JPH09312597A - Echo canceling method and apparatus for multi-channel audio communication conference - Google Patents

Abstract

(57) [Abstract] [PROBLEMS] Echo cancellation is possible even if there is cross-correlation between channels. In echo cancellation for a multi-channel audio communication conference, received signals of a plurality of channels are reproduced as acoustic signals from a plurality of regenerators, respectively, and a combined vector of the received signal vectors of the channels and at least those vectors. A rearranged reception signal vector is generated when two channels are exchanged, and the reception signal combination vector is input to a pseudo echo signal generation unit that simulates an echo path from the regenerator to at least one sound collector. Generate a reverberation signal, subtract the pseudo reverberation signal from the reverberation signal obtained from the sound collector to obtain the residual reverberation signal, and the received signal vector,
From the relationship between the corresponding residual echo signal and the relationship between the rearranged received signal vector and the corresponding approximate residual echo signal, obtain a modified vector that modifies the pseudo echo path vector that represents the impulse response of the pseudo echo signal generator. .

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a reverberation canceling method for a multi-channel voice communication conference and a reverberation canceling method for canceling a room reverberation signal which is a cause of howling and an auditory obstacle in a communication conference system having a multi-channel reproducing system. It relates to the echo canceller used.

[0002]

2. Description of the Related Art There is an echo canceller in order to provide a loud voice communication system which has excellent simultaneous call performance and little reverberation.
First, a conventional echo canceling method and device structure of the echo canceller for one channel will be described with reference to FIG.
In a voice call, the received signal obtained by the other party's utterance at the receiving terminal 11 is reproduced from the speaker (reproducer) 12 as it is, and before the speaker 12 is sent to the speaker 12, the amplitude and power of the received signal are large. In some cases, the reception signal processing unit 13 that automatically adjusts the gain in accordance with the above is inserted, and the reception signal is reproduced from the speaker 12 after being subjected to some processing. Therefore, in this specification, the reception signal x ₁ (k) is not limited to the reception signal itself from the other party, but refers to the reception signal after being processed when the reception signal processing unit 13 is inserted. And k represents a discrete time. The echo canceller 14 outputs the received signal x ₁ (k) from the speaker 12 to the echo path 1
The reverberant signal y ₁ (k) obtained by being collected by the microphone (sound collector) 16 via 5 is erased. Where the echo signal y
₁ (k) is the impulse response of the echo path 15 at time k
As h ₁₁ (k, n), y ₁ (k) ＝ Σh ₁₁ (k, n) x ₁ (k−n) (1) where Σ represents addition from n = 0 to L-1 It can be modeled as that obtained by simple convolution operation. L is the number of taps, which is a constant set in advance corresponding to the reverberation time of the echo path 15. First, in the reception signal accumulation / vector generation unit 17, the reception signal x ₁ (k)
Are stored up to L-1 time past. The accumulated signal is the received signal vector x ₁ (k), that is, x ₁ (k) = [x ₁ (k), x ₁ (k-1), ..., x ₁ (k-L + 1)] It is output as ^T (2). However, (*) ^T represents the transposition of the vector. The pseudo echo signal generation unit 18 calculates the inner product of the received signal vector x ₁ (k) of the equation (2) and the pseudo echo vector h ^ ₁₁ (k) obtained from the echo path estimation unit 19.

[0003]

[Equation 1] And the resulting pseudo-echo signal y ^ ₁ (k). This inner product operation is equivalent to the convolution operation like the expression (1). The echo path estimation unit 19 generates the pseudo echo path vector h ^ ₁₁ (k) used by the pseudo echo signal generation unit 18. The most general algorithm used for this echo path estimation is the NLMS algorithm (learning identification method). N
In the LMS algorithm, the received signal vector x ₁ (k) at time k and the residual echo signal e ₁ (k), that is, the following expression e ₁ (k) = the difference circuit 21 subtracting the pseudo echo signal from the output of the microphone 16 From y ₁ (k) −y ^ ₁ (k) (4), the pseudo echo path vector used at time k + 1
h ^ ₁₁ (k + 1) is calculated by the following equation.

[0004]

[Equation 2] However, α is called a step size parameter and 0 <α <
It is used for adjusting the adaptive operation in the range of 2. By repeating the above processing, the echo path estimation unit 19 gradually obtains the pseudo echo path vector h ^ ₁₁ (k) and the time series of the impulse response h ₁₁ (k, n) of the true echo path 15. Echo path vector h ₁₁ (k) as an element, that is, h ₁₁ (k) = [h ₁₁ (k, 0), h ₁₁ (k, 1), ..., h ₁₁ (k, L-1)] ^T ( 6), and as a result, the residual reverberation signal e ₁ (k) in equation (4) can be reduced.

Generally, a reproducing system of N ( > 2) channels and M
Elimination of echoes in the case of a communication conference system composed of ( > 1) channel sound collecting system is performed by the configuration shown in FIG. That is, echo cancellers that connect N-channel echo cancellers 22 ₁ , 22 ₂ , ..., 22 _M for processing N input 1 output time-series signals between all N channels on the reproducing side and each 1 channel on the sound collecting side. It is realized as the system 23. In this case, N × M echo paths 15 in the entire system
_nm (1 < n < N, 1 < m < M) exists. The N-channel echo cancellers 22 ₁ , 22 ₂ , ..., 22 _M connected between all N channels on the reproducing side and each one channel on the sound collecting side, which is a structural unit of this system, are shown in FIG. The echo canceller 14 is expanded to have the structure shown in FIG. This is for example B. Widrow and SDStearns, "Adap
tive signal processing, "Prince-Hall, Inc., pp. 19
8-200, (1985). Now, consider an N-channel echo canceller 22 _m whose sound collecting side is connected to the m-th sound collecting channel (1 < m < M). Since the echo signal collected by the m-th channel sound collector 16 _m is obtained by adding all the reception signals of the respective reproduction channels through the echo paths 15 _{1 m} to 15 _Nm on the sound collecting side. , It is necessary to devise an echo path estimation by uniformly evaluating only one residual echo signal e _m (k) for all reproduction channels. First, with respect to the reception signal of each reproduction channel, the reception signal accumulation / vector generation units 17 ₁ , 17 ₂ ,
…, 17 _N , the reception signal vector x ₁ (k) = [x ₁ (k), x ₁ (k-1), ..., x ₁ (k-L + 1)] ^T (7) x ₂ (k ) ＝ [x ₂ (k), x ₂ (k-1), ..., x ₂ (k-L + 1)] ^T (8): x _N (k) ＝ [x _N (k), x _N ( k−1), ..., X _N (k-L + 1)] ^T (9) is generated. However, L is the number of taps and is a constant set in advance corresponding to the reverberation time of each echo path. These vectors are calculated by the vector combiner 24.

[0006]

(Equation 3) Combine with Also in the echo path estimation unit 19 _m , each pseudo echo path vector h ^ for simulating N echo paths between each reproduction channel and the m-th sound collection channel.
_{Combining 1m} (k), h ^ _2m (k),…, h ^ _Nm (k)

[0007]

(Equation 4) Treat as When the NLMS algorithm is used to update the pseudo echo path coupling vector h ^ _m (k),

[0008]

(Equation 5) It is performed like. In the pseudo echo signal generator 18 _m ,
Inner product operation

[0009]

(Equation 6)By the echo signal y collected on the mth sound collection channel_m(k)
Pseudo echo signal for y ^ _mGenerate (k). in this way,
Combine the vectors for each channel into one vector
The basic processing flow is shown in Fig. 1.
It is similar to the one-channel echo canceller.

Among the drawbacks of the conventional echo canceling system used in the communication conference system including the N-channel reproducing system and the M-channel sound collecting system, the problems to be solved by the present invention will be described with reference to specific examples. Explain. As shown in FIG. 4, when a conventional echo canceling system is applied to a stereo audio conference apparatus that collects and reproduces two channels between point A and point B, the echo path on the point B side changes completely. Even if it does not, there is a problem that the echo sound from the point B increases with respect to the utterance each time the speaker moves or changes at the point A side. This is a problem that occurs because the echo path estimation is not performed correctly in the echo canceling system on the B point side.

