JPH11317695

Patent Translate
Powered by EPO and Google
Notice
This translation is machine-generated. It cannot be guaranteed that it is intelligible, accurate,
complete, reliable or fit for specific purposes. Critical decisions, such as commercially relevant or
financial decisions, should not be based on machine-translation output.
DESCRIPTION JPH11317695
[0001]
BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a
multi-channel acoustic echo cancellation method for canceling acoustic echo which is the cause
of howling and hearing impairment in a communication conference system having a multichannel reproduction system, its apparatus and its program recording It relates to the medium.
[0002]
2. Description of the Related Art In recent years, with the spread of digital networks such as
ISDN, LAN, and the Internet, and the development of high-efficiency coding technology for voice
and image, various types of telephone conversations have appeared in addition to conventional
telephones. With a TV conference system that allows you to make a call while looking at the
other party's face using a large screen TV or a personal computer or work station, a desktop
meeting system, etc., it is easy for more than one person to participate in a call and it is more
natural In many cases, a speech communication system capable of providing a speech
environment is employed. However, this system using a speaker and a microphone is
accompanied by the problem of the occurrence of echo and howling, and to avoid this, acoustic
echo canceller technology is indispensable.
[0003]
09-05-2019
1
In fact, in these situations, acoustic echo canceller devices are widely used, but most of them are
for single channel audio, and one system of loudspeakers from one channel to microphones of
one channel (channel). Only acoustic wraparounds can be eliminated. On the other hand, in many
TV broadcasts, music media, etc., stereo sound is common, and the implementation of such multichannel sound conversion is strongly demanded also for loud speech. Therefore, it is necessary
to realize an acoustic echo canceller device for multi-channel audio that enables cancellation of
acoustic interference from a plurality of (channel) speakers to a microphone, and in recent years,
technical problems and solutions therefor have been realized. The examination of has become
active.
[0004]
Acoustic echo cancellation in a teleconferencing system generally consisting of a receiver system
for N (N ≧ 2) channels and a transmitter system for M (M 送 1) channels is conventionally
performed according to the configuration shown in FIG. . That is, the N channel echo cancelers
221, 222,..., 22M constituting the echo cancellation unit 22 respectively between the receiving
terminals 111, 112,..., 11N of all N channels on the receiving side and the transmitting system of
the M channel .., 11 N are connected as acoustic signals by the respective speakers 121, 122,...,
12 N, and echoes (represented by impulse responses h mn) Reverberation) Acoustic echoes that
go around each of the microphones 161, 162,..., 16M via the path 15 nm (1 ≦ n ≦ N, 1 ≦ m ≦
M) are eliminated.
[0005]
The N-channel echo cancelers 221, 222,..., 22M described above have the same configuration for
each transmission channel, and for example, the configuration shown in FIG. This is described in
the document B. Widrow and S. D. Stearns, "Adaptive signal processing," Prince-Hall, Inc. pp. 198200, (1985), taking the case of two channels as an example. In the configuration of FIG. 2, the
reception signals x1 (k), x2 (k),..., XN (k) are respectively input to adaptive filters 2211, 2212,
and 221N forming N pseudo echo paths, and the adaptive filter 2211 is , 2212,..., 221N are
added by the adder 222 to generate a pseudo echo signal y'm (k), and the pseudo echo signal and
the collected signal (echo signal) ym (k) from the microphone 16m are generated. And the error
signal (residual echo signal) em (k), which is the output thereof, is fed back to the adaptive filters
2211 to 221N, and received signals x1 (k) to xN (). k), the filter coefficient vector is determined
by, for example, the NLMS algorithm so that the error signal em (k) becomes smaller, and the
adaptive filters 2211 to 221N are adaptively controlled.
09-05-2019
2
[0006]
Although not shown in FIG. 1, in the echo cancellers 221 to 22M, an acoustic signal zm (k) to be
originally transmitted is input to the microphones 161 to 16M, and is transmitted through the
echo cancellers 221 to 22M. , And to prevent the reproduced sound from the speaker picked up
by the microphone from being transmitted together with the signal zm (k) as an echo ym (k).
That is, the error signal em (k) output as a result of echo cancellation in FIG. 2 includes the signal
zm (k) to be originally transmitted. However, the present invention focuses only on the echo
signal ym (k) which is a pickup signal by the microphone of the reproduction sound from the
speaker, and does not particularly mention the signal zm (k) to be transmitted. The same applies
to the detailed description of the present invention.
[0007]
When the correlation between the respective reception signals x1 (k) to xN (k) is low, the
adaptive filters 2211, 2212,..., 221N can estimate the corresponding echo path with relatively
high accuracy, and therefore become the cancellation target It is possible to generate a pseudo
echo that accurately simulates an acoustic echo. However, in an actual communication
conference, in many cases, when the voice of one speaker is transmitted from the other party on
multiple channels and these become receiving signals, there is a very high correlation between
the receiving signals, and an adaptive filter Both the convergence speed and the convergence
accuracy may deteriorate, and the desired echo cancellation performance may not be obtained. In
order to solve this problem, it is possible to perform preprocessing to reduce or change the
correlation between the received signals before each received signal is input to the N channel
echo canceller 221, 222, ..., 22M. No. 5, 661, 813.
[0008]
In the configuration shown in the above-mentioned US Patent, as shown in FIG. 3, in the
configuration of FIG. 1, the pre-processing unit 30 having the above-described function receives
the receiving terminals 111 to 11N, the speakers 121 to 12N and the N channel echo canceller
221 It is added between ˜ 22M. A configuration example of this pre-processing unit 30 is shown
in FIG. Here, the received signals from the receiving terminals 111 to 11N and the additional
signals generated in the additional signal generation units 3011, 3012, ..., 301N are added by the
adders 3021, 3022, ..., 302N to process the processed signals. Output x1 '(k), x2' (k), ..., xN '(k).
09-05-2019
3
When generating the additional signal, there are cases where the information of the reception
signals x1 (k),..., XN (k) is used and cases where it is not used. By increasing the size of the
additional signal, it is possible to improve the convergence characteristics of the adaptive filters
2211, ..., 221N. A similar approach is also shown in U.S. Patent No. 5,828,756. U.S. Patent Nos.
5,661,813; J. Benesty, DR R. Morgan, and M. M. Sondhi, "Abetter Understanding and an Improved
Solution to the Problems of Stereophonic Acoustic Echo Cancellation", Proc. The preprocessing
method proposed in ICASSP 97, vol. 1, pp. 303-306 (1997), etc. can be realized by the
configuration shown in FIG. For example, for N channel signals xi (k) (i = 1, 2,..., N) at discrete
time k, processing signals xi ′ (k) are processed by processing functions fi (i = 1, 2,. Xi '(k) = xi
(k) even if it is a preprocessing unit that outputs (i = 1, 2, ..., N) as xi' (k) = fi [xi (k)] (1) Since it
can be transformed into +) (fi [xi (k)]-xi (k)) (2), additional signal fi [xi (k)]-xi (k) is added to the
original signal xi (k) It can be seen.
