JPH01220530A - Echo eraser - Google Patents

Abstract

PURPOSE:To improve a convergence characteristic by setting a step gain with prescribed algorithm by providing a memory circuit to store a step gain matrix with prescribed weight. CONSTITUTION:A reception signal is supplied to a correction information generation circuit 15 which performs a processing according to the estimation algorithm of an impulse response by a learning identification method or an LMS method, etc., via reception signal memory circuit 14 and a norm arithmetic circuit 13, respectively. At the time of performing processing, the step gain is set by referring to the step gain matrix weighted by the exponent damping characteristic of the impulse response of a sound field in a step gain matrix memory circuit 12, and a pseudo echo signal for erasing echo is formed. Therefore, estimation for the gradient of a mean square error curved surfaces can be performed with high accuracy, then, an echo erasable device with superior convergence characteristic can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION "Field of Industrial Application" The present invention relates to an echo canceling device for eliminating echoes that cause ringing and are a hearing impairment (mainly in public address communication systems).

``Prior Art'' Popularization of audio conferencing 1: It is desired to provide a loudspeaker communication device that has excellent simultaneous call performance and has less reverberation. There is an echo canceling device that satisfies this requirement. FIG. 5 is a block diagram showing an example of a conventional echo canceller, in which the received signal x
In the loudspeaker communication system, which consists of a receiving system from the receiving terminal 1 to the speaker 2 that receives the receiving signal X(t), and a transmitting system from the microphone 3 to the transmitting output terminal 4, is converted into a sample value, and the received signal x
(k) is supplied to the pseudo echo path 7, and the pseudo echo signal y(k) from the pseudo echo path 7 is subtracted from the echo signal y (kl) sampled by the A/D converter 5 in the subtracter 9. As a result, the echo signal )' (k) is eliminated.

Here, the pseudo echo path 7 needs to follow the temporal fluctuations of the echo path, and is successively estimated by the estimation circuit 6 so that the residual e (k) = y (k) - 9tk) approaches 0. 7 modifications are made to maintain optimal echo cancellation at all times.

The impulse response of this pseudo-echo path 7 is generally modified using the learning identification method (k) = (k), t ('2Φ)...11
'A(k))''X(kl= (, x(k), X(k
-1>-...X (k-N+1)) Tα Nistep gain (scalar amount, constant), N: number of taps, T:
Using the transposition of the vector, the norm of 11×1ji×, and k: discretized time, the impulse response Q tk of the pseudo echo path 7 is brought closer to the impulse response 1 h (kl) of the true echo path.

This learning identification method uses the residual e (k) = y (kl−
Estimate the gradient of the mean square error surface using (k)Y, and use the estimated gradient information to descend the mean square error surface and finally reach the point where the gradient = 0. It is something that will continue.

By the way, Figure 6 shows the time waveforms of two impulse responses measured with speakers and microphones installed in an actual conference room, their differences, and their Schrader method CM, R, 5.
chr6eder t J, A, S, A, 37,
p.

409,. 1965). From FIG. 6, it can be seen that the impulse response of the sound field is exponentially attenuated (attenuated linearly on the logarithmic axis), and the difference therebetween is also attenuated exponentially. Furthermore, it can be seen that the difference (state variation amount) does not depend on the impulse response (state amount). From the above, the fluctuation of the impulse response 1h (kl) of the sound field can be expressed as follows.

lh (k + 1) = lh(k)−+−a −
f(k) (2) However, lh(k)= (h, (k), h2(kl...
...hN(k)) Ta4 = exp (-6,9(
+ 1) Ts/Tu) (i=1, 2...
・N) Ta: Sampling time TR: Reverberation time f(kl= (!1(k), ξ2(k)−=・ξN(
k)) Te3(k), ξ2(k)...ξN(k
) : Noise with mean 0 and variance σ2 Now, the residual e(kl1) used for successive correction of tap coefficients is e(klr) = (+hTtk+B-Q”tk))・X(
kl1), and the impulse response of the pseudo-echo path is (k)
The impulse response of the true echo path 1h (kl) converges to kl (k) + lh (k)
(4), then equation (3) is +2+, (4
Using equation 1, e (kl1 )-4=(lh (kl1
)−1h(k))T−X(kl1)=l−ζmount))
T・X(kl1) =a1ξ1(k)x(kl1)+azξ2(k) x
(k) +...-+aNξN(kIX (k -N)
(5) becomes. Each term is each impulse response coefficient hl(k), h2(kl...h
This is caused by fluctuations in N(k), and ξ, (
kl, ξ2(k)...ξN (kl is 01 on average
It is noise with variance σ2, x(kl1), x(k)...
-x(k-N) are of the same magnitude, then the residual e(k
The proportion of each term in ll) is al, a2...
・It is equal to the rate of aN (exponential decay).

Although the contribution of each impulse response coefficient to the residual e(kl1) is exponentially attenuated, the conventional learning identification method ignores this, and the received signal X(kl1
) was simply weighted. As a result, the estimation of the gradient is incomplete and the coefficient correction contains many errors, resulting in a problem of poor convergence characteristics.

The present invention has been made in view of the above problems, and an object of the present invention is to provide an echo canceling device that can obtain good convergence characteristics by estimating the gradient of the mean square error surface with high precision.

"Means for Solving the Problem" This invention focuses on the fact that the amount of variation in the impulse response of a sound field decays exponentially with the same slope as the exponential decay characteristic of the impulse response, and uses a step gain in an adaptive algorithm using a learning identification method etc. ! It has a step gain matrix storage circuit for setting each tap, that is, each coefficient of the impulse response of the pseudo echo path, and the step gain is weighted by an exponential attenuation characteristic of the impulse response of the sound field. do.