Therefore, in order to explain this problem, the operation of the device 22b ₁ connected to the first sound collecting channel, which is one of the two-channel echo cancellers, which is a constituent unit of the echo canceller at the point B, is explained. Pay attention to. In the following, in order to clarify the operation of each reproduction channel, the elements for each reproduction channel of the combination vector used in the section of the prior art will be described in a clearly defined form. 2 at the point B side expressed by a vector for each reproduction channel without using a combined vector
The channel reception signal vectors are x ₁ (k) and x ₂ (k). When the echo path vectors of the true echo paths 15 ₁₁ and 15 ₂₁ for each reproduction channel are h ₁₁ (k) and h ₂₁ (k), the echo signal y ₁ collected through these echo paths 15 ₁₁ and 15 _21. (k) is

[0012]

(Equation 7) It is expressed as On the other hand, the pseudo echo signal y ^ ₁ (k) generated in the echo canceller 22b ₁ uses the pseudo echo paths h ^ ₁₁ (k) and h ^ ₂₁ (k) generated in the device,

[0013]

(Equation 8)Is obtained. When the same speaker speaks from point A
Is the received signal vector x₁(k) and x_Twobetween (k)
Have a very strong cross-correlation. x₁(k) and x _Two(k) and
If there is a constant cross-correlation between, then y ^₁(k) = y₁(k) The combined vector [h ^ as the solution of (16)₁₁ ^T(k), h ^
_{twenty one} ^T(k)]^TExist innumerably, x ₁(k) and x_TwoPhase with (k)
Subspace H peculiar to cross-correlation_xTo form Therefore, N
General sequential error minimization algorithm such as LMS algorithm
When the algorithm is used, the connection vector [h ^
₁₁ ^T(k), h ^_{twenty one} ^T(k)]^TIs the subspace H from the initial value_xUntil
Converge to a point where the distance of
₁₁ ^T(k), h_{twenty one} ^T(k)]^TDoes not converge to.

[0014] For ease of description, the constant scalar values p _1, p ₂ and the original signal vector s (k), the received signal vector _{x 1 (k), x 2} (k) is, x ₁ (k ) = P ₁ s (k), x ₂ (k) = p ₂ s (k) (17). Join vector [h ^
_The subspace H _x in which ₁₁ ^T (k), h ^ ₂₁ ^T (k)] can exist is

[0015]

[Equation 9] Can be regarded as the straight line in FIG. 5 that satisfies the above condition, and when the adaptation is started from the initial value 0, the convergence point [h ^ _11p ^T (k),
h ^ _21p ^T (k)] is

[0016]

(Equation 10) Is obtained as. Therefore, when the ratio between the scalar values p ₁ and p ₂ fluctuates, the equation (18) is no longer satisfied, so that the echo cannot be canceled and the echo returning to the other party increases.

[0017]

As can be seen from the above example, when the conventional echo canceller is applied to a communication conference system generally composed of an N-channel reproduction system and an M-channel sound collection system. However, if there is a strong cross-correlation between the received signals of the respective channels, the echo path is not correctly estimated, so that the echo increases every time the cross-correlation between the received signals changes.

An object of the present invention is to provide a conventional multichannel audio communication conference echo canceling system capable of canceling echoes even if there is cross-correlation between received signals as described above. To provide an echo canceling method and device.

[0019]

According to the principle of the present invention, the received signal of each channel is directly reproduced as an acoustic signal from the corresponding channel and input to the echo path, and the output of the echo path is output by the sound collector. In addition to the input / output relationship that is also used in the conventional method that can be actually observed, when some or all of the received signals of each channel are replaced and reproduced as an acoustic signal from a channel different from the corresponding channel An unknown input / output relationship between the input and the output that should be obtained when this input is observed as a reverberant signal by the sound collector through the reverberant path is approximately obtained, and the pseudo input / output relationship is used. It is used to derive the correction vector of the echo path.

The echo canceling method for a multi-channel audio communication conference according to the present invention includes the following steps: (a) Receiving signals of a plurality of channels are reproduced as acoustic signals by a plurality of corresponding reproducers, and (b) From each of the plurality of regenerators, through each echo path, added in at least one sound collector to obtain an observable echo signal, (c)
Accumulate the received signals of the above multiple channels for a predetermined time, generate the received signal vector from the received signals accumulated for each channel, and (d) combine the observed signal vectors of all channels and observe. (E) giving the received signal connection vector to a pseudo echo path having a pseudo echo path connection vector that simulates the connection vector of the echo path vectors represented by the impulse responses of the echo paths. By generating a pseudo echo signal that simulates the observable echo signal at its output, the received signal coupling vector and the corresponding echo signal corresponding to the observable input-output relationship for the plurality of echo paths. And (f) subtracting the pseudo echo signal from the echo signal to obtain a residual echo signal, (g) at least one of the plurality of channels. Interchange when two received signals are exchanged. An approximate pseudo echo signal is obtained when the receive signal combination vector is given to the above pseudo echo path, and (h) when the above received signals are exchanged in the same manner as in step (g) above. The echo signal to be obtained from the sound collector through the plurality of echo paths is obtained by approximation, and this is used as the approximate echo signal, and (i) the approximate pseudo echo signal is subtracted from the approximate echo signal to approximate the residual echo. Signal, (j) observable input / output relationship representing the relationship between the observed received signal combination vector and the corresponding observable residual echo signal, and the replaced received signal combination vector and the approximate residual echo signal The approximate input / output relationship representing the relationship is evaluated, the correction vector is obtained based on the evaluation, and (k) the pseudo echo path vector is corrected by the correction vector.

The echo canceller for a multi-channel audio communication conference according to the present invention includes the following: a plurality of regenerators for reproducing the received signals of a plurality of channels as acoustic signals and outputting the acoustic signals, and at least one transmission channel. The collected sound collector,
The acoustic paths from the plurality of regenerators to the sound collector form echo paths, and the acoustic signals collected by the sound collectors and synthesized are sent to the transmission channel as echo signals. A received signal vector generation unit that holds the received signals of the plurality of channels and temporarily generates each received signal vector, and a received signal vector of the plurality of channels are combined to generate a received signal combination vector. And the received signal combination vector are given, and a pseudo echo that simulates an echo path combination vector that combines echo path vectors representing the impulse response of the echo path from the plurality of regenerators to the sound collector. Pseudo echo signal generation that has a path coupling vector and outputs the pseudo echo signal that simulates the observable echo signal when the received signal coupling vector is input And a subtracter that subtracts the pseudo echo signal from the echo signal from the sound collector and outputs the difference as an observable residual echo signal, and at least two reception signals of the plurality of channels are exchanged and combined. The rearranged reception signal combination vector for generating the rearranged reception signal combination vector obtained by the above, at least the observed reception signal combination vector and the corresponding observable residual echo signal are given, and the rearranged reception signal combination vector is given. To obtain an approximate pseudo-echo signal when given to the pseudo-echo path, obtain an approximate echo signal that should be obtained from the sound collector when the received signal is replaced by an approximate echo signal, and from the approximate echo signal The approximate pseudo echo signal is subtracted to obtain the approximate residual echo signal, and the received signal combination vector and the corresponding observable residual echo signal are obtained. And an observable input / output relationship representing the relationship between the rearranged received signal combination vector and the corresponding approximate residual echo signal, and the approximate input / output relationship is evaluated, and the modified vector is based on the evaluation. And an echo path estimation unit that generates a pseudo echo path vector in which the pseudo echo path vector is corrected by the correction vector and gives the pseudo echo signal vector to the pseudo echo signal generation unit.

[0022]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Principle First, for simplicity, the principle of the present invention will be described in the case where a received signal has two channels. As described above with reference to FIGS. 4 and 5, when the received signals x ₁ (k) and x ₂ (k) are reproduced, these are the corresponding echo paths h _1m (k) and h _2m ( via k),
The input-output relation when the sound is collected as the reverberation signal y _m (k) in the m-th sound collection channel (however, m = 2 is shown in FIG. 4) is

[0023]

[Equation 11] When the cross-correlation between the received signal vectors x ₁ (k) and x ₂ (k) is constant, as described above, the pseudo echo path vector h ^ _1m (k), which satisfies the relationship of the equation (21), h ^ _{2 m} (k) cannot be uniquely determined. Therefore, now, the received signal vectors x ₁ (k), the echo signal in the m pickup channel to the received signal vector x different cross-correlation _{'1 (k), x'} 2 (k) and x ₂ (k) y If ' _m (k) is observed, then equation (21)
alike,

[0024]

(Equation 12) Is obtained. On the other hand, the received signal x ₁ (k), x
₂ (k) and _{x '1 (k), x} ' the echo replica against for the first m pickup channel when each set of ₂ (k) is given y ^ _m, When y ^ _'m,

[0025]

