JP2602750B2 - Echo canceller - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a reverberation canceller, and more particularly to a reverberation signal which causes howling and impairs hearing in a two-wire / four-wire conversion system, a loudspeaker system, and other signal transmission systems. The present invention relates to an echo canceller.

[0002]

2. Description of the Related Art With the spread of satellite communication, voice conference, and other transmission systems, it has been desired to provide a communication device which is excellent in simultaneous communication performance and has little reverberation. An echo canceller meets this requirement. FIG. 1 is a block diagram showing a conventional example of a reverberation canceling apparatus, and shows an example of a loudspeaker call. A receiving input terminal 1 for receiving a transmitting signal, for example, a receiving signal x (t)
In the communication system including the reception system from the speaker 2 to the speaker 2 and the transmission system from the microphone 3 to the transmission output terminal 4, an echo path is formed between the speaker 2 and the microphone 3. Here, the received signal x (t) is sampled binarized by the A / D converter 8, the received signal x (n) is supplied to the echo path 7, the estimated echo signal y ^f from the pseudo echo path 7 ( n) is A
Echo signal y sampled by the / D converter 5
By subtracting from (n) in the subtractor 9, the echo signal y (n) can be eliminated.

Incidentally, the simulated reverberation path 7 needs to follow the temporal variation of the reverberation path as described above. In this conventional example, the estimated echo path 7 constituted by using a digital FIR filter, as residuals e (n) = y (n) -y ^f (n) approaches 0, LMS method, NLMS Estimating circuit 6 which adopts a gradient adaptive algorithm such as ES, ES or projection method
, The residual e (n) is successively corrected by the pseudo echo path 7 through the filter coefficients. By performing the correction of the pseudo echo path 7, the optimal echo cancellation is always maintained.

[0004] Here, the ES method is an LMS method, a learning identification method, a projection method, or other gradient-type adaptive algorithm, ^hhe (n + 1) = ^hhe (n) + α [-Δ (n)] (1 ) However, h ^f (n) = (h ^f ₁ (n), h ^f ₂ (n), ..,
h ^f _L (n)) ^T : pseudo echo path (FIR filter) coefficient Δ (n): gradient vector of (mean) square error α: step size (scalar amount) L: number of taps ^T : transposition of vector n: In the discretization time, the step size α, which has been conventionally given as a scalar quantity, is extended to a diagonal matrix called a step size matrix A. ^hhe (n + 1) = ^hhe (n) + A [−Δ ( n)] (2) 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 (0 <λ <1) It is.

[0005] When the estimated echo path 7 is constituted by a digital FIR filter, the filter coefficient h ^f (n) directly simulates the room impulse response h (n). Therefore, the magnitude of the filter coefficient correction required according to the fluctuation of the echo path matches the fluctuation amount of the room impulse response. Therefore, the step size matrix A representing the correction width in the filter coefficient correction operation is weighted by the temporal variation characteristics of the impulse response. Generally, the amount of impulse response fluctuation in a room sound field is expressed as an exponential function using an attenuation rate λ. As shown in FIG. 2, the diagonal components α _i (i = 1, 2,..., L) of the step size matrix A are indexed from α _{0 as} i increases with the same slope as the exponential decay characteristic of the impulse response. Decays and approaches 0. Details regarding these points are disclosed in Japanese Patent Application Laid-Open No. H2-220530 . This algorithm makes use of the acoustic knowledge that when the impulse response fluctuates due to the movement of a person or an object, the amount of fluctuation of the impulse response (difference in impulse response) exponentially decay at the same rate of decay as the impulse response. is there. The initial coefficient of the impulse response having a large fluctuation is corrected by a large step, and the coefficient of the late part of the impulse response having a small fluctuation is corrected by a small step.

[0006] By applying ES method to the learning identification method, the estimated echo path 7 are sequentially corrected in accordance with equation (3), the impulse response h ^f of the pseudo echo path 7 (n) is the impulse response h of the true echo path approaching (n). However, e (n): estimated error (= y (n) -y ^f (n)) y ^f (n) = h ^{f (n) T x (n)} x (n) = (x (n), x (n-1), ..., x (n-L + 1)) ^T : Received signal vector The projection method removes the autocorrelation of the input signal within the algorithm, and converges on a correlated signal such as speech. It is based on the idea of improving speed. By the projection method, the convergence speed for the audio signal can be improved to about twice that of the learning identification method. The estimated echo path 7 by the secondary projection algorithm is sequentially corrected according to equation (4), the impulse response h ^f of the pseudo echo path 7 (n) approaches the impulse response h of the true echo path (n). h ^f (n + 1) = h ^f (n) + α [β ( n) x (n) + γ (n) x (n-1)] (4) However, β (n) x (n ) T x ( n) + γ (n) x (n-1) T x (n) = e (n) (5) β (n) x (n-1) T x (n) + γ (n) x (n-1) ^{T x (n-1) =} (1-α) e (n-1) (6) α: step size (scalar) beta (n), gamma (n) of the simultaneous equations (5) (6) solution It is a constant that can be obtained by

