JP2000174592A - Adaptive filter - Google Patents

Abstract

PROBLEM TO BE SOLVED: To suppress increase in an estimate error with respect to the adaptive filter that controls a block length. SOLUTION: A comparator circuit 130 estimates whether or not a level of a residual echo Ej reaches a level causing disturbance (1st threshold) when a coefficient of a noncyclic filter is updated even under the lock length control that keeps a prescribed estimate error even against power fluctuation in a reference signal on the basis of a numerical value newly added to an accumulated sum stored in a normalized power calculation circuit 110 for the reference signal Xj or like. Then only when the comparator circuit 130 discriminates no disturbance, a coefficient update circuit 140 conducts accumulation of a numeral corresponding to power of the reference signal Xj to normalize the cross correlation function between the residual echo Ej and the reference signal Xj according to a 1st control signal from the comparator circuit 130 to each of calculation circuits 110, 120 and uses the accumulated sum to update the coefficient.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive filter, and more particularly to estimating a coefficient of a non-recursive filter simulating an impulse response of a signal transmission system from a reference signal transmitted to a signal transmission system having unknown characteristics and its response signal. And an adaptive filter including a coefficient updating circuit.

FIG. 3 shows an acoustic echo canceller as an example of a system using such an adaptive filter. Although the configuration of the adaptive filter is arbitrary, a non-recursive type is shown here for the sake of simplicity. Now, the far-end speaker voice signal X _j [j: sample time ind
ex] is output from the speaker 201 to reach the ear of the near-end speaker, and as the echo g _{j the} microphone 2
Enter around 02.

On the other hand, the recursive filter 210 synthesizes a pseudo echo G _j approximated to the echo g _j ,
The echo g _j is canceled by giving the signal to 211. At this time, the coefficient updating circuit 220 sets the coefficient of the non-recursive filter 210 so that the residual echo E _j remaining after the cancellation is minimized.

[0004]

(Equation 1) To update. Here, I is the number of taps of the non-recursive filter 210.

Here, the power of the far-end speaker voice signal _Xj as a reference signal used for updating the coefficient _Hj may vary greatly. Ambient noise N _j input to the microphone 202 with the echo g _j as can be seen from the coefficient updating principle of the coefficient H _j (principle difference between the pseudo echo G _j and echo g _j is updated so as to minimize), the Coefficient H _j
Act as a disturbance to prevent the update of

In this case, the far-end speaker voice signal X is clearly_j
When the power drops, echo g_jRemains small
Echo E_jAmbient noise g_jIncreases,
Coefficient H _jExpected to increase errors in updating
It is. Therefore, even when the power of the reference signal drops, the coefficient
H_jSo that the error related to the estimation of
There is a need to.

[0007]

In the acoustic echo canceller shown in Prior Art FIG. 3, the FIR filter 210 to minimize the residual echo E _j
The most representative algorithm for estimating the coefficient is a learning identification method represented by the following equation.

[0008]

(Equation 2) Here, μ is a constant called a step gain, and E _j is a remaining signal obtained by canceling an echo g _j called a residual echo by a pseudo echo G _j , that is,

[0009]

(Equation 3) It is.

X _j is the output of the shift register constituting the non-recursive filter 210 at time j, and is represented by the following equation.

[0011]

(Equation 4) When the coefficient of the acyclic filter 210 is updated by H _j using this learning identification method, the root mean square of the estimation error remaining after the convergence of the coefficient H _j Has already been formulated as follows:

[0012]

(Equation 5) Here, σ _N ² is the average power of the ambient noise N _j σ _X ² is the average power of the reference signal.

As is apparent from equation (5), when the power σ _X ² of the reference signal decreases or the noise power σ _N ² increases, the error σ _d ² increases and the ability to cancel the echo g _j decreases. You can see that The most common countermeasure to be taken when the error σ _d ² is expected to increase is to stop updating the coefficient H _j . However, stopping the coefficient update and thus slow the convergence of the coefficients H _j.

Therefore, the present inventors have proposed the following method in Japanese Patent Application No. 9-218697 as a solution. That is, an algorithm for updating the coefficient of the non-recursive filter 210 is represented by the following equation.