(Equation 13) Becomes Moreover, echo signals y _m (k), and y _'m (k) and the estimated echo signal y ^ _m, y ^' respectively an error signal between the _m e _m (k), and e _'m (k), also, echo path _{h 1m (k), h 2m} (k) and the estimated echo path vector _{h ^ 1m (k), h} ^ 2m (k) and the echo path estimation error vector _{Δh 1m (k), Δh 2m} (k) Then,

[0026]

[Equation 14] Becomes I mean

[0027]

(Equation 15) Is obtained. Although the echo path estimation error vectors Δh _1m (k) and Δh _2m (k) cannot be uniquely calculated from Eq. (27), the minimum norm solution is Δh ^ _1m (k), Δ
h ^ _2m (k),

[0028]

(Equation 16) Is obtained. However, (*) ⁺ represents a generalized inverse matrix. From this, the pseudo echo path vectors h ^ _1m (k), h ^ _2m (k)
Is the minimum norm solution Δh ^ _1m (k), Δh ^ _2m (k) as a correction vector

[0029]

[Equation 17] Can be modified. However, μ _m is a step size parameter. At this time, the correction vector Δh ^
_1m (k) and Δh ^ _2m (k) are the received signals x ₁ (k) and x ₂ (k)
It is possible to have a cross-correlation different from that, and it is possible to estimate the characteristics of the true echo path without depending on the cross-correlation between the received signals. However, the above discussion is based on the received signal x ₁ (k),
x ₂ (k) and the received signal x _'1 different cross-correlations (k),
x _'2 (k) with respect to it than established based on the assumption that the input-output relationship is observable represented by the formula (22), in fact,
This input / output relationship is unknown. Therefore, in the present invention,
A state that approximately satisfies this input / output relationship is used.

First, as the reception signals x ′ ₁ (k) and x ′ ₂ (k), any signals having a cross correlation different from those of the reception signals x ₁ (k) and x ₂ (k) can be used. Therefore, assuming that the input / output relation of the equation (22) is considered to be approximately satisfied, by exchanging the reception signals of the two reception signal channels with each other, x ′ ₁ (k) = x ₂ (k) (31) x Consider the case of reproduction as' ₂ (k) = x ₁ (k) (32). That is, x ₁ (k): x
_When the correlation of ₂ (k) is, for example, 1: 2, by switching channels, x ′ ₁ (k): x ′ ₂ (k) becomes 2: 1 and the cross correlation is different. That is, according to the present invention, in general, when a multi-channel reception signal is reproduced, a correction vector is obtained by solving a combination of signals in which reception signals are exchanged for arbitrary two channels, as shown in Expression (28). Several methods for obtaining correction vectors Δh ^ _1m (k) and Δh ^ _2m (k) using this principle are shown below. Method 1 Method 1 uses the observable input / output relationship expressed by Expression (21) and Expression (22) that is estimated to be approximately obtained when the signals are exchanged by Expressions (31) and (32). Evaluate the input-output relationship of the with different weights. For example evaluates the error signal e _m (k) and e _'m (k) in different weights. There are two types of weighting, one is constant regardless of time and the other is temporal change.

Method (1-a) For example, in the case of reproducing two channels,
To evaluate the error e _m (k) and e _'m (k) at different weights,
Weighting factor gamma _m, introducing gamma _'m,

[0032]

(Equation 18) The correction vector is obtained by solving Method (1-b) The error e ′ _m (k) is separated into a component having a correlation with the error e _m (k) and other components, and then a weighting factor is introduced. That is, for example, focusing on the correlation between the received signals, x ′ _{1
(k), x '2 (} k) is expressed by the following equation

[0033]

[Equation 19] Such as x ₁ (k), x ₂ (k) (equation (3
4) The second term on the right side) and other parts (the first term on the right side of Equation (34))
And can be separated into In addition, add [Δ
h _1m ^T (k), Δh _2m ^T (k)] and using the relations of equations (25) and (26), the error e ′ _m (k) is

[0034]

(Equation 20) Next, separated from the point of view of the correlation of the received signal, the error e _m (k) and a correlation component (formula (37) the second term on the right side), and other components (formula (37) the first term) To be done. From this, the formula
(27) is

[0035]

(Equation 21) Becomes Here, with respect to the error e ′ _m (k), the component correlating with the error e _m (k) and the other components are independently multiplied by the coefficients v and w, respectively, so that the equation (38) further becomes

[0036]

(Equation 22) Then, the correction vector is obtained by solving this. As shown in equation (39), the error e _'m (k), a component having a correlation with the error e _m (k), then separated into the other ingredients, by weighting each component independently , V = 0, w = 1
Then, the obtained correction vector matches the correction vector obtained by the NLMS algorithm described in the prior art. Further, v = 1, if w = 0, of the error e _'m obtained for the approximation input-output relationship (k), the correlation between the error obtained for observable input-output relationships e _m (k) As can be seen from the fact that a certain component, that is, a redundant component is removed and evaluated, and a correction vector is obtained, it is possible to make the physical meaning clearer and perform weighting.

As one of the methods for changing the above weighting factors with time, basically, the degree of simulating the true echo path of the pseudo echo path vector is shown, as shown by the effect of the present invention described later. Change according to. For example, the method (1
In -a) and _{γ m (k) = 1,} γ 'm (k) small city than 1, the gamma when the poor simulation of the true echo path' large within a range that does not exceed the 1 _m (k) And In method (1-b), when the simulation is bad, v = 0 and w = 1 are set to match the conventional method, and the values of v and w are changed as the state improves.

In order to evaluate the degree of this simulation, the ratio of the echo signal y _m (k) to the residual echo signal (that is, the error signal) e _m (k), the average power ratio of the echo signal to the residual echo signal, and , At least one of the absolute value of the residual echo signal and the average power of the residual echo signal is used. Method 2 Equation (28) in the description of the principle of the present invention, or method (1-a) and
In equations (33) and (39) in (1-b), Δh _1m (k), Δ
A specific method for obtaining h _2m (k) will be described below. Input to the echo path when some or all of the received signals of each channel are exchanged and reproduced as an acoustic signal from a channel different from the corresponding channel, and this input collects sound through the corresponding echo path. The unknown input-output relationship with the output that should be obtained when it is observed as a reverberant signal by the instrument is obtained by approximation. In order to perform this approximation, here, the received signal x _m (k) of each channel is reproduced as an acoustic signal as it is, and the known input / output relationship actually observed by the sound collector via the corresponding echo paths, The similarity of the residual echo signal (error signal) and the echo path used for the evaluation of the input / output relationship is used. The specific method is as follows.

[0039]

(Equation 23) Method (2-a) Method (2-a) is a method that particularly utilizes the similarity of the echo paths, that is, a regenerator that has a similar physical layout relationship such as the distance between the sound collectors and the sound collectors, and the orientation. The echo paths between the sound collectors are assumed to be equal to each other. For example, if both playback and sound collection are 2 channels, the input / output relationship for each sound collection channel is

## EQU4 ## Here, as shown in FIG. 6, when the speakers 12 ₁ and 12 ₂ and the microphones 16 ₁ and 16 ₂ are symmetrically arranged including these directional characteristics, h ₁₁ (k) = h ₂₂ (k) (42 ) h ₂₁ (k) = h ₁₂ (k) (43) holds, and from these, and equations (40) and (41),

[0041]

(Equation 24) Is obtained. These are exactly the input / output relationships when the reproduction channels of the received signal are exchanged with respect to the equations (40) and (41), and therefore, e ′ ₁ (k) = e ₂ (k) and e ′ ₂ (k) ＝ e
₁ (k) holds. When these approximation results are applied to Eqs. (28), (33) and (39) together with Eqs. (44) and (45), the following three methods are obtained. Method (2-a-1) That is, if the channel replacement and the echo path approximation are applied to Equation (28), the following simultaneous equations are obtained.

[0042]

(Equation 25) And the correction vector can be obtained by solving this. Method (2-a-2) That is, applying the channel swapping and the echo path approximation to Equation (33) gives the following simultaneous equations.

[0043]

(Equation 26) And the correction vector can be obtained by solving this. Method (2-a-3) c (k) = r '(k) r ^-1 (k), and applying channel permutation and echopath approximation to Eq. (39), the following simultaneous equations are obtained.

