JPH07212278A - Echo erasing device - Google Patents

Abstract

PURPOSE:To stably operate the device even at a silent block such as the partition of words or breathing, to prevent convergent speed from being decelerated and to provide a large echo erasing amount. CONSTITUTION:Concerning an input signal x(n), respective circuits 151-154 respectively conduct the respective norm arithmetic of X(n)TX(n), X(n)TX(n-1), X(n-1)TX(n) and X(n-1)TX(n-1). A circuit 31 detects the noise level of an echo path from the x(n), echo signal y(n) and residual signal e(n) and corresponding to this noise level, a positive constant delta is decided. An arithmetic circuit 18 conducts arithmetic to solve associated equations at the shading method as echo path estimation algorithm by adding delta to that denominator, a circuit 19 prepares correction information by using those arithmetic results beta(n), gamma(n), X(n), X(n-1) and alpha, this is added to an estimated echo path impulse response h(n)' the last time, and a dummy echo path is set as h(n+1)'.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pseudo echo path which is set by estimating the characteristics of the echo path for the echo signal which causes the howling and impairs hearing in a two-wire four-wire conversion system and a voice communication system. The present invention relates to an echo canceling device that cancels a signal by using a signal from Eq.

[0002]

2. Description of the Related Art With the widespread use of satellite communication and voice conferences, it is desired to provide a communication device having excellent simultaneous communication performance and less reverberation. An echo canceller is one that meets this demand. FIG. 7 is a block diagram showing an example of a conventional echo canceller, showing a case of a voice call. An A / D converter is provided in a speech communication system including a reception system that receives a reception signal (input signal) x (t) from a reception input end 1 to a speaker 2 and a transmission system that extends from a microphone 3 to a transmission output end 4. The received signal x (t) is sampled by 5, and the received signal x (n) is sampled.
Is supplied to the pseudo echo path 6, and the pseudo echo signal y (n) ′ from the pseudo echo path 6 is input from the microphone 3 and A /
The echo signal y (n) is canceled by subtracting it from the echo signal y (n) sampled by the D converter 7 by the subtractor 8, and the canceled residual signal is analogized by the D / A converter 9. It is converted into a signal and output to the transmission output terminal 4.

Here, the pseudo echo path 6 needs to follow the temporal change of the echo path 11 from the speaker 2 to the microphone 3. In this configuration example, the pseudo echo path 6 is constructed by using a digital FIR filter, and the residual signal e (n) = y
so that (n) −y (n) ′ approaches 0, for example, the LMS method,
The estimation circuit 12 using a learning identification method, an ES method, a projection method, an ES projection method, or the like sequentially corrects the filter coefficient. By thus modifying the pseudo echo path 6, optimum echo cancellation is always maintained.

The projection method is based on the idea that the autocorrelation of the input signal is removed inside the algorithm to improve the convergence speed for a correlated signal such as a voice signal. Here, a simple second-order projection method will be described. The second-order projection method can improve the convergence speed for a speech signal to about twice that of the learning identification method. The pseudo echo path 6 is sequentially modified according to the following equation (1) by the quadratic projection method, and the impulse response h (n) 'of the pseudo echo path 6 approaches the impulse response h (n) of the true echo path 11. go.

H (n + 1) ′ = h (n) ′ + α [β (n) x (n) + γ (n) x (n-1)] (1) where h (n) ′ = (h ₁ (n) ', h ₂ (n)', ..., h _L (n) ')
^T : Impulse response of pseudo echo path (FIR filter) 6, that is, filter coefficient x (n) = (x (n), x (n-1), ..., x (n-L + 1)) ^T :
Received signal vector α: step size (scalar amount) e (n): residual signal (= y (n) -y (n) ') y (n)' = h (n) ' ^T x (n) L: number of taps ^T of the echo path 6: transpose of a vector n: discretization time second order projection method, the past two input signal vector x
Modify h (n) 'to obtain the correct output y (n), y (n-1) for (n), x (n-1). That is, when α = 1, x (n) ^T h (n + 1) '= y (n) (2) x (n-1) ^T h (n + 1)' = y (n-1) (3 ) Is established, the constants β (n) and γ (n) in Eq. (1) are determined. That is, by substituting equation (1) into equation (2), β (n) x (n) ^T x (n) + γ (n) x (n-1) ^T x (n) = e (n) (4) Becomes Similarly, by substituting equation (1) into equation (3), β (n) x (n-1) ^T x (n) + γ (n) x (n-1) ^T x (n-1) = (1 −α) e (n-1) (5)

The simultaneous equations (4) and (5) are converted into β (n) and γ (n)
If you solve about

[0007]

[Equation 1] Becomes Substituting β (n) and γ (n) into equation (1) yields (2)
H (n + 1) ′ satisfying the equation (3) is obtained. FIG. 8 shows an example of the inside of the estimation circuit 12 using the quadratic projection method. The received signal (input signal) x (n) is received by the received signal storage circuits 14 ₁ and 14 ₂ and received signal vectors x (n) and x
(n-1). Norm calculation circuit 15 ₁ , 15 ₂ , 15 ₃ , 15
_{In 4} respectively, x (n) ^T x (n) and x (n) ^T x (n-
1), x (n-1) ^T x (n), x (n-1) ^T x (n-1) is calculated. These calculated norm and residual signal e (n)
And the residual signal e (n-1) from the residual storage circuit 16 and the step size α from the step size storage circuit 17 are supplied to the β (n), γ (n) operation circuit 18 (6 )
Equation (7) is calculated to obtain constants β (n) and γ (n). These β (n), γ (n) and the above x (n), x (n-1), α
Is supplied to the correction information generation circuit 19 to calculate α [β (n) x (n) + γ (n) x (n-1)] (8), and its output is supplied to the adder 21 and tapped. It is added to h (n) ′ from the coefficient storage circuit 22 to obtain h (n
+1) ′ is obtained. The calculation result h (n + 1) ′ is output to the pseudo echo path 6 and at the same time, the tap coefficient storage circuit 22
Update the value of.

By the above operation, the pseudo echo path 6 becomes (1)
The impulse response h (n) ′ of the pseudo echo path 6 is corrected in accordance with the equation, and the impulse response h (n) of the true echo path 11 is
Get closer to (n). The ES projection method is a method in which the step size α, which is conventionally given as a scalar quantity in the projection method, is expanded to a diagonal matrix called a step size matrix A, and a second step size μ (scalar quantity) is introduced. This is a method that makes use of the respective advantages of the ES method focusing on the fluctuation characteristics of No. 11 and the projection method focusing on the characteristics of the input signal (details are described in Japanese Patent Application No. 4-44649). Here, for simplicity, the second-order ES projection method will be described. The second-order ES projection method can improve the convergence speed for a voice signal to about four times that of the learning identification method. Second ES
The pseudo echo path 6 is sequentially modified by the projection method according to the following equation (9), and the impulse response h (n) of the pseudo echo path 6 approaches the impulse response h (n) of the true echo path 11. h (n + 1) ′ = h (n) ′ + μA [β (n) x (n) + γ (n) x (n-1)] (9) where A = diag [α ₁ , α ₂ , ... , Α _L ]: Step size matrix α _i = α ₀ λ ^i-1 (i = 1, 2, ..., L) λ: Attenuation rate of impulse response fluctuation amount (0 <λ <1) μ: Second step Size (scalar amount) In the second-order ES projection method, correct output y (n), y (n-1) is obtained for the past two input signal vectors x (n), x (n-1).
Modify h (n) 'to obtain That is, x (n) ^T h (n + 1) '= y (n) (10) x (n-1) ^T h (n + 1)' = y (n-1) (11) Determine the constants β (n) and γ (n). Therefore, by substituting equation (9) into equation (10), β (n) x (n) ^T Ax (n) + γ (n) x (n-1) ^T Ax (n) = e (n) (12 ). Similarly, by substituting equation (9) into equation (11), β (n) x (n-1) ^T Ax (n) + γ (n) x (n-1) ^T Ax (n-1) = (1 −μ) e (n-1) (13).