[0007] The ES method focuses only on the fluctuation characteristics of the acoustic echo path, whereas the projection method focuses only on the properties of the input signal. Therefore, by combining the ES method and the projection method, it is expected to provide an echo canceller with a high convergence speed that makes use of the respective advantages. FIG. 3 shows an example of the inside of the estimation circuit 6 combining the ES method and the secondary projection method.

The received signal x (n) is stored in the received signal storage circuit 14 ₁ ,
In 14 ₂ received signal vector x (n), are x (n-1). Norm operation circuits 13 ₁ , 13 ₂ , 13 ₃ , 13 ₄
X (n) ^T x (n), x (n) ^T x (n-1), x (n-
^{1) T x (n),} x (n-1) T x (n-1) is calculated. The calculated norm, the residual e (n), the residual e (n-1) from the residual storage circuit 17 and the step size α are β (n), γ (n)
It is supplied to the arithmetic circuit 18 to form the simultaneous equations (5) and (6). By solving equations (5) and (6), the constant β
(n) and γ (n) are obtained. Step size matrix storage circuit 1
2 stores a step size matrix A.

A, β (n), γ (n), x (n), x (n-1) are supplied to a correction information generating circuit 15, and A [β (n) x (n) + γ (n) x (n-1)] ( 7) is calculated, and its output is fed to adder 16, which adds to the h ^f (n) from tap coefficient storage circuit 11 h ^f (n + 1) is obtained Can be Calculation result h ^f (n + 1) updates the value of the tap coefficient storage circuit 11 is outputted to the estimated echo path 7.

[0010] By the above operation, the estimated echo path 7 is sequentially corrected according to equation (8), the impulse response h ^f of the pseudo echo path 7 (n) is the impulse response h of the true echo path (n)
Approaching. ^hhe (n + 1) = ^hhe (n) + A [β (n) x (n) + γ (n) x (n-1)] (8) where β (n) and γ (n) are simultaneous It is obtained by solving equations (5) and (6). Equation (8) is obtained by replacing the step size α of the projection method defined by equation (4) with a step size matrix A based on the concept of the ES method.

[0011]

The above-described algorithm that simply combines the ES method and the projection method converges when the input signal is a stationary signal such as a white signal.
In the case of an unsteady signal such as an audio signal, there is a problem that it diverges. The present invention solves this problem. By correctly evaluating the norm, the advantages of the ES method and the projection method can be utilized, and by reflecting both the fluctuation characteristics of the acoustic echo path and the elimination of the correlation of the input signal. Another object of the present invention is to provide an echo canceller having a high convergence speed even when an input signal is an audio signal.

[0012]

[Means for Solving the Problems] A signal x to be transmitted to a reverberation path.
reverberation signal after passing through the reverberation path of (t) and transmission signal x (t)
A pseudo echo path 7 is generated from y (t) and a pseudo echo signal ^yhe (n) obtained by using the transmission signal x (t) as an input of the pseudo echo path 7 is subtracted from the echo signal y (n). In the echo canceller for canceling the echo signal y (n), the pseudo echo path 7 is constituted by a digital filter, and the coefficient of the digital filter is an algorithm which operates so as to minimize the error of the echo signal y (n). Has a first and a second correction width (step size) for adjusting the magnitude of the filter coefficient correction operation, and the first correction width is a function of the variation of the echo path. An echo canceller is provided which is weighted in proportion to the magnitude of the required filter coefficient correction and the norm calculation used in the algorithm is also weighted by the first correction width.