[0044]

[Equation 27] And the correction vector can be obtained by solving this. Method (2-b) Another method (2-b) is a regenerator having a similar physical arrangement as in the above method (2-a), for example, equations (42) and (43). This is a method that can be applied even when the assumption that the echo paths between sound devices are equal to each other does not hold with high accuracy. Similar to the above example, when both reproduction and sound collection are two channels, instead of equations (42) and (43), h ₁₁ (k) = h ₂₂ (k) + f ₁ (k) (49) h ₂₁ (k) = h ₁₂ (k) + f ₂ (k) (50) is used. Where f ₁ (k) and f ₂ (k) are the similarity error of the true echo path vectors h ₁₁ (k) and h _{2 2} (k) and the true echo path vectors h ₁₂ (k) and h ₂₁ ( It represents the similarity error of k). On the other hand, the similarity error f ^ ₁ (k) between the pseudo echo path vectors h ^ ₁₁ (k) and h ^ ₂₂ (k), and the pseudo echo path vectors h ^ ₁₂ (k) and h ^ ₂₁ (k). The similarity error f ^ ₂ (k) of h ^ 11 (k) = h ^ 22 + f ^ 1 (k) (51) h ^ 21 (k) = h ^ 12 (k) + f ^ 2 (k) (52), and the difference between the similarity error vectors of the true echo path and the pseudo echo path is expressed by the following equation Δf ₁ (k) = f ₁ (k) −f ^ ₁ (k) (53) It is defined by Δf ₂ (k) = f ₂ (k) −f ^ ₂ (k) (54), and these are called similarity error differences.
Therefore, by changing the reproduction channels, the received signal x ₁ (k),
When x ₂ (k) is reproduced, the output obtained in the first sound collection channel is y ′ ₁ (k) and the output obtained in the second sound collection channel is y ′.
₂ (k), by using the relationship between equations (49) and (50) and equations (40) and (41),

[0045]

[Equation 28] Is obtained. The outputs y ′ ₁ (k) and y ′ ₂ (k) are pseudo echo path vectors h ^ _{1 1} (k), h ^ ₂₁ (k) and h ^, respectively.
_{Using 12} (k), h ^ ₂₂ (k) and equations (51) and (52), y ^ ' ₁ (k) and y ^' ₂ (k) are simulated as in the following equation.

[0046]

(Equation 29) _{y '1 (k), y} ' 2 (k) and they simulated the _{y ^ '1 (k),} y ^' 2 (k) error e between the _{'1 (k), e'} 2 (k) is ,

[0047]

[Equation 30] However,

[0048]

[Equation 31] It is. From equations (63) and (64), the similarity error difference Δf ^
_{Since 1} (k) and Δf ^ ₂ (k) are represented by unknown similarity error vectors f ₁ (k) and f ₂ (k), respectively, e ′ ₁ (k) and e ′
_We cannot know ₂ (k) exactly. Therefore, equations (59) and (6
Solving (0) for known e ₂ (k) and e ₁ (k),

[0049]

(Equation 32) Becomes On the other hand, from equation (25),

[0050]

[Expression 33] It is. Therefore, from equations (65) to (68),

[0051]

(Equation 34) The following simultaneous equations are obtained. The modified vector of the pseudo echo path vector is obtained as the minimum norm solution of the simultaneous equations of Equation (69). Here, the similarity error difference Δf ^ ₁ (k), Δ
f ^ ₂ (k) does not actually have to be obtained. Method (2-b-1) When obtaining the pseudo echo path correction vector based on Equation (69), introducing the weighting coefficient in Equation (33) of Method (1-a) above,

[0052]

(Equation 35) Becomes Here, according to equations (59) and (60),

[0053]

[Equation 36] The weighting coefficient is applied as Method (2-b-2) This method (2-b-2) introduces a time-invariant or time-varying weighting factor that makes it possible to adjust the influence of the echo path similarity error vector. For example, if both playback and sound collection are for 2 channels,
The evaluations of the similarity error differences Δf ^ ₁ (k) and Δf ^ ₂ (k) are calculated as Δh ^ ₁₁ (k), Δh ^ ₂₁ (k), Δh ^ ₁₂ (k), Δ.
We weight the evaluation of h ^ ₂₂ (k). For example, equation (63)
Introducing the weighting factors κ ₁ , κ ₂ , κ ₃ , and κ ₄ into each element multiplied by Δf ^ ₁ (k) and Δf ^ ₂ (k), respectively,

[0054]

(37)Becomes In this case, from equations (46) and (47), ‖h ^₁₁(k)-
h ^_{twenty two}(k) ‖ / ‖h ^₁₁If (k) ‖ is small, f₁(k) is
Since it is estimated to be small, for example, a predetermined threshold K1
Less than κ₁ Is controlled to be less than 1, for example
I do. Similarly, ‖h ^_{twenty one}(k) -h ^₁₂(k) ‖ / ‖h ^
_{twenty one}If (k) ‖ is smaller than a predetermined threshold K2, then κ _Two For example
If it is less than 1, control is performed.

Method (2-b-3) In this method, in the formula (69) based on the method (2-b), the method (1-b
Introducing weighting factors v ₁ , v ₂ , w ₁ , w ₂ based on b), c
Setting (k) = r '(k) r ^-1 (k),

[0056]

(38)

[0057]

[Equation 39] Then, the correction vector can be obtained by solving the simultaneous equations (73). Method 3 The third method combines the observed input / output relationship obtained in the past and the input / output relationship obtained by approximation with the observed input / output relationship at the current time and the input / output relationship obtained by approximation. Then, the modified vector of the pseudo echo path vector is derived.

Method (3-a) Specifically, the projection algorithm is applied to the above-mentioned methods (2-a) and (2-b) using channel switching. The projection algorithm simultaneously satisfies the input / output relationship at the past p _m time including the present of the mth received signal channel,
This is a method of correcting the pseudo echo path vector. However, p _m is an integer that does not exceed the tap number L of the pseudo echo path vector. First, collectively received signal vector of the n-channel from the current time to p _m time ^{_{past, X n (pm) (k}} ) = [x n (k), x n (k-1), ..., x n (k-p + 1)] (77). Again, when the above-mentioned reproduction and sound collection are both for two channels, regarding the error signal,

[0059]

(Equation 40) And applying the projection algorithm to equation (73),

[0060]

[Equation 41] Becomes Here, the parameters that determine up to what time in the past the input / output relationship is used are min {p ₁ , p ₂ } ≧ max {p ' ₁ ,
Like p ₁ , p ′ ₁ , p ₂ , and p ′ ₂ that satisfy p ′ ₂ }, they can be set independently within a range that does not exceed the number L of taps of the pseudo echo path vector.

Method (3-b) Apply the projection algorithm using Equations (77) to (81) to Equation (73) based on Method (2-b-3), and calculate the matrix for cross-correlation.

[0062]

(Equation 42) If you put it

[0063]

[Equation 43] Get. A correction vector is obtained as the minimum norm solution of this simultaneous equation (84). However, the received signal vectors of the n-th channel up to the p time past are collected and X _n ^(p) (k) = [x _n (k), x _n (k-1), ..., x _n (k-p +1)] (85) and expressed, the parameter that determines whether to use the input-output relationship until past several times, and _{p 1, p '1, p} 2, p' 2, which are min {p _1, met _{p 2} ≧ max {p '} 1, p' 2},

[0064]

[Equation 44] It is. Method (3-b-1) In addition, with respect to the equation (84) of the above method (3-b), it is possible to adjust the effect of the echo path similarity error vector of the method (2-b-2) in time invariant or time. Introducing variable weighting factors κ ₁ , κ ₂ , κ ₃ , and κ ₄ as in the following equation as in equation (82),

[0065]

[Equation 45] The correction vector may be obtained as Method (3-b-2) Furthermore, the weighting factors λ ₁ and λ ₂ are introduced into the equation (94) as follows,

[0066]

[Equation 46] As weighting factors κ ₁ , κ ₂ , κ ₃ , κ ₄ and weighting factors λ ₁ ,
The correction vector may be obtained by adjusting the weighting depending on the balance with λ ₂ . Method 4 In each of the various methods described above, an ES algorithm that can give a different correction weighting coefficient to each tap based on the exponential decay of the impulse response of the echo path can be applied.