Solving simultaneous equations (12) and (13) for β (n) and γ (n)

[0010]

[Equation 2] Becomes Substituting β (n) and γ (n) into equation (9) gives (10)
H (n + 1) ′ satisfying the equation (11) is obtained. FIG. 9 shows an example of the inside of the estimation circuit 12 using the second-order ES projection method, and the parts corresponding to those in FIG.

The step size matrix storage circuit 24 has a first
The step size matrix A of is stored. Received signal x
(n) is used as reception signal vectors x (n) and x (n-1) in the reception signal storage circuits 14 ₁ and 14 ₂ . Norm arithmetic circuit 2
3 ₁ , 23 ₂ , 23 ₃ , 23 ₄ , respectively, norm x (n) ^T Ax (n) weighted by the first step size matrix A,
x (n) ^T Ax (n-1), x (n-1) ^T Ax (n)
, X (n-1) ^T Ax (n-1) is calculated. The calculated norm, the residual signal e (n), the residual signal e (n-1) from the residual storage circuit 16 and the second step size μ from the step size storage circuit 25 are β (n),
The constants β (n) and γ (n) are obtained by being supplied to the γ (n) calculation circuit 26 and subjected to the calculations of equations (14) and (15). These β (n)
, Γ (n), x (n), x (n-1), μ, and A are supplied to the correction information generation circuit 27 and μA [β (n) x (n) + γ (n) x (n- 1)] (16) is calculated, and its output is supplied to the adder 21 and added to h (n) ′ from the tap coefficient storage circuit 22 to obtain h (n
+1) ′ is obtained. The calculation result h (n + 1) ′ is output to the pseudo echo path 6 and at the same time, the tap coefficient storage circuit 22
Update the value of.

By the above operation, the pseudo echo path 6 becomes (9)
The impulse response h (n) ′ of the pseudo echo path 6 is corrected in accordance with the equation, and the impulse response h (n) of the true echo path 11 is
Get closer to (n).

[0013]

The projection method and the ES projection method exhibit stable operation if the input signal is a stationary signal such as a white signal. However, when a silent interval is generated due to word-to-word separation, breathing, etc., like an audio signal, equations (6) and (7) in the projection method and equations (14) and (15) in the ES projection method are input. Signal vector x (n),
Since x (n-1) becomes zero and division by zero is performed, the divergence occurs and the estimated impulse response h (n) ', that is, the filter coefficient is greatly disturbed. As a result, there is a problem that the amount of echo cancellation decreases in such a silent section. In the actual echo canceller, measures such as stopping adaptation (double-talk control) are taken when there is no input signal.
It is difficult to apply the double-talk control even to a silent period of a minute time, and a solution to it has been strongly demanded.

The present invention has been made in view of the above problems, and it is an object of the present invention to provide an echo canceling device having a large echo canceling amount without slowing the convergence speed even when there is a silent section of a minute time. .

[0015]

According to the present invention, a positive division constant δ for preventing division by zero is stored in the storage unit, and the estimating means uses a calculation formula for solving simultaneous equations included in the projection method or ES projection method. The zero division prevention constant δ is added to the denominator. Further, the magnitude of the constant δ is adaptively determined according to the ambient noise level.

[0016]

Since the present invention is configured as described above, it operates stably even when there is a silent section due to word-to-word segmentation, breathing, etc. like a voice signal, without slowing the convergence speed. An anti-erasing device with a large amount of echo cancellation can be obtained. In the projection method, by adding a positive constant δ to the denominator of equations (6) and (7), computing the following equations (17) and (18) has a silent period of a minute time, and x
Even when (n), x (n-1) becomes 0, division by zero is prevented.

[0017]

[Equation 3] In the ES projection method, by adding a positive constant δ to the denominator of the equations (14) and (15), the following equations (19) and (20) are calculated, and there is a silent period of a minute time, and x ( n),
Even when x (n-1) becomes 0, division by zero is prevented.