Since this invention incorporates the sound field fluctuation characteristics into the learning identification method as described above, it is based on the assumption that the amount of sound field fluctuation is uniform. The gradient of the mean square error surface can be estimated with high accuracy. Therefore, an echo canceler with good convergence characteristics can be obtained.

"Embodiment" Figure 1 shows an embodiment of this invention, and Figure 5.
- shows the inside of the estimation circuit 6 at
(k) is converted to ×(k) by the reception signal storage circuit 13 and is converted to 11×■11! by the norm calculation circuit 13! is calculated.

The step gain matrix storage circuit 12 stores a step gain matrix section. The step gain matrix is weighted by the exponential decay characteristic of the impulse response of the sound field. X(
kl, II X(k) II2. e (kl,
Ih is supplied to the correction information generation circuit 15 and calculated, and its output is supplied to the adder 16 and added to lh (k) from the tap coefficient storage circuit 11 to obtain Ih.
(kl1) is obtained.

Next, its operation will be explained. In order to reflect the contribution (exponential decay) of each coefficient of the impulse response to the residual e(kll), the step gain matrix Tara, which is a diagonal matrix expanded from the step gain α, which was conventionally given as a scalar quantity, is used αi ” (amax-αmln) eXp(-6,
9 (i 1 ) T8/TR) + "m1n (i = 1.2...N) N: Number of taps: Discretization time) Each coefficient of The diagonal component α!(
i=1.2...N) is shown in FIG.

As i increases, αi exponentially decays from amax to αml with the same slope as the exponential decay characteristic of the impulse response of the sound field.

Asymptotes to .

In addition, when configuring hardware using multiple DSPs (digital signal processors), etc., the exponential attenuation characteristic of the impulse response of the sound field is simplified and applied stepwise to each chip, as shown in Figure 3. , the present invention can be applied without increasing the amount of calculation or storage capacity at all.

Equation (6) can be regarded as having a step gain for each tap in the conventional learning identification method, and weighting the step gain with the exponential attenuation characteristic of the impulse response of the sound field.

As mentioned above, the step gain matrix is a diagonal matrix, and its element α1 (i=l, 2...N) decays exponentially with the same slope as the exponential decay characteristic of the impulse response of the sound field. Therefore, it can be determined if the reverberation time is given. If the reverberation time is not known, the present invention is used after estimating the reverberation time from the tap coefficients obtained through initial training.

The simulation results of this invention are shown in FIG. It is known that in the conventional method, the convergence speed is maximum when α=1.0, and as α becomes smaller than 1, the convergence speed becomes slower and the convergence value becomes larger. In this invention, αmax in FIG. 2 is 1.0. When using the step gain matrix Shira with αm10 = 0.5, α = 1 in the conventional method.
．． It can be seen that the convergence speed when α=0 and the convergence value when α=0.5 are both achieved. In addition, in this invention, amax = 1.5 tαml in FIG.
When using the step gain matrix section with = 0.5, the maximum convergence speed in the conventional method (when α = 1)
It can be seen that a convergence speed exceeding .

Also, amax = 1.9, αml, in Fig. 3,
When using a step gain train with = 0.5, "max = 1-Osαml" in Fig. 2
It can be seen that convergence characteristics that are almost the same as when using a step gain matrix with , = 0.5 can be obtained.

As is clear from the above results, an echo canceler with better convergence characteristics can be obtained when the convergence value is the same compared to the conventional technique. The convergence speed of sequentially correcting the tap coefficient once per sampling time (120μ) is 3 seconds to 2.4 seconds (90% value)
improve.

Although the case where the present invention is applied to the learning identification method has been described here, it may be applied to other algorithms such as the LMS (Least Mean Square) method. Further, the echo path may be a line system.

"Effects of the Invention" As explained above, focusing on the fact that the fluctuation characteristic of the impulse response of the sound field decays exponentially with the same slope as the exponential decay characteristic of the impulse response, the step gain is set at each tap in an adaptive algorithm using the learning identification method. Since the step gain is weighted by the exponential attenuation characteristic of the impulse response of the sound field, the gradient of the mean square error surface can be estimated with high accuracy. When the convergence values are made equal, an echo canceler is obtained that improves the convergence speed by about 20 steps (3 seconds for 2048 taps is 24 seconds (90% value)). Therefore, there is an effect that the quality of speech in loudspeaker calls is improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an explanatory diagram showing the diagonal component αi of the step gain matrix competition,
Figure 3 shows the diagonal component αi of the step gain matrix
4 is a diagram showing simulation results of the present invention and the conventional method, and FIG. 5 is a block diagram showing a conventional echo canceling device. FIG. 6 is an explanatory diagram showing the time waveforms of two impulse responses in an actual sound field, their difference, and their energy attenuation waveforms (Schrader method). Patent applicant: Nippon Telegraph and Telephone Corporation Representative: Takuya Kusano 1 Figure e (phantom χ (4) o 2 Figure 1
N-sai 3 figure

Claims (1)

[Claims] (1) From the transmitted signal to the echo path and the echo signal after the transmitted signal has passed through the echo path, the impulse response of the echo path is sequentially estimated to generate a pseudo echo path, and the transmitted signal is An echo cancellation device that eliminates the echo signal by subtracting from the echo signal a pseudo echo signal obtained by inputting the pseudo echo path, and uses a learning identification method or LMS method as an algorithm for estimating the impulse response of the echo path. has a step gain matrix storage circuit for setting a step gain in the algorithm for each coefficient of the impulse response of the pseudo echo path, and the step gain matrix is weighted by an exponential attenuation characteristic of the impulse response of the sound field. An echo canceling device characterized by: