (9) as follows. That is, the following equation is obtained by substituting the estimated error C ₀ to be ensured by the system into the estimated error C of the equation (9), and substituting the disturbance power Q ₀ observed in the system into the disturbance power P _N. Is multiplied by the number of taps I to the reference signal power.

[Equation 14]

Here, the reason why the threshold value P ₀ is multiplied by the number of taps in the equation (14) is that the denominator P _N is actually calculated by the following equation.

(Equation 15)

As is clear from this, the denominator P _N is obtained on average as the number of taps (I) times the power P _X of the reference signal.

[0025]

From the expression (13), it is apparent that
This method is effective only when the disturbance power P _N is fixed to the constant value Q ₀ as described above. When the disturbance power P _N becomes P _N > Q ₀ (for example, during double talk)
In order to maintain the estimation error C ₀ , the threshold value needs to be P ₀ or more.

[0026] Therefore, there is a problem that occurs the same problem as described above must also stop the coefficient update when the threshold value is P ₀ or more.

Accordingly, the present invention provides an addition algorithm as an adaptive algorithm for estimating a coefficient of a non-recursive filter simulating an impulse response of a signal transmission system from a known reference signal transmitted to a signal transmission system having unknown characteristics and its response signal. Normalized L
An object of the present invention is to maintain a constant estimation error in a coefficient updating circuit using the MS method without stopping coefficient updating even during disturbance fluctuation such as double talk.

[0028]

In order to achieve the above object, as shown in principle in FIG. 1, a coefficient updating circuit 220 according to the present invention is connected to a stable estimation error holding circuit 100. .

The circuit 100 for stably holding the estimation error
Signal g of the signal transmission system 200_jAnd the output of the acyclic filter 210
Force signal G_jDifference signal E from_jFrom the response signal g_jSuperimposed on
Disturbance N_jPower P_NApproximate disturbance power P
e_nAnd the approximate disturbance power Pe_n
And disturbance power P_nWhen the coefficient is updated based on
Error C that occurs in_nAs well as the estimate
Value C_nIs the desired estimation error C ₀Coefficient only when
Block in the additive normalized LMS method to perform an update.
An estimation error estimating circuit 110 for controlling the lock length J;
It is configured.

In the above invention, according to the equation (13),
In order to maintain the desired estimation error C ₀ when the power of the disturbance is P _N , the denominator P _{n of} the second term on the right side of the equation (10) is μ μ
The configuration may be such that the coefficient is updated when P _NI / (2-μ) C ₀ or more.

That is, the power P _N of the disturbance is measured, and μP _N I / (2-μ) C ₀ is obtained therefrom.
_It can be seen that the desired estimation error C ₀ can be maintained by using it instead of ₀ .

In the system shown in FIG. 6, the signal available for calculating the disturbance power P _N is either the input y _j or the residual E _j.
It is. However, since the response g _{j of the} unknown system is generally much larger than the disturbance N _J, it is difficult to directly calculate the disturbance power P _N from the input y _j .

Therefore, in the present invention, the power P _N of the disturbance is calculated (estimated) from the residual E _j . Now, the residual E _j is composed of the disturbance N _j and the residual echo (g _j −G _j ) according to the equation (5), and the power of the residual echo is μP _N according to the equation (9).
/ (2-μ). The size is inversely proportional to the step gain mu, the same size as the disturbance N _j at maximum mu = 1 (step gain is usually selected in the range 0 <μ <1) of.

In general, the step gain μ is selected to be smaller than 1 in order to keep the estimation error C small.

Therefore, in the present invention, the power P _N of the disturbance is approximated by the power of the residual E _j .

In this case, the power of the disturbance is estimated to be larger than the actual power by the amount of the residual echo (g _j -G _j ), but there is no problem since it works in the direction of keeping the estimation error C smaller.

That is, in the present invention, the same block length is used at the same time as the calculation of the denominator P _n by the following equation:

(Equation 16) Was calculated, the Pe _n / J This was divided by the block length J,
If to approximate the power Q _n in the n blocks disturbance varying the estimation error that occurs when the updating coefficients in the denominator P _n at this time is estimated as follows.

[Equation 17]

In this case, since the goal is to maintain the estimation error C ₀ , when C _n > C ₀ , the coefficient cannot be updated and the stop processing is performed as described above.

Therefore, in the present invention, when C _n > C ₀ , the block is extended. Estimate C _n of the estimation error becomes larger by P _n is to extend the block in the denominator of equation (17) approaches the C ₀ as a target by the extension.