For example, the equation (33) in the above-mentioned method (1-a)
about,

[0068]

[Equation 47] Then, equation (33) becomes

[0069]

[Equation 48] And the minimum norm solution of this system of equations is

[0070]

[Equation 49] Given by However, I is an identity matrix, and δ is a small constant for stabilizing the inverse matrix operation. To apply the ES algorithm to this,

[0071]

[Equation 50] _{^{However, α (r) i = α}} (r) (λ (r)) i-1, i = 1, ..., L (100) 0 <λ (r) Basic index represented by ≦ 1 decay matrix A ^{( r)} Alternatively, the L number of the sequence given by equation (100) is divided into equal intervals or unequal intervals in the block, the value of each alpha ^(r) _i in each block, the head of the same block alpha ^{( r)} _i
The basic exponential damping matrix A ^{(r) is} simplified by making it equal to the value of
make. And

[0072]

(Equation 51) And applying this to equation (98),

[0073]

[Equation 52] As a result, the correction vector of each pseudo echo path vector is obtained by the obtained minimum norm solution. Similarly, for example, with respect to the equation (82) in the method (3-a) to which the above projection algorithm is applied,

[0074]

(Equation 53) Then, equation (82) becomes

[0075]

(Equation 54) And the minimum norm solution of this system of equations is

[0076]

[Equation 55] Given by To apply the ES algorithm to this, the following matrix consisting of the basic exponential decay matrix A ^{(r) of} equation (99)

[0077]

[Equation 56] And apply this to equation (105) to

[0078]

[Equation 57] As a result, the correction vector of each pseudo echo path vector is obtained by the obtained minimum norm solution. Similarly, it is apparent that the ES algorithm can be applied to other methods as well, and a description thereof will be omitted. As described above, in the multi-channel echo canceling method according to the present invention, some or all of the multi-channel received signals are exchanged, and a virtual input when reproduced as an acoustic signal from a channel different from the corresponding channel , Approximately, an unknown input / output relationship between the virtual input and the virtual output that should be obtained when the virtual input is observed as a reverberation signal by the sound collector through the corresponding reverberation paths, and the approximate input / output relationship is obtained. Both the input and output relationships that are also used in the conventional method and the received signal of each channel is reproduced as an acoustic signal as it is and actually observed by the sound collector through the corresponding echo path are both evaluated. Since the correction vector of the pseudo echo path is obtained, the estimated speed of the true echo path in the echo path estimation unit can be increased. Echo Canceller Next, an embodiment of the echo canceller for carrying out the above-described method of the present invention will be described in comparison with a conventional device configuration. First, regarding the overall configuration, in the case of constructing an echo canceling system having an N-channel receiving system and an M-channel transmitting system, in the conventional method, as shown in FIG. The M-channel echo cancellers 22 ₁ , 22 ₂ , ..., 22 _M are provided to constitute the echo canceller. On the other hand, when the method according to the present invention is configured as an apparatus, as shown in FIG. 7, an echo canceling system having an N channel receiving system and an M channel transmitting system is used.
As shown in FIG. 8, it is configured as a × M channel echo canceller 23, and the echo path estimation can be performed by exchanging information between the channels.

The structure of the echo canceller of FIG. 8 shows the basic structure applied to any of the embodiments of the methods 1 to 4 according to the present invention described above, and of the conventional type shown in FIG. 3 as described below. Compared with the echo path estimation unit 19 _m , it is characterized in that a vector rearrangement / combination unit 41 is provided. In this vector rearrangement / combining unit 41, the arrangement of the reception signal vectors x ₁ (k), x ₂ (k), ... Is rearranged by exchanging the reception signal vectors with other channels, so that there are D arrangements. Then, the D combination vectors x ⁽¹⁾ (k), x obtained by combining them are obtained.
⁽²⁾ (k), ..., x ^(D) (k) are generated, and using the similarity of the echo paths, the connection vectors x ⁽¹⁾ (k), x ⁽²⁾ (k), ...,
It is possible to obtain the pseudo echo path correction vector Δh ^ ₁ (k), ..., Δh ^ _M (k) by approximating the input-output relationship with respect to x ^(D) (k).

As shown in FIG. 8, the N × M channel echo canceller has N reception signal storage / vector generation units 17 _{1 to} 17 _N corresponding to N channels. Similarly to the vector generators 17 _{1 to} 17 _N , the L reception signals up to the present time (k) determined for each reception signal channel are held, and the reception signals represented by the equations (7) to (9) are held. Generate signal vectors x ₁ (k) to x _N (k). Further, the received signal vectors x ₁ (k) to x _N (k) are combined by the vector combination unit 24 to generate the combined vector x (k) represented by the equation (10), and the pseudo echo signal generation unit 18 Give ₁ to 18 _M each. The pseudo echo signal generation unit 18 _m (m = 1, ..., M) for each m-th sound collection channel uses the given combination vector x (k) and pseudo echo path combination vector h ^ _m (k) to obtain an equation ( The pseudo echo signal y ^ _m (k) is calculated by 13), and the subtractor 21
give to _m . Each subtractor 21 _m, the difference of the echo signal y _m (k) and the estimated echo signal y ^ _m (k) the residual echo signal (or error signal) e _m
Output as (k).

As described above, the echo canceller to which the method of the present invention is applied is characterized by the vector rearrangement / combining unit 41, and the reception signal vector x ₁ (k) to which N channels are given. The array of x _N (k) is arranged by rearranging the receiving signal vectors of the channels, and the D kinds of receiving signal vector arrays obtained by combining are arranged by the equation (10) in the same manner as the vector combination unit 24 to obtain D. The individual connection vectors x ⁽¹⁾ (k), x ⁽²⁾ (k), ..., X ^(D) (k) are generated and given to the echo path estimation unit 19. The echo path estimation unit 19 uses the D combined vectors x ⁽¹⁾ (k) to x ^(D) (k), the combined vector x (k) from the vector combining unit 24, and the subtractor 21 ₁
To 21 and the error signal _{e 1 (k) ~e M (} k) from _M, echo signal as indicated by a broken line _{y 1 (k) ~y M (} k) is given as necessary, various previously described The pseudo echo path correction vector Δh ^ ₁ (k) to Δh ^ _M (k) is calculated by any one of the above methods, and the next time point (k + 1) is calculated in the same manner as the equations (29) and (30). Pseudo echo path vectors h ^ ₁ (k + 1) to h ^ _M (k + 1) are obtained.

[0082] Figure 9 is the error signal e _m (k) and e _'m (k) coefficient to γ _m, γ' aforementioned methods (1-a) for weighting the _{m, (2-a-1} ),
The functional block diagram of the echo path estimation unit 19 when performing (2-a-2), (2-b-1), (2-b-2), (3-a), etc. is shown. Combined vector x ⁽¹⁾ (k) from the vector rearrangement / combining unit 41
x ^(D) (k) and the combined vector x (k) from the vector combination unit 24 are supplied to the correction vector calculation unit 19A and also to the weighting coefficient addition unit 19B.

[0083] weighting factor controller 19E are respectively introduced weighting factor gamma _m, the gamma _'m between the approximate input-output relationship between the measured input-output relationship according to the method (1-a), the weighting coefficients, estimated echo The weight is added to the weighting coefficient adding units 19B, 19C, 19D by changing the path vector according to the degree of simulation of the true echo path. That is, the weighting factor controller 19E, the average thereof in takes in each sound collecting channel with echo signal y ₁ in this example (k) ~y _M (k) an error signal _{e 1 (k) ~e M (} k) Power ratio r _pm = Av | y
_m (k) | ² / Av | e _m (k) | ² (where Av represents the average value) is calculated as an index representing the simulated degree of the pseudo echo path vector, and the weighting factor γ _m , determine the γ _'m. Equation (33), (7
Γ _m in 0), (73), or (82) is, for example, always set to γ _m = 1 and if the power ratio r _p becomes larger than a predetermined positive value r _th , the weighting factor γ ′ _m Is set to a large positive predetermined value that does not exceed 1, and if r _pm ≦ r _th , the weighting coefficient γ ′ _m is set to a small positive predetermined value that does not exceed 1. The weighting factor evaluation and control unit 19G is used in the case of the method (2-b-2), and based on the correction vector Δh ^ (k), the echo path similarity error vector Δ
Weighting factors κ ₁ , κ ₂ , κ ₃ , κ ₄ for adjusting the influence of f ^ are determined and given to the weighting factor adding unit 19B. Weight coefficient adding unit 19B, as shown in equation (73), the received signal vector x ^T _1, x ^T ₂ weighting factor κ ₁ ~κ ₄ for gamma _'1, gamma' ₂
The weight is given by the combination of
Give to 9A. The weighting factor adding units 19C and 19D apply the weighting factors γ ′ ₁ , γ ′ ₂ and γ _{1 to} the error signals e ₁ (k) and e ₂ (k),
γ ₂ is given to the correction vector calculation unit 19A. The correction vector calculation unit 19A uses the given signal to calculate the equation (3
3), (70), (73), (84) is solved to obtain the correction vector Δh ^ _mn (k), and the adaptive filter updating unit 1
Give to 9F. The adaptive filter updating unit 19F uses the equations (29) and (3
(0), the pseudo echo path vector is modified, and the time (k + 1)
Is output as the pseudo echo path vector of the above and is given to the pseudo echo signal generation units 18 ₁ to 18 _M of FIG.