[0015]

(Equation 6) In this method, the coefficient can be constantly updated without increasing the estimation error by controlling the block length J. Here, ρ is a step gain and is a constant corresponding to μ in Expression (2).

As can be seen from equation (6), in the block execution type learning identification method, the coefficients of the acyclic filter 210 are updated in units of the J sampling period. Equation (6) shows how the coefficient in the n-th block is updated. To explain this block implementation NLMS More specifically, the estimation error is first left even after the convergence of the coefficients H _n is calculated by the following equation.

[0017]

(Equation 7) It is apparent from this equation that the estimation error σ _d ² is set to the required value C irrespective of the fluctuation of the power σ _N ² of the ambient noise and the power σ _X ² of the reference signal.
Adjust the block length J to keep it below ₀

[0018]

(Equation 8) It can be seen that the coefficient should be updated only when Naturally, for this control, it is necessary to estimate the power σ _N ² of the ambient noise and the power σ _X ² of the reference signal. This can be done as follows.

First, the average power σ _X ² of the reference signal is given by the following equation (6).
Is the following formula,

[0020]

(Equation 9) From the approximation. Further, the ambient noise N _j
Since the average power σ _N ² of the residual echo E _j is approximated to the ambient noise N _j when the echo g _j is sufficiently canceled, the following equation is estimated.

[0021]

(Equation 10) That is, using these two observation results, the following equation:

[0022]

[Equation 11] If updating coefficients H _n by adjusting the block length J such that, it becomes possible to constantly updated coefficients while ensuring the required estimation error C ₀ to the power variation of the ambient noise and the reference signal. Obviously, the constant update does not delay the convergence, unlike the method of stopping the coefficient update.

On the other hand, controlling the step gain ρ with respect to the power (σ _X ² ) fluctuation of the reference signal from the equation (6) is a norm. This is not applicable because the learning identification method is returned to the LMS by canceling out the effect of the normalization by.

However, the power of the ambient noise (σ _N ² )
It is effective to control the step gain ρ with respect to the fluctuation. That is, in equation (8), the block length J is controlled with respect to the power fluctuation of the reference signal to make σ _X ² J a constant value,
It is clear that the estimation error can be kept constant by adjusting the step gain ρ to a constant value of σ _N ² ρ with respect to the power fluctuation of the ambient noise.

Then, an appropriate threshold value P ₀ is determined for the power fluctuation of the reference signal, and

[0026]

(Equation 12) The execution of the coefficient update is determined when.

At this point, the estimation error is

[0028]

(Equation 13) Is guaranteed. This inequality, on the other hand, gives the step gain ρ as

[0029]

[Equation 14] It is shown that the estimation error is kept at or below C ₀ when given so as to satisfy the relationship.

That is, using the step gain ρ as a variable

[0031]

(Equation 15) Or using the current addition norm P _n

[0032]

(Equation 16) , The estimation error is kept at C ₀ or less even with respect to the power fluctuation of the ambient noise.

The advantage of the combined use of the step gain control is that the frequency of coefficient update is maintained high because the block length does not increase due to an increase in ambient noise, and therefore, when the reference signal is a colored signal. In addition, the convergence is faster than when only the block length control is performed. This indicates that the method is suitable for application to an acoustic echo canceller using voice as a reference signal.

FIG. 4 shows a commercially available digital signal processor (D
SP) constitutes an acoustic echo canceller based on the block length control described above, and cancels the echo when the filter coefficient is constantly updated as a news sound with a reference signal recorded in an environment where the ambient noise level is constant (step gain is constant) It is the result of measuring the performance. From this result, it can be seen that, even when the news is silent and the reference signal is only background noise, the residual echo is stably eliminated (the difference from the echo is large) by the block length control.

[0035]

When such a block length control method is applied to an actual acoustic echo canceller,
Especially when the speaker is separated from the microphone, the speaker output may be weak and may not reach the microphone as an echo or may not reach the level at which analog signals can be converted to digital signals (A / D conversion). . This is more likely when the noise level is low in sections where there is no sound, such as when used in a quiet room.