[0018]

[Equation 4] If δ is too small, no effect can be expected, and if it is too large, the convergence speed becomes slow. The optimum value of δ is related to the ambient noise level. The optimum value of δ is small when the ambient noise level is low, and the optimum value of δ is large when the ambient noise level is high. Therefore, the magnitude of the constant δ is adaptively determined according to the ambient noise level.

[0019]

FIG. 1 shows an embodiment in which the present invention is applied to the projection method.
The parts corresponding to those in FIG. 8 are designated by the same reference numerals.
In the δ decision circuit 31, the received signal (input signal) x (n), echo
No. y (n) and residual signal e (n) are used to detect the ambient noise level.
Output, and a positive constant (for prevention of division by zero) according to the noise level
Constant) δ is determined. That is, when there is no received signal x (n)
The residual signal e (n) or the echo signal y (n) (sending signal zero) is not reflected.
It was noise on the echo path 11, and the echo signal was sufficiently suppressed.
The residual signal e (n) at time (sending signal zero) is also the noise on the echo path 11.
Almost matches. Level of transmission signal to microphone 3
Is usually well above the noise level in its surroundings. Yo
Therefore, the noise obtained as described above, that is, the residual signal e
Ratio y (n) of signal y (n) from microphone 3 to (n)
 / E (n) corresponds to S / N. Depending on this S / N
The larger this is, the larger the constant δ is determined. With this S / N
The relationship with δ is stored in advance as a table, and the calculated S
It is sufficient to find δ by referring to the table with / N. this
The table will be described later. This is stored in the δ memory circuit 32.
The determined positive constant δ is stored. The received signal x (n) is
Received signal storage circuit 14₁, 14₂Incoming signal vector
x (n), x (n-1). Norm calculation circuit 15₁，
15₂, 15₃, 15_FourThen x (n)^Tx
(n), x (n) ^Tx (n-1), x (n-1)^Tx (n), x (n-
1)^Tx (n-1) is calculated. These calculated nor
The residual signal e (n), the residual signal e from the residual storage circuit 16
(n-1), the step size from the step size storage circuit 17
The positive constant δ from the α and δ storage circuit 32 is β (n),
It is supplied to the γ (n) calculation circuit 18 and calculates equations (17) and (18)
Then, the constants β (n) and γ (n) are obtained. These β (n) and γ (n)
And x (n), x (n-1), α are stored in the correction information generation circuit 19.
Is supplied and α [β (n) x (n) + γ (n) x (n-1)] (8) is calculated, and the output is supplied to the adder 21 and tapped.
It is added to h (n) ′ from the coefficient storage circuit 22 to obtain h (n
+1) ′ is obtained. The calculation result h (n + 1) ′ is the pseudo echo path
6 is output to the tap coefficient storage circuit 22 at the same time.
To update.

By the above operation, the pseudo echo path 6 becomes (1)
The impulse response h (n) ′ of the pseudo echo path 6 is corrected in accordance with the equation, and the impulse response h (n) of the true echo path 11 is
Get closer to (n). FIG. 2 shows an embodiment in which the present invention is applied to the ES projection method, and the portions corresponding to those in FIG. 9 are designated by the same reference numerals.

The δ determination circuit 31 detects the ambient noise level using the received signal x (n), the echo signal y (n) and the residual signal e (n), and a positive constant δ is determined according to the noise level. decide.
The determined positive constant δ is stored in the δ storage circuit 32. The step size matrix storage circuit 24 stores the first step size matrix A. The reception signal x (n) is received by the reception signal storage circuits 14 ₁ and 14 ₂ and the reception signal vector x
(n), x (n-1). Norm operation circuit 23 ₁ , 23
₂ , 23 ₃ , and 23 ₄ , respectively, norm x (n) ^T Ax (n), weighted by the first step size matrix A,
x (n) ^T Ax (n-1), x (n-1) ^T Ax (n),
x (n-1) ^T Ax (n-1) is calculated. From these calculated norm, residual signal e (n), residual signal e (n-1) from residual memory circuit 16, and second step size μ and δ memory circuit 32 from step size memory circuit 25. The positive constant δ of is supplied to the β (n), γ (n) calculation circuit 26 and the equations (19) and (20) are calculated to obtain the constants β (n), γ (n). μ, A, β (n), γ (n), x (n), x (n-1) are supplied to the correction information generation circuit 27 and μA [β (n) x (n) + γ (n) x (n-1)] (16) is calculated, and its output is supplied to the adder 21 and added to h (n) 'from the tap coefficient storage circuit 22 to obtain h (n
+1) ′ is obtained. The calculation result h (n + 1) ′ is output to the pseudo echo path 6 and at the same time, the tap coefficient storage circuit 22
Update the value of.