[0009]
SUMMARY OF THE INVENTION In order to improve the convergence characteristic of the
adaptive filter in the above N channel echo canceller 221, 222,..., 22M, as shown in FIG.
However, since the preprocessed signal is actually output from the speakers 121, 122, ..., 12N, if
the preprocessing shown in FIG. 4 is performed, the magnitude of the additional signal is
processed As compared with the previous reproduced sound, it is necessary to be limited within a
range that does not cause auditory discomfort. For this reason, the improvement amount of the
convergence characteristic of the adaptive filters 221 to 22N is also limited, and the
improvement of the echo cancellation performance is also limited.
[0010]
Although the above example has been described with reference to echo cancellation in a multichannel call conference system, the principle of echo cancellation originally simulates the echo
path from the speaker to the microphone in FIG. 1 in the echo canceller (ie impulse response of
echo path) The estimation is intended to cancel the actual echo signal ym (k), and the reception
signal does not necessarily have to be the reception signal from a remote place in the call
conference system. For example, in a hall, a theater, a dome or the like provided with a loud
sound system, this echo cancellation technology is used even in the case where an acoustic signal
from a target sound source is picked up by a microphone and background sound emitted from a
speaker is removed. Can be applied. Therefore, including the present invention, the reception
signal in the following description may be a signal from any signal source as long as it is a
reproduction electric signal to be supplied to the reproduction channel to be reproduced from
09-05-2019
4
the speaker.
[0011]
SUMMARY OF THE INVENTION An object of the present invention is to provide a new multichannel acoustic echo cancellation method for increasing the amount of improvement in echo
cancellation performance even when the magnitude of a signal added by preprocessing is small.
It is providing the device and the program recording medium. According to the present invention,
it has N reproduction channels each including a speaker for outputting reproduction sound, and
at least one pickup channel including N being an integer of 2 or more and including a
microphone for collecting an acoustic signal. The multi-channel acoustic echo cancellation
method in an acoustic system in which the N speakers and the microphone are disposed in a
common sound field includes the following steps: (a) for reproduction respectively input to the N
reproduction channels (B) adding the reproduction signals of the N reproduction channels and
the addition signals for each reproduction channel to generate a processing signal of each
reproduction channel; (c) reproducing the processed signals of the N reproduction channels from
the speakers of the corresponding reproduction channel, and (d) selecting one of the N
reproduction channels From each of the speakers, an acoustic echo that is looped and
synthesized to the microphone of the pickup channel is collected and input as an acoustic echo
signal to the pickup channel, (e) the N reproduction signals and the N Individual processing is
performed on each additional signal to generate a pseudo echo that simulates the acoustic echo
signal in the pickup channel, and acoustic echo cancellation is performed by subtracting the
pseudo echo from the acoustic echo.
[0012]
According to the present invention, it has N reproduction channels each including a speaker for
outputting reproduction sound, and at least one pickup channel including N being an integer of 2
or more and including a microphone for collecting an acoustic signal. In a multi-channel acoustic
echo canceller in an acoustic system in which N speakers and a microphone are arranged in a
common sound field, N generates an additional signal for reproduction signals respectively input
to the N reproduction channels. N additional signal generation means, N processing signal
generation means for adding the reproduction signals of the N reproduction channels and the
addition signals of the N reproduction channels to each corresponding reproduction channel to
generate processing signals, and N N above-mentioned speakers, each of which is provided in
each of the reproduction channels, reproduces the processing signal of each reproduction
channel, and outputs reproduction sound The microphones that collect acoustic echoes that are
looped and synthesized from the N speakers and input them as the acoustic echo signals to the
pickup channel, and the N reproduction signals and the N additional signals individually And
09-05-2019
5
processing means for generating a pseudo echo simulating the acoustic echo signal in the pickup
channel, and canceling the acoustic echo signal by subtracting the pseudo echo from the acoustic
echo signal.
[0013]
In the recording medium according to the present invention, the processing steps for carrying
out the multi-channel acoustic echo cancellation method are recorded as a computer program.
[0014]
BEST MODE FOR CARRYING OUT THE INVENTION In the multi-channel acoustic echo
cancellation method, the echo cancellation performance can be improved by a method in which
an additional signal is added to the received signal, but the amount of improvement improves the
additional signal due to auditory limitations. It is limited because it can not be enlarged.
From another point of view of this problem, it can be said that there is a lack of sufficient
information exchange between the pre-processing unit 30 and the N-channel echo cancelers
221,..., 22N in FIG.
In the processing signal which is the sum of the reception signal and the small additional signal,
the information of the additional signal is buried in the information of the reception signal.
Therefore, in the conventional N-channel echo canceller, it is difficult to effectively extract and
use the information of the additional signal from the input processed signal.
[0015]
Therefore, in the present invention, the information on the additional signal can be directly given
to the echo canceller. FIG. 5 is a block diagram showing the basic configuration of the multichannel acoustic echo cancellation apparatus according to the present invention. Also in the
device according to the present invention, additional signal generation units 3011 to 301N and
adders 5011 to 501N are provided corresponding to the reception channels, respectively, and
microphones 161 to 16M and an echo canceller 401 are provided corresponding to the
respective transmission channels. To 40M, and the received signals x1 (k) to xN (k) and the
additional signals a1 (k) to aN (k) are added by the adders 5011 to 501N, and the addition result
09-05-2019
6
(processed signal) x1 ' (k) to xN '(k) are applied to the speakers 121 to 12N in the same manner
as the conventional configuration of FIG.
[0016]
A different point is the individually used echo canceller in which the reception signals x1 (k) to
xN (k) before addition and the additional signals a1 (k) to aN (k) before addition are respectively
provided corresponding to the transmission channels. 401 to 40M, and based on them, the echo
signals y1 (k) to yM (k) of each transmission channel are estimated. That is, in the conventional
FIG. 4, the reception signals x1 (k), x2 (k),..., XN (k) are added to the additional signals a1 (k), a2
(k),. After the addition, the addition result (ie, the preprocessed signal) of those additions is given
to echo cancelers 221 to 22M of the transmission channel as shown in FIG. 3 to generate pseudo
echo signals. As shown in FIG. 5, before the addition of the additional signal, the received signals
x1 (k), x2 (k),..., XN (k) and the additional signal are sent to the individually used echo cancelers
401, 402,. a1 (k), a2 (k),..., aN (k) are separately input, and separately from them, received signals
xi (k) (i = 1, 2,..., N) and additional signals ai (k) ) Is added by the adder 501i and reproduced from
the speaker 12i. Therefore, even if the magnitude of the additional signal is limited due to a sense
of hearing, the information on the additional signal embedded in the reception signal in the
conventional configuration is used in the individually used echo canceller 401, 402, ..., 40M. Can
be used directly to improve echo cancellation performance.