As described above, when the echo is low, only the ambient noise N _j remains in the residual echo E _j , and if the coefficient is updated in this state, the estimation error greatly increases. FIG. 5 shows an echo canceling effect when the acoustic echo canceller used for the measurement of FIG. 4 is operated by using a sound recorded on a compact disc having a low noise level in a section where no sound is made as a reference signal. is there.

Obviously, both signals are in close proximity,
In such a case, if the coefficient is constantly updated by the block length control, the echo canceling effect by the echo canceller is lost. Accordingly, an object of the present invention is to provide an adaptive filter capable of suppressing an increase in estimation error even when the speaker output of the reference signal is at such a low level that does not appear in the residual echo.

[0038]

FIG. 1 shows an example of the configuration of a coefficient updating circuit in an adaptive filter according to the present invention. In FIG. 1, the normalized power calculation circuit 110
It receives the reference signal X _j, the cross-correlation function between the reference signal X _j and the residual echo E _j are calculated by cross-correlation calculation circuit 120 in the circuit for calculating a value corresponding to the power of the reference signal to normalize is there.

When the comparison circuit 130 compares a numerical value added to the accumulated value stored in the normalized power calculating circuit 110 with a first threshold value prepared in advance, and the numerical value becomes equal to or less than the first threshold value, The numerical value is not added to the cumulative addition value stored in the normalized power calculation circuit 110 and the cross-correlation calculation circuit 120. On the other hand, when the numerical value is equal to or more than the first threshold value, the first value is added to the cumulative addition value. Control signal comparison circuit 13
0 is sent to each calculation circuit 110 and 120.

Then, the comparison circuit 130 compares a second threshold value prepared separately in advance with the cumulative addition value stored in the normalized power calculation circuit 110 as a conventional block length control (Equation (12)). see), when the cumulative addition value is updated coefficients H _n by using the stored accumulated value to the normalized power calculation circuit 110 and the cross-correlation calculation circuit 120 and the coefficient update circuit 140 when the above second threshold Simultaneously normalized power calculation circuit 1
The second control signal for performing the block length control by resetting the accumulated addition value stored in 10 and the cross-correlation calculation circuit 120 is given to the calculation circuits 110 and 120 and the coefficient update circuit 140.

The coefficient updating circuit 140 updates the filter coefficient H _n receives the second control signal, and outputs the new coefficients H _{n + 1.} That is, the difference (residual echo) E _j between the response of the unknown signal transmission system and the output of the non-recursive filter is determined for the block length control in which the predetermined estimation error is maintained even in the power fluctuation of the reference signal X _j. The cumulative addition stored in the normalized power calculation circuit 110 for the reference signal or the like determines whether or not the magnitude (the first threshold) at which the disturbance occurs when the coefficient of the non-recursive filter is updated also by the block length control. The comparison circuit 130 estimates from the magnitude of the numerical value newly added to the value.

Then, only when the comparison circuit 130 determines that there is no disturbance, the coefficient updating circuit 140 allows the cross-correlation function between the residual echo E _j and the reference signal X _j and the power of the reference signal for normalizing the cross-correlation function. The cumulative addition of the numerical value corresponding to
If the coefficient is updated by the cumulative addition value in accordance with the first control signal from 130 to calculation circuits 110 and 120 and the cumulative addition value is performed in accordance with the second control signal, a non-recursive filter can be used even for power fluctuations in which the level of the reference signal is greatly reduced. Is not performed, so that the filter coefficients are updated without being disturbed.

[0043]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS As an embodiment of an adaptive algorithm for estimating a filter coefficient in an adaptive filter according to the present invention, it is possible to use block control by a block execution type learning identification method shown in equation (6). it can. Less than,
This will be described with reference to FIG.

In an acoustic echo canceller or the like using the adaptive filter of the present invention, the speaker output level of the far-end talker voice and the microphone input level of the near-end talker voice are usually determined with reference to ITU-T Recommendation P.340. Then, the gain of the amplifier that does not clip the peak of the digital signal is determined from the speaker output level, the microphone input level, and the arrangement of the speaker and the microphone. A / D at this point
In the conversion, the lower limit of the input level of the echo converted from the analog signal to the digital signal is determined.