[0084] Figure 10, the aforementioned methods for weighting coefficients v and w are divided error signal e _'m (k) to the component (1-b), (2 -a-3),
A functional block diagram of a multi-channel echo canceller when performing (2-b-3), (3-b-1), and (3-b-2) is shown. In this example, as in the case of FIG. 10, the weighting factor evaluation / control unit 19G
Determines the weighting coefficients κ _{1 to} κ ₄ which adjust the influence of the similarity error difference Δf ^ (k) of the echo path based on the correction vector Δh ^ (k), and gives it to the weighting coefficient adding unit 19B. The correction vector calculation unit 19A includes the received signal combination vector x (k)
And the received signal combination vector x with the channels replaced
⁽¹⁾ (k), ..., X ^(D) (k) are given. Wherein the error signal e _'m (k) in this embodiment (39), a component having a correlation with (74), (84), (94) or e _m (k) as indicated by (95), In order to give the weighting factors w and v separately to the other components,
In the weighting factor control unit 19E, the echo signal
From y ₁ (k) to y _M (k) and error signal e ₁ (k) to e _M (k), their average power ratio in each sound collection channel r _pm = Av | y _m (k) | ² / Av | e
The value of w and v is determined by calculating _m (k) | ² as an index representing the degree of simulation of the pseudo echo path vector. Further, the cross-correlation calculation unit 19H calculates the matrix c (k) = r 'regarding the cross-correlation of the received signal.
(k) r ^-1 (k) or C ₁ (k), C ₂ (k) in the equations (86) to (89) is obtained and given to the component separation unit 19J.

[0085] As the component separating unit 19J shown in formula (87), the error signal e _'m (k) is separated into components and other components that are correlated with e _m (k), respectively weighting factor Additional unit 19K,
19L and the weighting factors w and v. The correlation component to which these weighting factors are given and the other components are added by the adder 19M, and the correction vector calculation unit 19A
Given to. The correction vector calculation unit 19A calculates the simultaneous equations (37), (74), (84), (94) or (95) from the given signal.
To obtain the correction vector Δh ^ (k), which is applied to the adaptive filter updating unit 19F and the weighting coefficient evaluation and control unit 19G. The adaptive filter updating unit 19F calculates and outputs the pseudo echo path vector by the equations (29) and (30) based on the given correction vector.

FIG. 11 shows that in each echo path estimation unit 19 shown in FIGS. 9 and 10, the echo path has good symmetry, and therefore the method
When (2-a) is used, the functional block configuration of the correction vector calculation unit 19A which makes it possible to apply the p-th order projection algorithm of method 3 and the ES algorithm of method 4 is shown.
In order to implement the p-th order projection algorithm, the received signal vector is given to the signal storage unit 19A2, the past (p-1) received signal vectors are held, and they are a set of the received signal vector of k at the present time. Received signal matrix X shown in equation (85)
^(p) _n is given to the exponential weighting coefficient adding unit 19A1, and equations (99) and (10
The weighting coefficient is added by the basic exponential damping matrix A ^(r) represented by 0) and is given to the simultaneous equation calculation unit 19A3. At the same time, the reception signal vector weighted by the weighting factor addition unit 19B of FIGS. 9 and 10 is also given to the simultaneous equation calculation unit 19A3.

On the other hand, the error signal e weighted by γ _m
Similarly, the error signals e ₁ (k), ..., E _M (k) weighted by ₁ (k), ..., E _M (k) and γ ′ _m or v, w are also stored in the signal accumulating units 19A4 and 19A5 in the past. A pair of error signals at (p _m -1) time points are respectively held and given to the simultaneous equation calculation unit 19A3 together with the error signal at the current time point. The simultaneous equation calculation unit 19A3 solves the simultaneous equations (47) from these given signals and corrects the correction vector Δh ^.
Calculate (k) and output. In the case of FIG. 10, the past (p _m -1)
The set of the measured error signals e ₁ (k), ..., e _M (k) at the time point is the signal storage unit 19
It is held in A4 and given to the simultaneous equation calculation unit 19A3 together with the set of actual measurement error signals at the present time, and the set of component separation weighted approximate error signals from the adder 19M is given to the signal storage unit 19A5 and past (p _m- 1) The set of weighted approximation error signals at the time point is retained. The (p _m -1) sets of the weighted approximation error signals and the current error signal are given to the simultaneous equation calculation unit 19A3. The simultaneous equation calculation unit 19A3 uses, for example, an equation from these signals.
Solve (74) to get the correction vector.

FIG. 12 shows a case where a correction vector is obtained by the method (2-b) in the echo path estimation unit 19 shown in FIGS. 9 and 10 in consideration of the poor symmetry of the echo path (that is, similar error). The functional block configuration of the correction vector calculation unit 19A is shown. Also in this case, the p-th order projection algorithm of the method 3 and the ES algorithm of the method 4 can be executed. The difference from FIG. 11 is that the simultaneous equation calculation unit 19A3
Is a system of equations (69), (70), (7
3), (82), (84), (94), (95) to obtain a correction vector, and the weighting coefficient evaluation / control unit 19G in FIGS. Δ
The point is that a signal accumulating unit 19A6 and an exponential weighting coefficient adding unit 19A7 are further provided for the reception signal vector given a weighting coefficient for adjusting the influence of f ^ (k). Other configurations are the same as those in FIG. 11, and thus description thereof will be omitted.

[0089]

According to the method of the present invention, by using a multi-channel transmission system and a terminal device having a multi-channel sound collection / reproduction system between two points, it is possible to transmit both acoustic spatial information. It can be applied to eliminate echoes when realizing high-presence voice communication.

For simplification, as shown in FIG.
Connected by a channel transmission system, a two-channel sound collection / playback system
When performing stereo voice communication using your own terminal device
think of. For example, one person from different points from point A
If you speak from each seat,
Each speaker and two micros are included in the Leo audio signal.
There is a cross-correlation that depends on the positional relationship of the phones. Those sounds
When the voice signal is reproduced at the B point side, the conventional echo
In the case of using the canceller, the echo path estimation is performed on the stereo signal.
Because it depends on the cross correlation of the
A big echo returns to the point side. However, this invention
In the echo cancellation method, the stereo signal in the echo path estimation is
The echo path can be estimated faster by reducing the effect of cross-correlation of
Therefore, the increase in the reverberation that occurs every time the speaker is replaced is reduced.
it can. Fig. 13 shows the spoken voice with the body and head fixed.
Computer stain using the signal collected by tereo as a reproduction signal.
And (b) this invention.
Of cross-correlation due to previous application by the applicant (Japanese Patent Application No. 7-50002)
Method by variation extraction, and (c) method and odor of the present invention
, The pseudo echo path coupling vector generated by the echo path estimation unit.
And the magnitude of the error vector between the true echo path coupling vector and
This is a comparison when the echo path estimation start time is 0. This
Here, if the conventional method uses the NLMS algorithm,
In the method of extracting cross-correlation variation according to the previous application, the method
This is the case when the projection algorithm of is used. This invention method
Then, when the correction vector is obtained by Eq. (107),
, Γ₁, Γ '₁, Γ_Two, Γ '_TwoIs time-varying according to the amount of echo cancellation,
κ₁= Κ_Two= Κ_Three= Κ_Four= 1.0, p₁= P ' ₁= P_Two= P '_Two= 2, A
⁽¹⁾= A⁽²⁾= A⁽³⁾= A^(Four)= A^(Five)= A⁽⁶⁾so
Both are divided into 4 blocks [1.0, 0.5, 0.25, 0.125]
Weighted. As is clear from this result,
The light method causes the echo path error vector to occur more rapidly than other methods.
Has decreased, and its size has also decreased. Follow
Can be said to be effective.