[0017]
As the nature of the additional signal suitable for the configuration shown in FIG. 5, it is required
to take advantage of the ability to handle the reception signal xi (k) and the additional signal ai
(k) individually. That is, if the additional signal ai (k) contains many components correlated with
the incoming signal xi (k), even if the additional signal is processed individually, the influence of
the incoming signal xi (k) on the processing result Is included. Therefore, the set of reception
signals xi (k),..., XN (k) of all the reception channels and the set of additional signals a1 (k),. Is
desirable. Therefore, with regard to the method of generating the additional signal, the number of
received channels is generated as the additional signal since the number of low cross-correlation
signals having low cross correlation with the received signals of all channels, that is, close to
zero.
[0018]
09-05-2019
7
Alternatively, when the individual additional signals have high correlation among the additional
signal sets of all channels, in the conventional method, for the same reason as in the case where
the correlation between the reception signals is high, Improvement is considered to be limited.
Therefore, as a method of generating an additional signal, instead of the low cross correlation
signal, signals having low cross correlation (near zero) with received signals of all channels and
low cross correlation (near zero) are mutually low. The number of receiving channels is
generated as a correlation signal, and these are allocated as additional signals. Next, a method of
generating a low cross correlation additional signal will be described. Let the normal speech
signal be x (k) and the expected value be
[0020]
It will be expressed as: The following equation E [x2 (k)] ¦ x (k)> 0 = E [x2 (k)] ¦ x (k) <0 (4)
approximately holds for this expected value, so x (k) and The absolute value ¦ x (k) ¦ is
approximately uncorrelated, that is, for any time difference n, approximately E [x (k) ¦ x (k−n) ¦]
= 0 (5) It holds. Therefore, assuming that i = 1, 2,..., N are numbers assigned to the reception
channels, the reception signal xi (k) of each reception channel and its absolute value signal ¦ xi
(k) ¦ Also, if the received signals xi (k) and xj (k) of any two received channels are highly
correlated, the absolute value signals ¦ xi (k) ¦ and ¦ xj (k) ¦ are also highly correlated with each
other, It is considered that the reception signal xi (k) of any reception channel has low correlation
with the absolute value signal ¦ xj (k) ¦ of any other reception channel. Therefore, αi ¦ xi (k) ¦ is
generated with the adjustment coefficient for each receiving channel as αi, and this is used as a
low cross correlation signal, that is, an additional signal. It is preferable to select the adjustment
coefficient αi as large as possible within the range that the additional signal component in the
reproduced sound from the speaker is not offensive in order to converge the adaptive filter
coefficient in a short time. Additional Signal Generation Method (1) When the received signals x1
(k) to xN (k) have high correlation with each other, the low cross-correlation signals αi ¦ xi (k) ¦
Improvement is considered to be limited. Therefore, in each reception channel, the absolute value
of the reception signal is taken, and the zero crossing at which the sign of the original reception
signal changes is detected, and the above absolute value is synchronized with the detection point
according to different rules for each reception channel. After giving the signal a positive or
negative sign, the low cross-correlation signal is obtained by multiplying the adjustment factor.
[0021]
Here, as a method of giving a positive or negative sign to the absolute value signal, there are the
09-05-2019
8
following methods. In one receiving channel, the zero cross detection point is counted, and the
sign given to the absolute value signal is exchanged at the zero cross point every predetermined
count number C. In accordance with this rule, the sign given to the absolute value signal is
inverted at the zero crossing point for each of the other different reception channels in the same
way in the same way in different count values C. However, the count value C is an integer of 0 or
2 or more, and C = 0 means fixing the code. In this way, even when the received signals x1 (k) to
xN (k) are highly correlated with each other, the obtained low cross-correlation signals are less
correlated with each other, unlike the above case.
[0022]
6A to 6D are examples in which the absolute value signal is given a sign according to the above
rule. FIG. 6A shows an original signal, and the sign changes every zero crossing, so C = 1, and as
shown in FIG. 6B, the sign of the absolute value signal does not change, and C = 0. Assuming that
C = 2, as shown in FIG. 6C, the code changes every time zero crossing is performed twice, and
when C = 3, the code changes every time zero crossing is performed three times as shown in FIG.
6D. It can be seen that the signals shown in FIGS. 6A to 6D are all uncorrelated with one another
including the original signal. Hereinafter, the value of C will be referred to as a code switching
cycle. Additional Signal Generation Method (2) A method of generating yet another low cross
correlation signal will be described. That is, in each receiving channel, the absolute value signal
of the receiving signal is taken, and the above-mentioned absolute value signal is multiplied by
the adjustment coefficient which becomes nonzero for one receiving channel and zero for the
other receiving channel, thereby adding low cross correlation. Get a signal. As a method of
selecting a receiving channel giving a non-zero adjustment coefficient, there is a method of
switching periodically or randomly with time, or a method of giving priority to a receiving
channel having a large signal level. In this method, a valid additional signal is added to only one
receiving channel at any time, so the time taken to generate an accurate pseudo echo is
considered to be longer than in the above method (1). There is an advantage that it is possible to
reduce the operation taking into account that the additional signal of the reception channel of the
above is zero.
[0023]
Thus, according to the present invention, an additional signal to be added to each received signal
in order to fluctuate or reduce the cross correlation between the received signals of multiple
channels can be used separately for estimating the pseudo echo without adding it to the received
signal. As a result, it is possible to generate a pseudo echo with high estimation accuracy by
09-05-2019
9
effectively using the uncorrelated component contained in the additional signal. Individually
Used Echo Canceller FIG. 7 is an embodiment of the individually used echo canceller 40m (m = 1,
2,..., M) in FIG. In this configuration, received signals x1 (k) to xN (k) from the receiving terminals
111, 112,..., 11N are input to the received signal adaptive filters 4011, 4012,. Ru. In the present
invention, the additional signal adaptive filters 4021 to 402N are provided corresponding to the
respective reception channels, and the additional signals a1 (k) to aN (output terminals 181, 182,
..., 18N of the additional signal generation units 3011 to 301N are provided. k) are input to the
additional signal adaptive filters 4021, 4022, ..., 402N, respectively. The sum of the outputs of
the respective adaptive filters 4011, ..., 401N and 4021, ..., 402N is taken by the adder 403 to be
a pseudo echo ym '(k), and an acoustic echo ym (k) collected by the microphone 16m. Then, this
pseudo echo ym '(k) is subtracted by the subtraction circuit 404 to realize echo cancellation.
Also, in this embodiment and the following embodiments, two different appropriate weights are
applied to the error signal em (k) obtained by subtracting the pseudo echo ym '(k) from the
acoustic echo ym (k). The coefficients w1 and w2 are given to generate two weighted error
signals w1em (k) and w2em (k), the former is given to the adaptive filters 4011,..., 401N, and the
latter is given to the adaptive filters 4021,. Adaptive filters 4011 to 401N and 4021 to 402N are
based on weighted error signals w1em (k) and w2em (k) and input signals x1 (k) to xN (k) and a1
(k) to aN (k). For the currently set filter coefficients, calculate the filter coefficient vector that
minimizes the error signal em (k) by the well-known NLMS algorithm etc., thereby updating each
filter coefficient of the adaptive filter, and the next input Prepare for The specific calculation
method of the filter coefficient update will be described later.