If this echo is lower than the lower limit and cannot be converted into a digital signal, the difference between the echo and the pseudo echo (residual echo) is only ambient noise, and
As described above, if the coefficients of the non-recursive filter are updated at this time, the coefficients are greatly disturbed.

The fact that the level of the echo is equal to or lower than the lower limit (first threshold) is determined by an equation calculated for each sampling period.
(6) Norm that constitutes the denominator of the second term

[0047]

[Equation 17] Expected from the size of. This norm is calculated by the normalized power calculation circuit 110 that receives the reference signal _Xj .

Therefore, a threshold value P _{min at} which the level of the echo is expected to be equal to or lower than the lower limit value with respect to the norm is prepared in the comparison circuit 130.

[0049]

(Equation 18) Are compared by the comparison circuit 130.

When this relationship holds, the norm and the cross-correlation E _j X _j between the residual echo and the reference signal are added to the denominator and the numerator of the second term of the equation (6) so as not to be added (accumulated). If the control signal of 1 is given to the normalized power calculation circuit 110 and the cross-correlation calculation circuit 120, the coefficient will not be disturbed because the coefficient update circuit 140 does not update the coefficient.

Therefore, when the relationship of the expression (18) is not established, the echo is not expected to be lower than the lower limit value.
10 and 120 perform cumulative addition. Therefore, the coefficient updating circuit 140
Will update the coefficient. As another embodiment of the adaptive algorithm for estimating the filter coefficient in the adaptive filter according to the present invention, an individual normalized LMS (INLMS: Individuall
y Normalized Least Mean Square) block length control can also be used. This individually normalized LMS method is suitable for fixed-point arithmetic and is more likely to be applied to actual systems. This embodiment will be described below.

First, if the basic expression of the individual normalized LMS method is not a vector expression but an element expression, for example, an expression for updating the m-th coefficient H _n (m)

[0053]

[Equation 19] Is expressed as

Since the individual normalized LMS method is based on block operation processing as can be seen from equation (19), this block length J is adjusted in the block execution type learning identification method shown in equation (6). A similar effect can be obtained by controlling according to. However, when applying the present invention to this adaptive algorithm, some modifications are required as follows. First, the determination as to whether or not to perform the cumulative addition of the second term in equation (19) is different from the block execution type learning identification method.
9) It is necessary to perform this processing depending on the magnitude of the reference signal X _j (m) as a numerical value added to the normalized power serving as the denominator of the second term.

That is, the lower limit value X _min of the amplitude of the reference signal X _j (m), which is expected to constitute an echo of a low level that does not have a value as a digital signal at the time of A / D conversion, is set as the first threshold value. Determined,

[0056]

(Equation 20) The comparison circuit 130 determines whether or not the relationship is established.

Then, a modification is made such that the denominator and the numerator of the second term of the equation (19) are calculated by the calculation circuits 110 and 120 by the first control signal only when the relation of the equation (20) is established. The addition of the component related to the reference signal in which the coefficient of the non-recursive filter is disturbed will not be performed.

On the other hand, this reference signal X _j (m)
+1 is the (m + 1) th tap output X _{j + 1} (m + 1). From here, the same signal up to the final tap is given by equation (20)
Is repeatedly performed by the comparison circuit 130. This repetition of the same determination causes an increase in the amount of processing when actually executed by a digital signal processor (DSP).

Therefore, a second shift register (not shown) is provided in the comparison circuit 130 separately from the shift register constituting the non-recursive filter 210 in FIG. 3, and is referred to at the input of the second shift register. The determination of Expression (20) is performed on the signal _Xj . as a result,

[0060]

(Equation 21) If it becomes zero,

[0061]

(Equation 22) , X _j is input to the second shift register, and the tap output of the second shift register is supplied to the calculation circuits 110 and 120 as the first control signal.

In each of these calculation circuits 110 and 120, if the denominator and the numerator of the second term of the equation (19) are calculated, the waste of repeating the same judgment can be eliminated. Furthermore, individual normalization
In the LMS method, the normalized power is shared by all taps in equation (19) as shown in the following equation to reduce the amount of computation

[0063]

(Equation 23) Can also be transformed. Here, J ≧ I.