By the above operation, the pseudo echo path 6 becomes (9)
The impulse response h (n) 'of the pseudo echo path 6 approaches the impulse response h (n) of the true echo path, which is corrected sequentially according to the equation. FIG. 3 shows an embodiment in which the present invention is applied to a subband echo canceller, and parts corresponding to those in FIG. 7 are designated by the same reference numerals.

The subband echo canceller divides a signal into a plurality of frequency bands, provides a pseudo echo path (adaptive filter) in each band, and cancels the echo signal independently in each band. The reception signal x (t) from the reception input terminal 1 is sent to the frequency band division circuit 35 to obtain N signals for each frequency band.
Are divided into real number signals x _k (m) (k = 0, 1, ..., N−1). Similarly, the echo signal y (t) from the microphone 3
Is divided into N real number signals y _k (m) for each frequency band by the frequency band division circuit 36.

The received signal x _k (m) divided into the pseudo echo path 6 _k corresponding to each frequency band is input, and the pseudo echo signal y _k (m) from the pseudo echo path 6 _k is input to the echo signal y _k.
By subtracting from (m) with a subtractor 8 _k , the echo signal y
_k (m) is deleted. Here, the pseudo echo path 6 _k is the echo path 1
It is necessary to follow the temporal change of 1 and the residual signal e _k (m)
= Y _k (m) −y _k (m) is sequentially estimated by the estimation circuit 12 _k using the projection method or ES projection method so that the pseudo echo path 6 _k is corrected so that it approaches 0. Optimal echo cancellation is always maintained.

The residual signal e _k (m) of each frequency band is combined by the frequency band combining circuit 37 into the residual signal e (t) of the entire frequency band. The power of the audio signal is concentrated in the low frequency band and is low in the high frequency band. Therefore, the higher the frequency band, the greater the number of silent sections. As a result, the problem that the amount of echo cancellation decreases due to division by zero becomes more serious in the higher frequency band. Therefore, the constant δ is used in the estimation circuit 12 _k of the _kth frequency band, as in the case shown in FIG. 1 or 2. The magnitude of the constant δ is
It is adaptively determined according to the ambient noise level in each frequency band. The determination of each constant δ is performed by the δ determination circuit 3 in FIG.
It may be performed in the same manner as described in 1. in this case,
A common relation table between S / N and δ is used for each frequency band. By using δ for each frequency band in this way, divergence due to division by zero is prevented in each frequency band, and it is possible to provide an echo canceller with a large echo canceling amount without slowing the convergence speed.

A computer simulation of the convergence characteristic in the device of the present invention was performed. The measured impulse response (512 taps, sampling frequency 8 kHz) was used for the computer simulation. Japanese and English voice signals are used for the reception signal and S / N ratio = 7 for the echo signal.
Near-end noise was added so as to be ˜41 dB. FIG. 4 is a computer simulation result of the convergence characteristic of the echo cancellation amount. In the conventional second-order projection method (solid line), the echo cancellation amount deteriorates to 0 dB or less in the silent section of the voice. On the other hand, in the present invention (broken line), the echo canceling amount of 10 dB or more is held even in the silent section of the voice without slowing the convergence speed. As a result, it can be seen that the amount of echo cancellation after the silent section can be increased by 10 dB or more as compared with the conventional one.