[0024]
The weighting for the error signal fed back to each of the reception signal adaptive filters 4011,
..., 401N and the additional signal adaptive filters 4021, ..., 402N is, for example, small for each of
the reception signals with large power. A weighting factor w1 is given, and a large weighting
factor w2 is given to each of the above-mentioned small-power additional signals. As described
above, in the pseudo echo generation of the individually used echo canceller shown in FIG. 7,
there is an advantage that the precision of the pseudo echo can be controlled by processing the
information on the reception signal and the information on the additional signal with different
weightings. . In practice, when reproducing the processing signal in which the additional signal is
added to the receiving signal in FIG. 5 from the speaker, it is necessary to adjust the range in
which the additional signal has an effect on the audibility. Often small. In such a case,
information on a small additional signal can be effectively used by the above weighting process.
However, as a basic configuration of the present invention, w1 = w2 in FIG. That is, the weighting
unit 411 may be omitted, and the same em (k) may be given to the adaptive filters 4011 to 401N
and 4021 to 402N. The same can be said for all the other modified embodiments below.
Modification 1 of Echo Canceller FIG. 8 shows an alternative to the echo canceller of FIG. 7
09-05-2019
10
applied to the individually used echo canceller 40m (m = 1,..., M) in FIG. The embodiment of FIG.
7 shows the case where the adaptive filters 4011 to 401N for the received signal and the
adaptive filters 4021 to 402N for the additional signal are separately provided and processed
separately in each N channel echo canceller 40m. As shown in FIG. 8, N additional signals a1 (k)
to aN (k) and received signals x1 (k) to xN (k) are subjected to another process and then
combined to obtain N combined signals x. "1 (k) to x" N (k) may be processed by N adaptive
filters 4011 to 401N.
[0025]
That is, as shown in FIG. 8, the received signals x1 (k) to xN (k) are multiplied by gain coefficients
g11 to g1N by multipliers 4121 to 412N, respectively, to be g11x1 (k) to g1NxN (k). The
additional signals a1 (k) to aN (k) are multiplied by gain coefficients g21 to g2N by multipliers
4101 to 410N, respectively, to be g21a1 (k) to g2NaN (k). The gain coefficients g21 to g2N for
the additional signal are larger than the gain coefficients g11 to g1N for the reception signal of
the corresponding reproduction channel. These multiplication results are added by the adders
4121 to 412N for each corresponding reproduction channel, and N combined processed signals
{g11x1 (k) + g21a1 (k)}, {g12x2 (k) + g22a2 (k)}, , And {G1NxN (k) + g2NxN (k)} are respectively
input to N adaptive filters 4011 to 401N.
[0026]
The sum of the outputs of the adaptive filters 4011 to 401N is taken by the adder 403, and the
sum is given to the subtractor 404 as a pseudo echo y'm (k), and is subtracted from the acoustic
echo signal ym (k). k) generate. The error em (k) is fed back to the adaptive filters 4011 to 401N,
and the combined processed signal {g11x1 (k) + g21a1 (k)}, {g12x2 (k) + g22a2 from the error
em (k) and the adders 4121 to 412N. The filter coefficients of the adaptive filters 4011 to 401N
are updated based on (k)},..., {g1NxN (k) + g2NxN (k)}.
[0027]
What is important in the embodiment of FIG. 8 is that the additional signal an (k) and the
reception signal xn (n) as components of the processing signal x'n (k) reproduced by each of the
speakers 12n (n = 1,..., N). The level ratio of the additional signal component g2nan (k) and the
reception signal component g1nxn (k) in each combined processed signal is made larger than the
09-05-2019
11
level ratio k). That is, the additional signal component an (k) in the reproduced sound from the
speaker 12n is reduced to an extent that it does not become offensive, while the additional signal
component uncorrelated in the synthetic processing signal {g1nxn (k) + g2nan (k)} The filter
coefficient of the adaptive filter 401 n can be converged in a short time by sufficiently increasing
the level of the signal component g 2 nan (k). Second Embodiment of Echo Canceller FIG. 9
shows another embodiment of each individually used echo canceller 40m shown in FIG. 5 in
place of that of FIG. In the same way as shown in FIG. 7, for each transmission channel, reception
signal adaptive filters 4011, 4012,..., 401N corresponding to all reception channels and
additional signal adaptive filters 4021, 4022,. On the other hand, respective reception signals
and respective additional signals respectively corresponding thereto are input, and a total sum
y'm (k) of the outputs of all the adaptive filters 4011 to 401N and 4021 to 402N is obtained by
the adder 403. The output y'm (k) of the adder 403 is subtracted from the acoustic echo ym (k)
collected by the microphone to obtain an error signal em (k). Given w 2 to generate weighted
error signals w 1 em (k) and w 2 em (k), give the former to the adaptive filters 4011,..., 401 N,
and the latter to the adaptive filters 402 1,. Adaptive filters 4011 to 401N and 4021 to 402N
update the filter coefficients by the NLMS algorithm or the like based on the given signals w1em
(k) and w2em (k), and prepare for the next input. In this embodiment, the error signal em (k) is
only used to determine the filter coefficients of the adaptive filters 4011 to 401N and 4021 to
402N and is not output as an echo cancellation result.
[0028]
In this embodiment, adders 4061 to 406N that add received signals x1 (k) to xN (k) and
additional signals a1 (k) to aN (k) for each corresponding received channel, and those adders
4061 to 406N. An echo canceling filter 4051 to 405 N which respectively gives the addition
result of 406 N, and an adder 407 which adds the outputs of the echo canceling filters 4051 to
405 N to obtain a pseudo echo signal y "m (k), an echo signal ym ( There is further provided a
subtractor 408 which cancels the echo by subtracting the pseudo echo signal y "m (k) from k)
and uses the residual as the output of the transmission channel. The adders 4061 to 406N add
the reception signals x1 (k) to xN (k) of the corresponding reception channels and the additional
signals a1 (k) to aN (k), respectively, and the adders 5011 to 501N in FIG. Are the same as the
processing signals x'1 (k) to x'N (k) which are the addition results of Therefore, the adders 4061
to 406N are not provided in the individual use type echo cancellers 401 to 40M, and the
processing signals x'1 (k) to x'N (k) which are the outputs of the adders 5011 to 501N in FIG. It
may be provided to the echo canceling filters 4051 to 405N of the echo canceller 40m (m = 1,...,
M). The same can be said for the examples described below.