The feature of this expression form is the normalized power of the second term denominator on the right side.

[0065]

(Equation 24) Uses the tap “1” in common for all taps, and the filter coefficient is limited to one within one sampling period. As a result, the amount of calculation is reduced. Further, in this format, the reference signal X _j stored in the second shift register is used only in the calculation of the numerator of the expression (23) in the calculation circuit 120 according to the determination result of the expression (22).

Further, in the above-described individual normalized LMS method, the calculation of the numerator of the equation (23) in the calculation circuit 120 is performed not on the reference signal _Xj itself but on its polarity (±), and the calculation circuit 11 for the remaining denominator is used.
Polarized-x IN which obtains the calculation at 0 by the absolute value of the reference signal X _j
It can be modified to the LMS (Individually Normalized Least Mean Square) method.

Here, the following equation is used to show an update equation in which the normalized power calculated by the calculation circuit 110 is shared by all taps.

[0068]

(Equation 25) In this case, in place of the reference signal _Xj stored in the second shift register, zero is set when the determination result of the previous equation (21) satisfies the relation of the equation (21), and otherwise, At times, the same effect can be obtained by storing ± 1.

However, this ± 1 is the same as ± α. It simply multiplies the step gain by α. Of course, the adjustment of the block length in this case is calculated by equation (25)
Is determined based on the size of the denominator. FIG. 2 shows the result of confirming the effect when the above control method is applied under the same conditions to the acoustic echo canceller used for the measurements of FIGS. 4 and 5. Also, the same low-level audio signal as in FIG. 5 is used as the reference signal.

From these results, it can be seen that the disadvantage of the conventional method which occurs when the echo is at a low level has been eliminated, and the echo has been stably canceled out. When it is necessary to use the above cumulative addition number (block length) for some other processing as in equation (16), and when the control according to each of the above embodiments is executed in a fixed-point DSP, When the input of the small reference signal continues, the addition number J may exceed the upper limit value that can be expressed by the fixed-point type DSP.

At that time, the number returns to zero after the addition number J exceeds the upper limit value. If this reset to zero occurs, the correct cumulative addition number cannot be used for another process. Therefore, when a very long block length is required until P _n ≧ P ₀ shown in the equation (12), the coefficient update becomes invalid or a large number of reference signals having a size close thereto become large. Means. Therefore, in this case, even if the cumulative addition for updating the coefficients is forcibly reset, the influence on the updating of the coefficients is small.

That is, an upper limit value is set for the block length J, and the calculation circuit 110 or 120 monitors this. When the block length reaches the upper limit value, the respective circuits 110, 12
By resetting the operations at 0 and 140, it is possible to prevent the block length J from exceeding an upper limit value that can be represented by a fixed-point type DSP.

[0073]

As described above, according to the adaptive filter according to the present invention, the residual echo is non-cyclic even when the block length is controlled by the block length control in which the predetermined estimation error is maintained even in the power fluctuation of the reference signal. The magnitude of a numerical value that is newly added to the cumulative addition value stored in the normalized power calculation circuit such as a reference signal as to whether or not the magnitude (first threshold value) at which the disturbance occurs when the coefficient of the type filter is updated is determined. The coefficient updating circuit predicts the cross-correlation function between the residual echo and the reference signal and the numerical value corresponding to the power of the reference signal for normalizing the cross-correlation function only when the comparison circuit predicts and does not disturb. Is performed in accordance with the first control signal from the comparison circuit to each calculation circuit, and the coefficient is updated based on the cumulative addition value, so that an echo that does not reach the microphone is generated. It is possible to remove the issue from the coefficient update,
The stability of the estimation error is further improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing in principle the configuration of an adaptive filter according to the present invention.

FIG. 2 is a graph showing the echo canceling effect of the adaptive filter according to the present invention.

FIG. 3 is a block diagram showing a configuration example of a general acoustic echo canceller.

FIG. 4 is a graph showing an echo canceling effect in a conventional adaptive filter.