FIG. 5 shows the average value of the convergence characteristics of the echo canceling amount 50 times. It can be seen that the steady echo cancellation amount can be increased by 10 dB or more without slowing the convergence speed as compared with the conventional second-order projection method (solid line). FIG. 6 shows δ when the average echo cancellation amount becomes maximum for 3 seconds after the start of learning as an optimum value and displays it as a relationship with the S / N ratio. δ
Is normalized by the mean of the denominator term and is displayed in decibels with respect to the mean value of the denominator term. The optimum value of δ has a simple correspondence relationship with the S / N ratio, and optimum echo cancellation can be achieved by obtaining the S / N ratio and determining δ according to the relationship of FIG. By obtaining the relationship of FIG. 6 in advance and storing it as a table as described above, δ can be adaptively changed according to the noise level.

Although the zero division preventing constant δ is adaptively changed in the above description, the noise level may be within a certain range depending on the environment in which the echo canceller is used. In such a case, the constant δ according to the average noise level may be fixedly set. In the subband echo canceller, δ in the low band is small and δ in the high band is fixedly set to a large value. In a voice call system, there are many fluctuations in the echo path due to the movement of people, and being able to quickly adapt to this is a great advantage.

Although the FIR filter has been described above as the digital filter of the pseudo echo path, any other digital filter may be used. Further, as the echo path estimation algorithm of the estimation circuit 12, the quadratic projection method or the quadratic ES projection method has been described.
The S projection method may be used.

[0030]

As described above, according to the present invention, in order to prevent division by zero when solving simultaneous equations included in the projection method or ES projection method, a positive constant δ is added to the denominator of the calculation formula. This prevents division by zero even when there is a silent period of a minute time, and further, the magnitude of the constant δ is adaptively determined according to the ambient noise level.
It is possible to obtain an anti-erasing device having a large echo canceling amount, which operates stably even when a silent section is generated due to word-to-word segmentation, breathing or the like like an audio signal, without slowing the convergence speed. Therefore, there is an effect that the call quality is improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an estimation circuit in which the present invention is applied to a projection method.

FIG. 2 is a block diagram showing an embodiment of an estimation circuit in which the present invention is applied to the ES projection method.

FIG. 3 is a block diagram showing an embodiment in which the present invention is applied to a subband echo canceller.

FIG. 4 is an explanatory diagram showing a simulation result of a convergence process of the present invention and a conventional device.

FIG. 5 is an explanatory diagram (50 average values) showing simulation results of a convergence process of the present invention and a conventional device.

FIG. 6 is an explanatory diagram showing a simulation result of the relationship between the optimum value of δ and the S / N ratio.

FIG. 7 is a block diagram showing an example of a conventional echo canceller.

FIG. 8 is a block diagram showing an example of the inside of an estimation circuit 12 using a conventional projection method.

FIG. 9 is a block diagram showing an example of the inside of an estimation circuit 12 using a conventional ES projection method.

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Yoichi Haneda 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Nihon Telegraph and Telephone Corporation

Claims (3)

[Claims] 1. An input signal to an echo path is passed through a pseudo echo path formed by a digital filter to obtain a pseudo echo signal, and the pseudo echo signal is obtained from the echo signal whose input signal has passed through the echo path. The echo signal is subtracted to eliminate the echo signal, and the residual signal obtained by subtracting the pseudo echo signal from the echo signal and the input signal are estimated by the projection method or ES.
In the echo canceller that sequentially estimates the impulse response of the echo path by the projection method and sets it in the pseudo echo path, a storage unit that stores a constant for division by zero is provided, and the estimation means is the projection method or ES. An echo canceller, which is a means for adding the zero division prevention constant to a denominator in a solution calculation of simultaneous equations included in the projection method. 2. The input signal and the echo signal are each divided into frequency bands into a plurality of subbands, and the estimation means, the pseudo echo path, and the storage unit are provided for each corresponding subband to eliminate the echo signal. The echo canceller according to claim 1, wherein the zero division prevention constant of the storage section is set to be larger in a high frequency band than in a low frequency band. 3. The echo canceller according to claim 1, further comprising means for adaptively determining the zero division prevention constant according to the noise level of the echo path and storing the constant in the storage unit. apparatus.