[0029]
09-05-2019
12
Usually, the correlation between the additional signals a1 (k) to aN (k) is smaller than the
correlation between the reception signals x1 (k) to xN (k). It is expected to converge in a short
time with higher accuracy than the -401N filter coefficient. Therefore, in this embodiment, for
example, when the power of the error signal em (k) becomes sufficiently smaller than the power
of the echo signal ym (k), it is determined that the filter coefficients of the adaptive filters 4021
to 402N have converged to some extent, The filter coefficients of the adaptive filters 4021 to
402N are transferred to the echo cancellation filters 4051 to 405N and used.
[0030]
For that purpose, a short time average power calculation unit 413 and a transfer control unit
414 are provided. The short-time average power calculation unit 413 receives the error signal
em (k) and the echo signal ym (k), and calculates, for example, average values Pem (k) and Pym
(k) of power over a predetermined length within one frame period. . These values are given to the
transfer control unit 414, and the transfer control unit 414 obtains the ratio Pem (k) / Pym (k) of
these values, and the ratio of the filter coefficient is determined in advance as a transfer
condition (for example, If it is satisfied, the transfer control signal TC is applied to the adaptive
filters 4021 to 402N to transfer the filter coefficients of the adaptive filters 4021 to 402N to the
echo cancellation filters 4051 to 405N. Transfer coefficients. From then on, the ratio Pem (k) /
Pym (k) is always calculated, and the filter coefficients of the adaptive filters 4021 to 402N are
transferred to the echo cancellation filters 4051 to 405N only when the aforementioned transfer
conditions are satisfied. If necessary, it may be added as a necessary condition that the reception
signal level as the transfer condition, for example, the sum of short-time average power of all the
reception signals is a predetermined value or more.
[0031]
The outputs of these echo cancellation filters 4051,..., 405N are added by an adder 407 to obtain
the sum of them as a pseudo echo y "m (k). By subtracting the pseudo echo y "m (k) from the
acoustic echo ym (k) by a subtractor 408, an echo canceled signal is obtained. Variation 3 of
Echo Canceller FIG. 10 shows another variation of the individually used echo canceller 40m in
FIG. As described above, in the embodiment of FIG. 9, the pseudo echo signal generated by the
echo cancellation filter by transferring the coefficient of the additional signal adaptive filter
which is expected to be rapidly converged to the echo cancellation filter. As compared with the
embodiment of FIG. 7, y ′ ′ m (k) can be brought closer to the echo signal ym (k) in a shorter
09-05-2019
13
time, so that the effect of echo cancellation can be expressed in a shorter time. The filter
coefficients of the additional signal adaptive filters 4021 to 402N are more accurate than the
filter coefficients of the reception signal adaptive filters 4011 to 401N at the same elapsed time
point k, and the true echo impulse response (vector) h1 (k), h2 (k), ..., hN (k) (approximate the
impulse response vector, that is, the coefficient vector in the following description, not in bold
but in bold in the formula). It is estimated to be. In other words, the filter coefficients of the
adaptive filters 4011 to 401N and 4021 to 402N are obtained by the adaptive algorithm based
on the sum em '(k) of the sum of these outputs y'm (k) and the echo signal ym (k) Therefore, it
can be considered that the filter coefficients of the reception signal adaptive filters 4011 to 401N
contribute to the error of the filter coefficients of the additional signal adaptive filters 4021 to
402N currently obtained. Therefore, in the modified embodiment of FIG. 10, the additional filter
adaptive filter coefficients with high accuracy are also transferred to the reception signal
adaptive filter for use.
[0032]
The echo canceller 40m of FIG. 11 is provided corresponding to each transmission channel as
shown in FIG. 5, and as in the case of FIG. 9, an adaptive filter 4011 individually provided for
each reception signal and each additional signal. , 4012,..., 401N and 4021, 4022,..., 402N, the
corresponding reception signals x1 (k) to xN (k) and the additional signals a1 (k) to aN (k) are
input, respectively. The sum y 'm (k) of the filter outputs is obtained. From the acoustic echo ym
(k) collected by the microphone, the sum y'm (k) of the output of this adaptive filter is subtracted
to obtain an error signal em (k). This error signal em (k) is weighted by weighting factors w1 and
w2 by weighting section 411 to generate error signals w1em (k) and w2em (k), which are fed
back to adaptive filters 4011 to 401N and 4021 to 402N, respectively. Update the filter and
prepare for the next input.
[0033]
In the actual echo cancellation, the coefficients of the adaptive filter for the additional signal are
transferred to the corresponding echo cancellation filters 4031, 4032,..., 403N, and these echo
cancellation filters receive the reception signal and the additional signal for each reception
channel. The sum is input, the sum of the echo cancellation filters is obtained as a pseudo echo y
"m (k), and the echo cancellation is realized by subtracting this pseudo echo y" m (k) from the
acoustic echo ym (k) . The coefficients of the adaptive filter for the additional signal are
transferred to the corresponding echo cancellation filters 4051, 4052, ..., 405N, and
simultaneously with the adaptive filters 4011, 4012, ..., 401N provided for the respective
09-05-2019
14
reception signals. Can be used to improve the overall accuracy of the adaptive filter.
[0034]
The conditions for coefficient transfer may be the same as in the embodiment of FIG. As in the
case of FIG. 9, based on the power Pem (k) of the error signal calculated by the short-time
average power calculator 413 and the power Pym (k) of the acoustic echo, the transfer controller
414 determines whether the transfer condition is satisfied. If satisfied, the transfer control signal
TC is applied to the adaptive filters 4021 to 402N, and the filter coefficients are transferred to
both of the adaptive filters 4011 to 401N and 4051 to 405N. In this modified embodiment, when
receiving the transfer of the filter coefficients, the reception signal adaptive filters 4011 to 401N
are newly updated from the filter coefficients. Echo Canceller Modification 4 FIG. 11 shows still
another modification of the individually used echo canceller shown in FIG. In FIG. 9 described
above, in general, the correlation between the reception signals x1 (k) to xN (k) is large, so the
coefficients of the adaptive filters 4021 to 402N to which the low correlation additional signal is
given are the reception signal adaptive filters 4011 to 401N. The coefficients of the additional
signal adaptive filters 4021 to 402N are transferred to the echo cancellation filters 4051 to
405N on the premise that the filter coefficients are converged more rapidly and with higher
precision than the filter coefficients of the above. However, the reception signals x1 (k) to xN (k)
are not always always highly correlated. When signals from a plurality of completely different
sources are given as multi-channel reception signals, the correlation between the reception
signals may be sufficiently small. In that case, it is expected that the coefficients of the reception
signal adaptive filter converge at high speed and can be obtained with high accuracy. Therefore,
in the modified embodiment of FIG. 11, the maximum correlation value between the reception
signals x1 (k) to xN (k) is calculated, and if the maximum correlation value is larger than a
predetermined value, addition is performed as in the embodiment of FIG. The coefficients of the
signal adaptive filter are transferred to the echo cancellation filter, and if less than a
predetermined value, the coefficients of the reception signal adaptive filter are transferred to the
echo cancellation filter.