FIG. 5 is a graph showing an echo canceling effect when a voice recorded on a CD having a low noise level in a section where no voice is generated by a conventional adaptive filter is used as a reference signal.

[Explanation of symbols]

 Reference Signs List 110 Normalized power calculation circuit 120 Cross-correlation calculation circuit 130 Comparison circuit 140, 220 Coefficient update circuit 201 Speaker 202 Microphone 210 Non-recursive filter 211 Subtractor In the drawings, the same symbols indicate the same or corresponding parts.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Tsutomu Hoshino 4-1-1 Kamikadanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Inside Fujitsu Limited (72) Inventor Toshiro Oga 4-1-1 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture No. 1 Fujitsu Limited F-term (reference) 5D020 CC06 5H004 GA40 HA20 HB15 KA32 KB21 KC12 KC44 KC45 KD70 MA11 5J023 DA05 DB03 DC07 DD01 DD07 5K046 HH19 HH37 HH42 HH53 HH79 9A001 EE02 FF01 FF05 GG05 FF01 H05 GG

Claims (7)

[Claims] 1. A coefficient updating circuit for estimating a coefficient of a non-recursive filter simulating an impulse response of a signal transmission system from a reference signal transmitted to a signal transmission system whose characteristics are unknown and a response signal thereof, A cross-correlation calculating circuit for calculating a cross-correlation function between a residual echo and the reference signal, and a cross-correlation function for normalizing the cross-correlation function. A normalized power calculation circuit for calculating a value corresponding to the power of the reference signal; and comparing a numerical value newly accumulated in the cumulative addition value stored in the normalized power calculation circuit with a first threshold value, A first control signal for giving an instruction to accumulate the numerical value to the accumulated value stored in the normalized power calculating circuit and the cross-correlation calculating circuit only when the value is equal to or more than the first threshold value; The product addition value is compared with a second threshold value, and only when the cumulative addition value is equal to or greater than the second threshold value, the cumulative addition value stored in the cross-correlation calculation circuit and the normalized power calculation circuit and the coefficient update A comparison that updates the coefficient with the coefficient stored in the circuit and generates a second control signal for resetting the accumulated sum stored in the normalized power calculation circuit and the cross-correlation calculation circuit. And an adaptive filter, wherein the coefficient updating circuit receives the second control signal, updates the coefficient, and outputs a new coefficient. 2. The method according to claim 1, wherein when the block execution type learning identification method is used as an adaptive algorithm for estimating the coefficient, the comparison circuit uses the magnitude of the norm of the reference signal as the numerical value. Adaptive filter. 3. An individualized LM according to claim 1, wherein said adaptive algorithm for estimating said coefficient is an individualized LM.
An adaptive filter, wherein when the S method is used, the comparison circuit uses the magnitude of the reference signal accumulated in the normalized power calculation circuit as the numerical value. 4. The comparison circuit according to claim 3, wherein the comparison circuit has a shift register, and compares the magnitude of the reference signal with the first threshold value when the reference signal is input. For a large reference signal, the signal value is directly input to the shift register, and for a reference signal smaller than the first threshold value, zero is input to the shift register, and the result input to the shift register is input to the first register.
An adaptive filter characterized in that the cumulative addition is executed by giving the control signal to both calculation circuits as a control signal. 5. The method according to claim 4, wherein when the normalized power in the individual normalized LMS method is shared by all taps, the signal value stored in the shift register is supplied only to the cross-correlation calculation circuit. An adaptive filter characterized in that a normalized power calculation circuit uses a common normalized power value for all taps. 6. An adaptive algorithm for estimating said coefficient, wherein Polarized-x
When the INLMS method is used, a signal value input to the shift register is used as a signal value of a reference signal larger than the first threshold, and a polarity of the reference signal or a value obtained by multiplying the polarity by a constant is used. Adaptive filter. 7. The circuit according to claim 1, wherein said calculation circuit has an upper limit value of the cumulative addition number, and when said cumulative addition number reaches said upper limit value, said calculation circuit sends a signal to each circuit for updating a coefficient. An adaptive filter, further comprising means for resetting all stored values.