In the method of the invention using equation (70),
FIG. 14 shows the results of examining the effects of γ ′ ₁ and γ ′ ₂ with γ ₁ = γ ₂ = 1. In FIG. 14, the curve (a) is γ ′ ₁ = γ ′ ₂ = 1.0, and the curve (b) is γ ′ ₁ = γ ′ ₂ = 0.5. By comparing these, depending on the convergence state of the filter coefficient, appropriate γ _'1,
γ _'2 of the value is found to be different. Therefore echo cancellation amount by γ _'1, γ' to change the value of _2, i.e. the echo cancellation amount large in gamma _'1, gamma' when is ₂ small, characteristics are improved as shown by the curve (c), further E focusing on exponential decay of impulse response
Incorporation of the S algorithm gave curve (d). This is because the method of the present invention approximately utilizes the unknown input-output relationship expressed by the equations (55) and (57), and thus the approximation accuracy is improved by using the statistical property of the impulse response. Conceivable.

Further, FIG. 15 is an example in which the method (3-b) of the present invention is applied under the same conditions as in FIG. 14, and in the equation (84), the following coefficients are time-invariant and the echo path estimation is performed. Are the convergence characteristics of, and the curve (a) is v _{1
= V 2 = 0, w 1} = w 2 = 1, p 1 = p '1 = p 2 = p' 2 = 1, μ 1 = μ 2
= 0.5, the curve (b) has v ₁ = v ₂ = 1, w ₁ = w ₂ = 1 and p ₁ = p ' ₁ = p ₂
= P ' ₂ = 8, μ ₁ = μ ₂ = 0.05, the curve (c) is v ₁ = v ₂ = 0, w _{1
= W 2 = 1, p 1} = p '1 = p 2 = p' 2 = 8, μ 1 = μ 2 = 0.05, curve
(d) is v ₁ = v ₂ = 0, w ₁ = w ₂ = 0, p ₁ = p ' ₁ = p ₂ = p' ₂ = 8,
μ ₁ = μ ₂ = 0.05, curve (e) shows v ₁ = v ₂ = 1, w ₁ = w ₂ = 0, p ₁
= P ' ₁ = p ₂ = p' ₂ = 8, μ ₁ = μ ₂ = 0.05.
As can be seen from this result, in the method (3-b), the method (1-
Since the weighting method of b) is adopted, as shown by the curve (e) in FIG. 19, among the errors obtained for the approximate input / output relationship, the error obtained for the observable input / output relationship is redundant. It becomes possible to evaluate by removing the component, and even if the weighting coefficient is not changed with time as shown in FIGS.
Good results can be obtained. Application Example In a terminal having a stereo reproduction device, as described above, in addition to the method of receiving the stereo signals collected by the other party on two channels and reproducing them as they are, in the multipoint communication etc., the reception signal for each ground is transmitted. A method of making the receiving environment comfortable by arbitrarily performing sound image localization processing on the receiving side has been considered.
The method and apparatus of the present invention can also be applied to such a multipoint communication terminal. FIG. 16 shows the configuration of communication between four points. At each point, the sound collection is one channel (monaural). Now, the point D will be described as an example. At the point D, the received image signals from the points A, B, and C are subjected to sound image localization processing so as to be localized to the right, center, and left, respectively, to generate a new 2-channel stereo reproduction signal to generate a 2-channel stereo reproduction signal. Stereo playback shall be performed. In this case, the response y ₁ (k) to the two-channel received signals x ₁ (k) and x ₂ (k) at a certain point
Is

[0093]

[Equation 58] It is. Here, applying the echo path similarity error vector f ₁ (k) between h ₁₁ ^T (k) and h ₂₁ ^T (k),

[0094]

[Equation 59] Then, the response y ′ ₁ (k) obtained when the reproduction channels of the received signals x ₁ (k) and x ₂ (k) are exchanged is

[0095]

[Equation 60] Is obtained. From this, a correction equation of the pseudo echo path vector based on the method of the present invention can be obtained by establishing a simultaneous equation based on the equations (108) and (110). When a conventional echo canceling method is applied to a communication conferencing system that is generally composed of a multi-channel reproduction system and a sound collection system of one or more channels, and when the cross-correlation between the received signals of each channel is strong,
Since the echo path is not correctly estimated, there is a problem that the echo increases every time the cross-correlation between the received signals changes. In the method of the present invention, apart from the input / output relationship of the multi-channel reception signals having a constant cross-correlation, the input / output relationship of the multi-channel signals having a different cross-correlation is used approximately. In other words, when some or all of the multi-channel received signals are exchanged and reproduced as an acoustic signal from a channel different from the corresponding channel, the unknown output that should be obtained by the sound collector via the corresponding echo paths Is approximately calculated. From this, since the simultaneous equations are solved for these input / output relationships to estimate the echo path, the solution does not become indefinite, and there is an effect of improving the above problems.

[Brief description of drawings]

FIG. 1 is a block diagram showing a one-channel echo canceller.

FIG. 2 is a block diagram illustrating a multi-channel echo canceller system.

FIG. 3 is a block diagram showing a conventional multi-channel echo canceller.

FIG. 4 is a block diagram showing a stereo voice communication system.

FIG. 5 is a diagram showing an echo path estimation operation in the case where there is cross-correlation between received signals.

FIG. 6 is a diagram showing an example of symmetrical arrangement of microphones and speakers in a stereo communication system.

FIG. 7 is a block diagram showing an example of a multi-channel echo canceller system to which the present invention is applied.

FIG. 8 is a block diagram showing a configuration example of a multi-channel echo canceller to which the method of the present invention is applied.

9 is a functional block diagram showing a configuration of an echo path estimation unit in FIG.

10 is a functional block diagram showing another configuration of the echo estimation unit in FIG.

FIG. 11 is a functional block diagram showing the configuration of a correction vector calculation unit in FIGS. 9 and 10.

FIG. 12 is a functional block diagram showing another configuration of the correction vector calculation unit in FIGS. 9 and 10.

FIG. 13 is a diagram showing a convergence characteristic of an error between a true echo path characteristic and a pseudo echo path characteristic.

FIG. 14 is a diagram showing an influence of a weighting factor given to an approximate error signal on a convergence characteristic of an error signal.

[15] the weighting coefficients _{v 1, v 2, w 1} , w 2, order _{p 1, p '1, p} 2, p' 2,
The figure which shows the influence on the convergence characteristic of the echo path estimation of step size μ ₁ and μ ₂ .

FIG. 16 is a block diagram showing a voice communication conference system having a sound image localization function in communication between four points.

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] October 8, 1996

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claims

[Correction method] Added

[Correction contents]

[Claims]

Claims (23)