[0035]
The echo canceller 40m of FIG. 11 is provided for each of the transmission channels as shown in
FIG. 5, and the reception signals x1 (k) to xN (k) and additional signals a1 (k) to aN are provided
similarly to FIG. (6) Input the corresponding received signals x1 (k) to xN (k) and additional
signals a1 (k) to aN (k) to the adaptive filters 4011 to 401N and 4021 to 402N individually
provided for (k). The adder 403 obtains the sum y'm (k) of the outputs of all the adaptive filters.
09-05-2019
15
The sum y'm (k) of the output of this adaptive filter is subtracted from the acoustic echo ym (k)
collected by the microphone to obtain an error signal em (k), and this error signal em (k) is
weighted by the weighting unit 411. The signals w1em (k) and w2em (k) generated with the
weighting factors w1 and w2 are fed back to the adaptive filters 4011 to 401N and 4021 to
402N, the coefficients of these adaptive filters are updated, and the signals are prepared for the
next input.
[0036]
As a difference from the configuration of FIG. 9, in the actual echo cancellation, the correlation
determination unit 430 obtains the correlation between the reception signals x1 (k) to xN (k) for
all combinations, and If the value is larger than the predetermined value, the coefficients of the
adaptive filters 4021 to 402N to which the additional signals a1 (k) to aN (k) are input are input
to the processed signals x'1 (k) to the corresponding receiving channels. It transfers to the echo
cancellation filter which receives x'N (k), and when the maximum correlation value between the
above-mentioned receiving signals is smaller than a predetermined value, the adaptive filters
4011 to 401 N receiving the above-mentioned receiving signals as input A transfer selection
control signal TS is applied to the transfer control unit 414 so as to transfer the coefficients to
the echo cancellation filters 4051 to 405N to which the processing signals x'1 (k) to x'N (k) are
input. The transfer control unit 414 transfers the transfer control signal TC2 to the adaptive
filters 4021 to 402N when transferring the coefficients of the additional signal adaptive filters
4021 to 402N according to the transfer selection control signal TS when the transfer conditions
are satisfied as described above. When the coefficients of the reception signal adaptive filters
4011 to 401N are transferred, the transfer control signal TC1 is applied to the adaptive filters
4011 to 401N. Processing for the echo cancellation filters 4051 to 405N in which the
transferred filter coefficients are set is the sum of the reception signals x1 (k) to xN (k) and the
additional signals a1 (k) to aN (k) for each reception channel Signals x'1 (k) to x'N (k) are input,
and the sum y "m (k) of the outputs of the echo cancellation filters 4051 to 405N is obtained as a
pseudo echo, and from the acoustic echo ym (k), Echo cancellation is realized by subtracting the
pseudo echo y "m (k).
[0037]
In the method of FIG. 10, when the correlation between the reception signals x1 (k) to xN (k) is
small, the reception signal has a higher level than the additional signal, so the accuracy of the
adaptive filter for the reception signal is better. Are trying to improve the performance. Echo
Canceller Modification 5 FIG. 12 shows another configuration of the individually used echo
09-05-2019
16
canceller. In this configuration, similarly to the configuration of FIG. 9, each of the reception
signals x1 (k) to xN (k) corresponding to the adaptive filters 4011 to 401N and 4021 to 402N
individually provided for the reception signals and the additional signals, respectively. And each
additional signal a1 (k) to aN (k) to obtain the sum y "m (k) of the outputs of all the adaptive
filters, and this adaptation from the acoustic echo ym (k) collected by the microphone The sum
y'm (k) of the filter outputs is subtracted to obtain an error signal em (k). The error signal em (k)
is weighted by coefficients w1 and w2 by the weighting unit 411 to generate the signals w1em
(k) and w2em (k), which are fed back to the adaptive filters 4011 to 401N and 4021 to 402N.
Update the coefficients and prepare for the next input.
[0038]
As a difference from the configuration of FIG. 9, in the actual echo cancellation, the correlation
between the above-mentioned reception signals x1 (k) to xN (k) is obtained in the correlation
determination unit 480 as in the embodiment of FIG. The echo cancellation filters 4051 to 405N
receive the processed signal as input for the corresponding reception channels for the
coefficients of the adaptive filters 4021 to 402N to which the additional signal is input, and the
reception signal Are transferred to the respective adaptive filters 4011 to 401N with the input as
the input, and when the correlation between the reception signals is smaller than a
predetermined value, the coefficients of the respective adaptive filters The transfer control unit
414 is supplied with transfer selection signals TS for controlling transfer to the echo cancellation
filters 4051 to 405N to which the processing signal is input and the adaptive filters 4021 to
402N to which the additional signal is input.
[0039]
When the transfer control unit 414 transfers the coefficients of the additional signal adaptive
filters 4021 to 402N according to the given transfer selection signal TS when the transfer
conditions are satisfied, the transfer control signal TC2 is transferred to the adaptive filters 4021
to 402N. When the coefficients of the reception signal filters 4011 to 401N are transferred, the
transfer control signal TC1 is applied to the adaptive filters 4011 to 401N.
Sum of reception signals x1 (k) to xN (k) of reception channels corresponding to the filters for
echo cancellation 4051 to 405N and additional signals a1 (k) to aN (k) (that is, processed signals)
x'1 (k) .About.x'N (k) are input, and the sum y "m (k) of the outputs of the echo cancellation filters
4051 to 405N is obtained as a pseudo echo. Echo cancellation is realized by subtracting this
pseudo echo y ′ ′ m (k) from the acoustic echo ym (k). This embodiment not only has the
advantages of both FIG. 10 and FIG. 11, but also enables transfer of filter coefficients from the
09-05-2019
17
reception signal adaptive filter to the additional signal adaptive filter to obtain more accurate
filter coefficients. it can.
[0040]
FIG. 13 shows a configuration example of the correlation determination unit 430 in the
embodiments of FIGS. The correlation determination unit 430 evaluates the cross correlation
between the reception signals of the respective channels with respect to the sample points for
each frame width divided into the number of taps L of the adaptive filter. The received signals x1
(k) to xN (k) are always, for example, two frames long from the current point k to the past (ie,
time points k, k-1, ..., k-L + 1, kL, ..., k-2L + 1) A minute number of signal samples are held in the
buffer 431. The frame cutout unit 432 includes L samples xn (k), xn (k-1),..., Xn (k-L + 1)
consecutive from the reception signal in the buffer 431 in the past for each reception channel.