[0040] That is, to extend the block until C _n ≦ C _0, if updating coefficients at the time of a C _n ≦ C _0, regardless of the variation of the power P _N of the disturbance such as during double talk,
Estimation error C will be maintained at a desired constant value C _0.

The power calculating circuit 120 is reset by the coefficient updating circuit 220 as shown by a broken line every time the coefficient updating is executed.

[0042]

FIG. 2 shows an example in which a coefficient updating circuit according to the present invention is applied to an acoustic echo canceller. When applying the present invention to this acoustic echo canceller, a particular problem arises.

That is, the residual E _j is a _factor of double talk (increase in noise N _j , or near-end speaker's microphone 320
Input) and the change in the impulse response g _j of the echo path (path from the speaker 310 to the microphone 320), and the extension of the block for the former and the On the other hand, in order to accelerate convergence, a process of shortening a block is required.

When the present invention is applied to the acoustic echo canceller shown in FIG. 2, it is necessary to identify these two and execute processing suitable for both.

For example, in order to perform the discrimination, the discrimination method proposed by the present inventors [discrimination between double talk and echo path fluctuation, IEICE Technical Report, EA94-15 (1994-05)] is used. it can.

That is, this method is represented by the following equation.

(Equation 18)Is an identification parameter, and this parameter R_EYValue of (k)
Is double talk (E_jG_j= N_jG_j) Sharply reduced
And the echo path change (E_jG_j= ((G_j-G_j+ N _j) G_j
≒ -G_j ^Two) Uses both to settle down to a certain value.
Are identified.

FIG. 3 shows a parameter R for double talk.
FIG. 4 shows an example of the transition of _EY (k), and FIG. 4 shows an example of a change in the echo path. The near-end speaker signal and the far-end speaker signal shown in both figures respectively show the input N _j to the microphone 320 and the reference signal X _j , and in FIG. It is shown to settle to a certain value.

[0048] In this application example, in addition to the estimated error of stably holding circuit 100 shown in FIG. 1, calculates a parameter R _EY (k) in the calculation circuit 610, further the parameter R _EY (k)
Is used to control the timing of coefficient update in the determination circuit 620 as shown in the flowchart of FIG. Hereinafter, this will be sequentially described together with the stable holding circuit 100 described above.

First, in the estimation error stable holding circuit 100, according to equations (11), (12) and (16), A _n , P _n ,
Calculate Pe. At the same time, in the calculation circuit 610,
The parameter R _EY (k) given by the equation (18) is calculated by modifying the following equation so as to conform to the present application example (step S1).

[Equation 19]

Next, the stable holding circuit 100 for the estimation error is expressed by the following equation.
As shown in (13), it is monitored whether or not P _n ≧ P ₀ (the same S
2) If P _n > P ₀ , the magnitude of the estimation error C _n that would occur when the coefficient is updated at the present time is calculated using equation (17) (S 3).

Since the initial value of a flag E described later is "0", the process proceeds from step S4 to S5.

If C _n ≦ C ₀ , a predetermined estimation error C ₀ is secured even if the coefficient is updated at this time, so the coefficient updating circuit 220 is instructed to update the coefficient (S5 and S1).
4, S15).

If C _n > C _{0 in} step S5, there is a possibility of double talk or echo path fluctuation. Since the initial value of the flag D is “0” and the initial value of the block length J ′ is also “0”, J ′ <2 pitches (S1).
6).

Accordingly, the current identification parameter R _n (J) is stored in the memory (S7), and the calculation of J 'is started (S7).
8), it is determined that the double talk is not possible, D = 1 (S10), and the process returns to step S1.

The processing is performed in steps S1, S2, S3, S4
S5, S6 and S1 are repeated. During this time, if C _n ≦ C ₀ in step S5, normal processing can be performed, and the process proceeds to step S14.

When J ′ ≧ 2 pitches, the determination circuit 620 sets the identification parameter R _n (J ′) at this time to R
_n (J) (step S9), if it has greatly decreased, it is determined that double talk has occurred, and the process returns to step S1. If it has not decreased, it has been determined that echo path fluctuation has occurred.
(S11).

In step S9, the echo path fluctuation (E = 1)
When it is determined that, since the process proceeds to the The S12 step S4, whether the assumed value Q ₀ of the disturbance power when not in double-talk separately prepared, resulting in a block that has not been determined that echo path change and double-talk so far Using P ₀ calculated as P _n / J as disturbance power Q ₀ , the coefficient is updated every time P _n ≧ P ₀ (step S2) through steps S14 and S15 until C _n ≦ aC _0. I do.

After updating the coefficient in this manner, P _n
Estimation error C of at ≧ P ₀ was estimated using the Pe _n / J
_n (S3) satisfies the relationship of C _n ≦ C ₀ , the coefficients converge and the number of echoes remaining in the residual E _j decreases, and Pe _n / J
It can be determined that the disturbance power can be estimated using.

In practice, it is not necessary to wait for such complete convergence of the coefficients. Somewhat
Even if the echo is mixed with the residual E _j and the power of the disturbance is estimated to be large, the estimation error is reduced to a predetermined value C _0.
It only works in the direction that can be obtained smaller.

Therefore, a small constant a is prepared, and P
_When the estimated error error C _n estimated at the time of updating the coefficient at _n ≧ P ₀ is C _n ≦ aC ₀ (S12).
It is possible to return to the normal processing. At this time, the flag E
= 0 (S13).

[0061]

As described above, according to the coefficient updating circuit of the present invention, the power of the disturbance power superimposed on the response signal is obtained from the difference signal between the impulse response signal of the signal transmission system and the output signal of the acyclic filter. Calculate the magnitude as approximate disturbance power, estimate the estimation error that occurs when the coefficient is updated at the present time based on the approximate disturbance power and the disturbance power, and execute the coefficient update only when the estimated value is equal to or less than the desired estimation error. In this case, the block length in the addition normalized LMS method is controlled so that the estimation error can be kept at a constant value without stopping the coefficient update even during disturbance fluctuation such as double talk.

[Brief description of the drawings]

FIG. 1 is a block diagram showing in principle a coefficient updating circuit according to the present invention.

FIG. 2 is a block diagram showing an acoustic echo canceller to which a coefficient updating circuit according to the present invention is applied.

FIG. 3 is a graph showing identification parameters at the time of double talk in the acoustic echo canceller to which the coefficient updating circuit according to the present invention is applied, in relation to a near-end talker signal and a far-end talker signal;

FIG. 4 is a graph showing identification parameters when an echo path changes in an acoustic echo canceller to which a coefficient updating circuit according to the present invention is applied, in relation to a near-end speaker signal and a far-end speaker signal;

FIG. 5 is a flowchart illustrating a coefficient update timing determination procedure in an acoustic echo canceller to which the coefficient update circuit according to the present invention is applied.

FIG. 6 is a block diagram showing a general system to which a coefficient updating circuit according to the present invention is applied.

FIG. 7 is a block diagram showing a general acoustic echo canceller.

[Explanation of symbols]

 100 Estimation error stable holding circuit 110 Estimation estimation circuit 120 Power calculation circuit 200 Signal transmission system 210 Acyclic filter 220 Coefficient update circuit 230, 231 Addition circuit 310 Speaker 320 Microphone 610 Identification parameter calculation circuit 620 Judgment circuit Same in the figure Symbols indicate the same or corresponding parts.

──────────────────────────────────────────────────の Continued on the front page (51) Int.Cl. ⁶ Identification symbol FI H04B 3/23 H04M 1/60 C // H04M 1/60 G10K 11/16 H

Claims (4)

[Claims] 1. An addition-normalized LMS as an adaptive algorithm for estimating coefficients of a non-recursive filter simulating an impulse response of a signal transmission system from a known reference signal transmitted to a signal transmission system with unknown characteristics and its response signal. A coefficient updating circuit using a method for calculating the magnitude of the power of the disturbance superimposed on the response signal from the difference signal between the response signal and the output signal of the acyclic filter as an approximate disturbance power; A block in the addition-normalized LMS method for estimating an estimation error generated when a coefficient is updated at the present time based on the disturbance power and the disturbance power, and executing the coefficient update only when the estimated value is equal to or less than a desired estimation error. A coefficient updating circuit comprising: an estimation error estimation circuit for controlling a length; 2. A parameter according to claim 1, wherein said impulse response signal, an output signal of said non-recursive filter and a residual signal are inputted to calculate a parameter for discriminating between a change in said impulse response and an increase in said disturbance power. An estimation parameter estimation circuit for estimating an estimation error based on disturbance power prepared in advance in the case of a change in the impulse response based on the identification parameter, and an estimation error estimation circuit for an increase in the disturbance power based on the identification parameter; And a determining circuit for controlling a block length in the addition normalized LMS method based on the obtained value. 3. The method according to claim 2, wherein a value obtained by normalizing the correlation between the residual signal and the output signal of the non-recursive filter with the power of the response signal on which disturbance is superimposed is used as the identification parameter. Coefficient updating circuit. 4. A coefficient updating circuit according to claim 1, wherein said estimation error estimation circuit uses a value obtained by multiplying said desired estimation error by a constant.