[0013]

An embodiment of the present invention will be described with reference to FIG. First, in the second-order projection method, a correct output y for the past two input signal vectors x (n) and x (n-1) is obtained.
(n), to modify the h ^f (n) to obtain y (n-1). That, x (n) ^T h ^f (n + 1) = y ( n) (9) x (n-1) T h ^f (n + 1) = y ( n-1) (10) (8) if the substituted into equation ^{(9) x (n) T} h ^f (n) + β (n) x (n) T Ax (n) + γ (n) x (n) T Ax (n-1) = y ( n) (11)

[0014] x (n) ^T h ^f (n) = from a y ^f (n) (11) Equation β (n) x (n) T Ax (n) + γ (n) x (n) T Ax become (n-1) = y ( n) -y ^f (n) = e (n) (12). If Likewise substituting equation (8) into equation (10) x (n-1 ) T h ^f (n) + β (n) x (n-1) T Ax (n) + γ (n) x (n- 1) ^T Ax (n-1) = y (n-1) (13) Since h ^f (n) satisfies 1 time preceding equation (9) x (n-1) ^T h ^f (n) = y (n-1) from a (14) formula (13) is beta (n ) x (n-1) ^T Ax (n) + γ (n) x (n-1) ^T Ax (n-1) = 0 (15) Solving the simultaneous equations (12) and (15) yields a constant β
(n) and γ (n) are obtained and substituted into Expression (8) to obtain Expression (9),
(10) satisfying h ^f (n + 1) is obtained the.

Based on the above, the second step size μ
When (scalar amount) is introduced, the expression (1) is obtained except when μ = 1.
4) Carefully h ^f if Kakikudase this invention not be satisfied (n + 1) = h ^f (n) + μA [β ( n) x (n) + γ (n) x (n-1) (16) β (n) x (n) ^T Ax (n) + γ (n) x (n-1) ^T Ax (n) = e (n) (17) β (n) x (n-1) ^{T Ax (n) + γ (} n) x (n-1) T Ax (n-1) = (1-μ) e (n-1) (18) e (n) = y (n) -y ^f ( n) (19) y ^f (n) = h ^f (n) ^T x (n) (20) However, mu: a second step size (scalar quantity).

FIG. 4 shows an example in which the ES method and the secondary projection method are combined as an example of the inside of the estimation circuit 6. Members common to those shown in FIG. 3 are denoted by the same reference numerals. The reception signal x (n) is stored in the reception signal storage circuit 1
At 4 ₁ and 14 ₂ , the received signal vectors x (n) and x (n−
1) Norm operation circuits 19 ₁ , 19 ₂ , 1
In 9 ₃ and 19 ₄ , the norms x (n) ^T Ax (n) and x (n) ^T weighted by the first step size matrix A
Ax (n-1), x (n-1) ^T Ax (n), x (n-1) ^T Ax (n-
1) is calculated. The calculated norm, the residual e (n), the residual e (n-1) from the residual storage circuit 17 and the second step size μ from the step size storage circuit 21 are β
(n) and γ (n) are supplied to the arithmetic circuit 20 and the simultaneous equations (1
7) Construct (18). By solving equations (17) and (18), constants β (n) and γ (n) are obtained.

The step size matrix storage circuit 12 has a first
Is stored. When the estimated echo path 7 is composed of a digital FIR filter, the filter coefficient h ^f (n) has become as simulating the room impulse response h (n) directly. Therefore, the magnitude of the filter coefficient correction required according to the fluctuation of the echo path matches the fluctuation amount of the room impulse response. Therefore, the step size matrix A representing the correction width in the filter coefficient correction operation is weighted by the temporal variation characteristics of the impulse response. Generally, the amount of impulse response fluctuation in a room sound field is expressed as an exponential function using an attenuation rate λ. Diagonal component α _i of step size matrix A (i = 1, 2,..., L)
Exponentially attenuates from α ₀ with the same slope as the exponential decay characteristic of the impulse response as i increases, as shown in FIG.
Asymptotic to zero.

Μ, A, β (n), γ (n), x (n), x (n−
1) is supplied to the correction information generation circuit 22 to calculate μA [β (n) x (n) + γ (n) x (n-1)] (21), and the output thereof is supplied to the adder 16. , is added to h ^f (n) from tap coefficient storage circuit 11, h
^F (n + 1) is obtained. Calculation result h ^f (n + 1) with is output to the estimated echo path 7 updates the value of the tap coefficient storage circuit 11.

By the above operation, the pseudo echo path 7 is calculated by the equation (2)
Is sequentially corrected according to 2), the impulse response h ^f of the pseudo echo path 7 (n) approaches the impulse response h of the true echo path (n). h ^f (n + 1) = h ^f (n) + μA [β ( n) x (n) + γ (n) x (n-1)] (22) If an echo canceller is constituted by a plurality of DSP chips, FIG. As shown in FIG. 5, the exponential decay curve of the step size α _i is approximated stepwise, and a constant α _i
Set. As a result, the present invention can be implemented with a calculation amount and a storage capacity substantially equal to those of the conventional projection method.