The reception signal sequence is cut out as a vector x n (k) (n = 1, 2,..., N). Although the reception
signal vector is shown in bold in the following equation, it can be distinguished by those skilled
in the art, so in the following description, the reception signal vector is written in a standard font,
not bold. The inner product calculation unit 433 selects two channels (assuming n = i and n = j)
to be evaluated among all the channels, and calculates the inner product of their signal sequence
vectors, for example, xi (k) and xj (k). First calculate. The normalization unit 434 divides this
inner product result xiT (k) .xj (k) by the magnitudes of the two signal sequence vectors to be
evaluated, and the cross correlation evaluation amount EC, for example,
[0042]
To obtain However, with the equation (6) as it is, when the signals having high correlation with
each other at mutually shifted positions on the time axis, appropriate evaluation amounts can not
be obtained. Therefore, the time difference can be calculated by calculating the cross correlation
evaluation amount EC each time the time k of one vector, for example, xj (k) is sequentially
shifted over k−1, k−2,. Evaluate signals with high correlation. As described above, the
correlation evaluation unit 435 compares the cross correlation evaluation amounts EC calculated
for all combinations of received signals (i, j) in all channels, and the maximum value among them
is used as the entire system. Adopted as an evaluation amount of When the cross-correlation
evaluation amount EC between at least one set of reception signals is larger than a
predetermined value, the transfer selection unit 436 selects the coefficients of the adaptive filters
4021 to 402N for the additional signals as the adaptive filters for the reception signals. It is
determined that convergence is performed with higher accuracy than 4011 to 401N, and the
adaptive filter coefficients for each additional signal are transferred to the other filters 4011 to
09-05-2019
18
401N and 4051 to 405N, and if there is no correlation evaluation amount larger than a
predetermined value, reception is received. A transfer selection signal TS for transferring the
coefficients of the signal adaptive filters 4011 to 401N to the filters 4021 to 402N and 4051 to
405N is supplied to the transfer control unit 414. The multi-channel acoustic echo canceler
according to the present invention as described above comprises received signals x1 (k) to xN (k)
given to each part constituting it and acoustic echo signal y1 (k) from a microphone ) To yM (k)
may be realized as data processing by a computer, for example. In that case, the processing of
each unit shown in the above-described embodiment is described as a program, stored in
advance in a recording medium, and the program read out from the recording medium at the
time of use is executed by a computer. An example of such a configuration is shown in FIG. A
computer 100 operating as an acoustic echo canceler shown in FIG. 14 is a computer having a
very general configuration, and a CPU (central processing unit) 110, RAM 120, hard disk 130,
interface 140, etc. Is configured. A program for executing echo cancellation according to the
present invention is stored in advance in, for example, the hard disk 130, and the program is
read into the RAM 120 at the time of operation, and the CPU 110 executes processing in
accordance with the program.
The received signals x1 (k) to xN (k) and the echo signals y1 (k) to yM (k) from the microphone
are taken in via the I / O interface unit 140, subjected to the above-described echo cancellation
processing, and subjected to echo cancellation The residual signals e1 (k) to eM (k) are outputted
through the I / O interface 140. The program for echo cancellation may be recorded on an
arbitrary recording medium 170 as the external storage device 170, and read from there onto
the RAM 120 by the driver 160 of the storage device for execution.
[0043]
Implementation Example in the Case of Two Channels Here, as an example of FIG. 9 representing
the embodiment of the multi-channel echo canceller of the present invention described above,
generation of additional signal as a system of two receiving channels and one transmitting
channel, The operation of the adaptive filter and error weighting will be described more
specifically with reference to FIG.
[0044]
In FIG. 15, parts corresponding to those in FIGS. 5 and 9 are given the same reference numerals.
[Generation of Addition Signal] An addition signal is generated for the two-channel reception
09-05-2019
19
signal x1 (k), x2 (k). When a code switching cycle C is given, a code coefficient giving a code
inversion at each count C of zero cross points is represented by .sigma.c. It is σ C = 1, -1. First,
for the reception channel 1, the code is fixed with the code switching cycle C = 0, and the
adjustment coefficient is α1 to obtain the additional signal α1 ¦ x1 (k) ¦. Also, for the receiving
channel 2, as the code switching cycle C = 2, the value of the code coefficient σ C is alternately
switched between 1 and -1 every two zero cross detection points of x 2 (k) to adjust the
adjustment coefficient Let .alpha. @ 2 be the additional signal .alpha. @ 2 .sigma.C.vertline.x @ 2
(k) .vertline. Here, in the case of the code switching cycle C = 2, instead of counting the number
of zero crossing points in the timing detection of the above code switching, the time at which the
code of x2 (k) changes from positive to negative or It is also possible to switch the code
synchronously with the time when it changes from negative to positive. "Operation of adaptive
filter" In describing the operation of the adaptive filter, the signal is vectorized. That is, in the
following description, the reception signal vector is expressed as x 1 (k), x 2 (k), and the
additional signal vector α 1 x −1 (k), α 2 x − 2 (k). However, in the formula, the vector is
shown in bold. Here, x1 (k) = [x1 (k), x1 (k-1),..., X1 (k-L + 1)] T (7) x2 (k) = [x2 (k), x2 (k) -1), ...,
x2 (k-L + 1)] T (8) x-1 (k) = [¦ x1 (k) ¦, ¦ x1 (k-1) ¦, ..., ¦ x1 (k- L + 1) ¦] T (9) x-2 (k) = [σC ¦ x2 (k) ¦,
σC ¦ x2 (k-1) ¦, ..., σC ¦ x2 (k-L + 1) ¦] T (10), where T represents the transpose of the vector. L is
the number of taps of the adaptive filter, and the coefficient vectors of the adaptive filters 4011
and 4012 to which x1 (k) and x2 (k) are input are respectively h ^ 1 (k) and h ^ 2 (k), and α1x
The coefficient vectors of the adaptive filters 4021 and 4022 to which −1 (k) and α2x−2 (k)
are input are h−1 (k) and h−2 (k), respectively. Also, an impulse response, which is a transfer
characteristic between two speakers and microphones, is modeled as a true acoustic echo path
vector h1 (k), h2 (k) of length L and used in the following discussion.
[0045]
First, the acoustic echo y (k) that gets around from the two speakers to the microphone,
[0047]
It can be expressed as:
Adaptive filters 4011, 4012, 4021 and 4022 are used to simulate this acoustic echo y ^ (k).