[Claims] 1. A echo canceling method for a multi-channel voice communication conference, comprising the steps of: (a) reproducing received signals of a plurality of channels as sound signals by a plurality of corresponding reproducers, and (b). (C) The received signals of the plurality of channels are accumulated for a predetermined period of time, respectively, from the plurality of regenerators through the respective echo paths to be added as observable echo signals in at least one sound collector, The received signal vector is generated from the received signal stored in, and (d) the above received signal vectors for all channels are combined to generate an observable received signal combination vector, and (e) each echo path By adding the received signal coupling vector to the pseudo echo path having a pseudo echo path coupling vector that simulates the coupling vector of the echo path vectors represented by the impulse response of A pseudo echo signal that simulates the observable echo signal is generated at its output, and the received signal combination vector and the corresponding echo signal correspond to the observable input / output relationship for the multiple echo paths. , (F) subtracting the pseudo echo signal from the echo signal to obtain a residual echo signal, and (g) replacing the exchanged reception signal combination vector when at least two reception signals of the plurality of channels are exchanged with the pseudo echo. Obtain an approximate pseudo echo signal when given to the path, (h) the echo signal to be obtained from the sound collector through the echo paths when the received signal is replaced in the same manner as in step (g) above. Is calculated as an approximate echo signal, and (i) the approximate pseudo echo signal is subtracted from the approximate echo signal to obtain the approximate residual echo signal, and (j) the observed received signal combination vector and it Corresponding to the observable input / output relationship representing the relationship of the observable residual echo signal, and the approximate input / output relationship representing the relationship between the exchanged received signal combination vector and the approximate residual echo signal are evaluated, and evaluated. Based on the correction vector, (k) the correction vector is used to correct the pseudo echo path vector. 2. The method of claim 1, wherein the steps are
In the evaluation of (j), there is a step of weighting at least one of the approximate input / output relationship and the observable input / output relationship when the received signal of each channel is reproduced as it is. 3. The method according to claim 2, wherein the weighting includes echo signals corresponding to outputs of the approximate input-output relationship and the observable input-output relationship, pseudo echo signals that simulate the echo signals, and difference between the echo signals. And the residual echo signal is multiplied by a time-invariant or time-variant coefficient. 4. The method according to claim 2, wherein the weighting separates the approximate residual reverberation signal into a component correlated with the observable residual reverberation signal and a component other than the observable residual reverberation signal, and time-invariant to each component. Alternatively, it includes the step of independently multiplying the time-varying coefficient. 5. The method according to claim 4, wherein the actually observed known input / output relationship, the observable residual echo signal used for evaluation of the input / output relationship, and the plurality of echo paths. Similarity of the echo paths between the sound collectors and the sound collectors is obtained by using the similarity and the approximate input / output relationship is obtained, and the physical arrangement relationships such as distance and direction between the sound collectors and sound collectors are similar. , And introduce the difference in similarity as an echopath similarity error vector. In addition to the approximate input-output relationship and the observable input-output relationship, the past input-output relationship is calculated using a projection algorithm. The step of obtaining a correction vector of the pseudo echo path is included. 6. The method of claim 5, including the step of introducing a weighting factor that allows the effect of the echopath-like error vector to be adjusted. 7. The method according to claim 2, 3 or 4, wherein the ratio of the echo signal to the residual echo signal, the average power ratio of the echo signal to the residual echo signal, and the average power ratio of the echo signal to the residual echo signal at each calculation time.
Determining a value of the weighting coefficient using at least one of an absolute value of the residual echo signal and an average power of the residual echo signal. 8. The method according to claim 1, wherein the known actual input / output relationship observed, the residual error signal used for evaluation of the input / output relationship, and the echo path similarity are utilized. The method includes a step of obtaining the approximate input / output relationship. 9. The method according to claim 2, wherein a known input / output relationship actually observed, a residual error signal used for evaluation of the input / output relationship, and echo path similarity are utilized. The method includes a step of obtaining the approximate input / output relationship. 10. The method of claim 4, wherein the ratio of the reverberation signal to the residual reverberation signal, the average power ratio of the reverberation signal to the residual reverberation signal, the absolute value of the residual reverberation signal,
Calculate at least one of the average power of the residual echo signal, and determine the convergence state of the residual echo signal based on the value, if the convergence state of the residual echo signal is poor, the observable The weight coefficient for the component having a correlation with the residual echo signal is set to a predetermined value other than 0, the weight coefficient for the remaining components is set to 0, and both components are set to 0 when the convergence state is good.
And a step of weighting each of the predetermined values including. 11. The method according to claim 8, wherein, in addition to the approximate input / output relationship and the observable input / output relationship, a past relationship between them is used to obtain the correction vector of the pseudo echo path. including. 12. The method according to claim 11, including the step of utilizing the past input / output relationship by using a projection algorithm. 13. The method according to claim 8, 9 or 11, wherein the similarity of the echo paths is between the regenerator and the sound collector having a similar physical arrangement relationship such as a distance between the regenerator and the sound collector and a direction. The echo paths in 1 are used assuming that they are equal to each other. 14. The method according to claim 4, 8, 9 or 11, wherein the regenerators and sound collectors having similar physical arrangements such as distance and direction between the sound collectors and the echo paths between the sound collectors. Assuming similarity, the step of introducing and utilizing the difference in similarity as an echo path similarity error vector is included. 15. The method of claim 14 including the step of introducing a weighting factor that allows the effect of the echopath-like error vector to be adjusted. 16. The method according to claim 8, 9 or 11, wherein when a correction vector of the pseudo echo path is obtained, a different correction weight coefficient is given to each tap based on exponential attenuation of impulse response of the echo path. Applying the ES algorithm. 17. An echo canceller for a multi-channel audio communication conference, comprising: a plurality of regenerators for reproducing and outputting received signals of a plurality of channels as acoustic signals, and at least one transmission channel. The connected sound collector and the acoustic path from the plurality of regenerators to the sound collector form a reverberant path, and the sound signals collected by the sound collector and synthesized are A reception signal vector generator, which is input to the transmission channel as an echo signal and temporarily holds the reception signals of the plurality of channels to generate each reception signal vector, and a reception signal vector of the plurality of channels are combined. And a vector combination unit for generating a reception signal combination vector, and the reception signal combination vector is given, and the impal of the echo path from the plurality of regenerators to the sound collector is given. It has a pseudo echo path coupling vector that simulates an echo path coupling vector that is a combination of echo path vectors that represent responses, and the received signal coupling vector is input to output a pseudo echo signal that simulates the observable echo signal. Pseudo echo signal generation unit, a subtracter that subtracts the pseudo echo signal from the echo signal from the sound collector, and outputs the difference as an observable residual echo signal, and at least two received signals of the plurality of channels A vector permutation combining unit that generates a permuted reception signal combination vector obtained by exchanging and combining, and at least the observed reception signal combination vector and the corresponding observable residual echo signal are given, and the arrangement When an approximate pseudo echo signal is obtained when a replacement reception signal combination vector is applied to the pseudo echo path, and when the reception signal is replaced, the above The echo signal that should be obtained from the sounding device is obtained as an approximate echo signal by approximation, the approximate pseudo echo signal is subtracted from the approximate echo signal to obtain the approximate residual echo signal, and the received signal combination vector and the corresponding observable signal The observable input / output relationship that represents the relationship with the residual echo signal and the approximate input / output relationship that represents the relationship between the rearranged received signal combination vector and the corresponding approximate residual echo signal are evaluated, and based on the evaluation An echo path estimating unit that obtains the correction vector by using the correction vector, generates the pseudo echo path vector by correcting the pseudo echo path vector by the correction vector, and gives the pseudo echo signal vector to the pseudo echo signal generating unit. 18. The apparatus according to claim 17, wherein the echo path estimation unit gives residual echo signal weight adding means for giving relative weight to the observable residual echo signal and the approximate residual echo signal. A correction vector calculation unit that obtains the correction vector from the relationship between the received signal combination vector and the rearranged reception signal vector and the observable residual echo signal and the approximate residual echo signal, and the pseudo echo path using the corrected vector. And an adaptive filter updating unit which corrects the vector and gives the pseudo echo signal generating unit. 19. The apparatus according to claim 17, wherein the echo path estimation unit separates the approximate residual echo signal into a correlation component correlated with the observable residual echo signal and a non-correlation component other than the correlation component. It includes a component separating section and a component weighting coefficient adding means for giving a relative weight to the correlated component and the non-correlated component. 20. The apparatus according to claim 19, wherein the echo path estimation unit includes a cross-correlation calculation unit that obtains a cross-correlation of the rearranged reception signal vector with respect to the reception signal vector, and the component separation unit includes the approximate residual echo. It includes means for calculating a product of the cross-correlation obtained by the cross-correlation calculation unit and the observable residual echo signal as a component of the signal having a correlation with the observable residual echo signal. 21. The apparatus of claim 18 or 19,
The echo path estimation unit, for each calculation time, the ratio of the echo signal and the residual echo signal, the average power ratio of the echo signal and the residual echo signal, the absolute value of the residual echo signal, the residual echo signal At least one of the average powers of the above is calculated as an index representing the convergence state of the observable residual echo signal, and a weighting factor control unit that controls the weighting factor based on the value is included. 22. The device of claim 18 or 19,
The reverberant path estimation unit determines the influence on the residual reverberant signal caused by the similar error difference vector that is the difference between the similar error vectors between mutually similar reverberant paths and the corresponding similar error vectors between the pseudo reverberant paths. A similar error weighting coefficient adding unit that gives a weighting coefficient for adjustment to the received signal vector is included. 23. The apparatus of claim 18 or 19, wherein
The correction vector calculation unit, the received signal vector,
The rearranged reception signal vector, a signal storage unit for holding the residual echo signal retroactively a predetermined number of times in the past from the present time, and the respective reception signal vector and rearranged reception signal vector from the present time to the past time , A simultaneous equation calculation unit that obtains the correction vector by solving a simultaneous equation using the relationship between the residual echo signal and the correction vector. 24. In the apparatus of claim 22, the correction vector calculation unit includes the reception signal vector, the rearranged reception signal vector, a signal storage unit that holds the residual echo signal backward a predetermined number of points in the past from the present time, and these current times. From the past to each of the past reception signal vector and rearranged reception signal vector, the residual echo signal, the correction vector and the similar error difference vector by using the relationship between the correction vector by solving the simultaneous equations And a simultaneous equation calculation unit for obtaining the simultaneous equations. 25. In the device of claim 23 or 24, the correction vector calculation unit weights each of the reception signal vector and the rearranged reception signal vector from the present time point to the past a predetermined number of time points by an exponential attenuation coefficient vector having the same number of elements. The simultaneous equation calculating unit includes the exponential weighting coefficient adding means for calculating the correction vector by using the ES algorithm.