FIG. 6 shows a computer simulation result of the convergence characteristic of the present invention. For the computer simulation, the measured impulse response (512 taps, sampling frequency 8 kHz) was used. A voice signal was used as the reception signal, and near-end noise was added to the echo signal so that the S / N ratio was 35 dB. FIG. 6 shows the average value of the convergence characteristics of the echo cancellation amount for 50 times. The solid line indicates the case using the present invention, the one-dot broken line shows the projection method, the two-dot broken line shows the ES method, and the broken line shows the case using the learning identification method.
Here, the step sizes of the respective methods were set so that the amounts of stationary echo cancellation were substantially equal. FIG.
As shown in the figure, the convergence speed at which the echo cancellation amount reaches 20 dB is about twice that of the ES method and the projection method as compared with the learning identification method.
It can be seen that the method of the present invention is about four times as large.

Conference room (Reverberation time 300 ms (500 Hz)
FIG. 7 shows the results of a real-time evaluation experiment of the convergence characteristics performed in step (1) using the residual echo level (= e (n) power level). The acoustic echo canceller used in the experiment is D
The 7 kHz band was divided into two, and each band (sampling frequency: 8 kHz) had 3072 taps (echo cancellation time: 384 ms). Audio signals were used as input signals. It can be seen that the real-time evaluation results (FIG. 7) of the convergence characteristics are almost the same as the computer simulation results (FIG. 6).

In a loudspeaker system, the echo path is often fluctuated due to the movement of people and objects, and it is a great advantage that it can quickly adapt to this. The case where the fluctuation amount of the impulse response of the sound field has an exponential decay characteristic has been described above, but may have another arbitrary fluctuation characteristic. Further, the diagonal component α _i of the step size matrix may be set by approximating the fluctuation characteristics of the impulse response.

Although the description has been given of the case where the FIR filter is used as the digital filter, any other digital filter can be used. Also, the case where the algorithm is the projection method has been described, but another algorithm may be used, and the correction width may be determined according to the magnitude of the variation exerted on the parameter estimated by the variation characteristic of the impulse response.

[0024]

As described above, the second step size is used, and the norm calculation is also weighted by the fluctuation characteristic of the impulse response, thereby making use of the advantages of the ES method and the projection method. Since the fluctuation characteristics of the echo path and the correlation removal of the input signal are both reflected, even if the input signal is a voice signal, it operates stably, and compared to the conventional echo canceller using the learning identification method, the An echo canceller whose convergence speed was increased about four times was obtained. Therefore, the call quality is improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a conventional example of an echo canceller.

FIG. 2 is a diagram showing a diagonal component α _i of a step size matrix A.

FIG. 3 is a block diagram showing the inside of an estimation circuit 6;

FIG. 4 is a block diagram showing the inside of an estimation circuit used in the present invention.

FIG. 5 is a diagram showing an example in which a diagonal component α _i of a step size matrix A is approximated in a stepwise manner.

FIG. 6 is a view showing a computer simulation result of a convergence characteristic.

FIG. 7 is a diagram showing a result of a real-time evaluation experiment of convergence characteristics.

[Explanation of symbols]

 7 pseudo echo path

 ──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-1-220530 (JP, A) JP-A-2-298119 (JP, A) IEICE Technical Report Vol. 92, no. 346, EA92-74 (1992-11-26) P. 41-52 IEICE Technical Report Vol. 92, no. 191, EA92-48 (1992-8-25) P. 9-20

Claims (1)

(57) [Claims] 1. A pseudo echo signal obtained by generating a pseudo echo path from a transmission signal to an echo path and an echo signal of the transmission signal after passing through the echo path, and inputting the transmission signal to the pseudo echo path. In a reverberation canceller that eliminates the reverberation signal by subtracting the reverberation signal from the reverberation signal, the pseudo-echo path is constituted by a digital filter, and the coefficients of the digital filter are sequentially determined by an algorithm that operates to minimize the reverberation error of the reverberation signal. A first correction width (step size) for adjusting the magnitude of the filter coefficient correction operation, wherein the first correction width is required for echo path variations Note that the weighting is proportional to the magnitude of the filter coefficient correction, and that the norm calculation used in the algorithm is also weighted by the first correction width. A reverberation canceling device.