[0049]
09-05-2019
20
Calculate: From the comparison of the equations (12) and (13), the coefficient vector of each
adaptive filter is h ^ 1 (k) → h1 (k) (14) h ^ 2 (k) → h2 (k) (15) h- It should converge with 1 (k)
→ h1 (k) (16) h-2 (k) → h2 (k) (17), so ideally, h-1 (k) = h ^ 1 ( k), h-2 (k) = h ^ 2 (k). However, in
many cases, there is a high correlation between x1 (k) and x2 (k), and as in the conventional
method, the mutual channels affect each other, so h ^ 1 (k), h ^ 2 The convergence of (k)
stagnates. On the other hand, since α1x-1T (k) and α2x-2T (k) have low correlation with each
other and also have low correlation with x1 (k) and x2 (k), h-1 (k), h The convergence of -2 (k) is
not affected by the signals of other channels. In other words, it is considered that at least the
coefficients of h-1 (k) and h-2 (k) converge with high accuracy. Therefore, even if y '(k) is used as
a pseudo echo, some performance improvement can be expected, but the effect of h ^ 1 (k) and h
^ 2 (k) does not affect the generation of the pseudo echo, Accurate h-1 (k) and h-2 (k) are
transferred as coefficient vectors to the echo cancellation filters 4051 and 4052, respectively,
and adders 4061 and 4062 are respectively transmitted to the echo cancellation filters 4051 and
4052. By inputting x1 (k) + α1x-1 (k) and x2 (k) + α2x-2 (k) from 4062, respectively, and
generating the pseudo echo from the adder 407 as the sum of their outputs, The accuracy of the
echo can be increased. In the following, h-1 (k) and h-2 (k) are first simulated with respect to true
impulse responses h1 (k) and h2 (k) of echo paths 1511 and 1512 from the speakers 121 and
122 to the microphone 161, respectively. It is called echo path impulse response, and h ^ 1 (k)
and h ^ 2 (k) are called second pseudo echo path impulse response. Also, for these true echo
paths 1511 and 1512, adaptive filters 4011 and 4012 are called first pseudo echo paths, and
adaptive filters 4021 and 4022 are called second pseudo echo paths. Error Weighting Method of
weighting differently by the weighting unit 414 in feeding back the error signal e (k) between the
acoustic echo y (k) and the adaptive filter output y ^ (k) for updating the coefficients of each
adaptive filter Describe.
[0050]
First, each adaptive filter coefficient vector h ^ 1 (k), h ^ 2 (k), h-1 (k), h-2 (k) is expressed by the
following equation using the NLMS algorithm
[0052]
Suppose that it is updated to
Here, the adjustment coefficient is α1 = α2 = α. Further, e (k) = y (k) −y ′ (k), and μ is a
parameter called a step size. From the left on both sides, with μ = 1 for equation (18)
09-05-2019
21
[0054]
By multiplying y (k) = y '(k) + e (k) / (1 + α2) + α2 e (k) / (1 + α2) = y' (k) + e (k) (20) It
becomes. That is, if the input signal at time k is input again to the updated adaptive filter, the
error signal e (k) is compensated, and the sum of the outputs is equal to y (k). Here, the second
and third terms on the middle side of equation (20) are the sum of the inner products when the
second term on the right side of equation (18) is multiplied by equation (19) from the left
[0056]
(20) is the adaptive filter coefficient vector h ^ 1 (k + 1), h ^ 2 for x1 (k), x2 (k). The
compensation term according to (k + 1), and the third term of the middle side is the adaptive
filter coefficient vector h-1 (k), h-2 (for the additional signals α1x-1 (k), α2x-2 (k) It turns out
that it is the compensation part by k). Normally, the value of α is small, for example, about α =
0.2. Therefore, equation (18) is more susceptible to the correlation of the reception signals than
h−1 (k + 1) and h−2 (k + 1) which are expected to have high convergence accuracy. It is
understood that updating is performed with emphasis on (k + 1) and h ^ 2 (k + 1).
[0057]
Therefore, the weighting unit 411 performs e in the update of h ^ 1 (k + 1), h ^ 2 (k + 1) and the
update of h-1 (k + 1), h-2 (k + 1). (k) are respectively multiplied by different weighting factors w1
and w2. That is, the adaptive filter update equation
[0060] 【0060】とする。 Thus, e (k) = (1 / (1 + α2)) w1 e (k) + (α2 / (1 + α2)) w2 e (k)
(24) is obtained. Then, if, for example, w1 = (1 + α2) / 2 (25) w2 = (1 + α2) / (2α2) (26),
equation (24) is e (k) = (1/2) e (k) + (1/2) e (k) (27), and the adaptive filter coefficient vector h ^
1 (k + 1), h ^ 2 for x1 (k), x2 (k) of the first term on the right side Adaptive filter coefficient
vector h-1 (k + 1), h-2 (k +) for the compensation by (k + 1) and the additional signal α1x-1 (k),
α2x-2 (k) of the second term on the right side The compensation amount according to 1) is
equal. Further, since w1 and w2 in the equations (25) and (26) are w1: w2 = 1: (1 / α2), the
power of x1 (k) and x2 (k), α1x-1 (k), respectively. , Α 2 x − 2 (k) are weighted in inverse
proportion to the power.
09-05-2019
22
[0061] When α is too small, if w1 and w2 are selected as in the equations (25) and (26), w2 may
become very large and the calculation may become unstable. Therefore, it is preferable to
introduce a relaxation coefficient β and set the update formula of the adaptive filter as the
following equation.
[0064] Equations (28) and (29) show computer simulation results in the case where the adaptive
filter is updated, in the case where the input is Gaussian white noise and for the case of male
speech in FIGS. 16A and 16B. Curve (a) represents the normalized squared error of h ^ 1 (k), h ^
2 (k) according to the invention with respect to h1 (k), h2 (k), curve (b) is h- according to the
invention Represents a normalized squared error of 1 (k), h-2 (k). Curve (c) is an adaptive filter h
^ 1 (h) for h1 (k) and h2 (k) in the conventional method when x1 (k) + α1x-1 (k) and x2 (k) +
α2x-2 (k) are input. k), the normalized squared error of h 2 (k). The parameters were α = 0.2, β
= 0.7, μ = 0.5, the adaptive filter length was 1000 taps each, and noise was added to the acoustic
echo so that the SN ratio was 30 dB. From this, it can be seen that the convergence performance
of h-1 (k) and h-2 (k) for additional signals is excellent.
[0065] Although the embodiments of the present invention described above have described echo
cancellation in the teleconference system as an example, as described above, the principle of the
present invention is that the reproduction from the speaker superimposed on the target acoustic
signal picked up by the microphone It is effective in removing sound. Therefore, in an acoustic
system in which a plurality of reproduction channels each including a speaker and at least one
pickup channel including a microphone for picking up a target sound are provided in a common
sound field, the speaker picked up by the microphone The present invention can be applied to
any system for the purpose of erasing reproduced sound.
[0066] According to the multi-channel acoustic echo cancellation method, the echo cancellation
performance can be improved by the previously proposed method of applying the additional
signal to the reception signal, but the additional signal can be made larger due to the auditory
restriction. Also, since the processing signal which is the sum of the reception signal and the
additional signal is used in the generation of the pseudo echo for echo cancellation, the
important information contained in the additional signal is buried in the reception signal, and the
echo cancellation performance The amount of improvement was limited.
[0067] According to the method of the present invention, first, after generating an appropriate
additional signal, instead of using the above processing signal in the generation of a pseudo echo
for echo cancellation, the receiving signal and the additional signal are processed separately. The
important information contained in the additional signal can be easily used. Thereby, the pseudo
09-05-2019
23
echo for echo cancellation can be generated with high accuracy. Therefore, the echo cancellation
performance can be improved as compared with the conventional method.
09-05-2019
24