Pulse or digital communications – Equalizers – Automatic
Reexamination Certificate
2000-03-08
2003-11-25
Pham, Chi (Department: 2631)
Pulse or digital communications
Equalizers
Automatic
C375S222000, C370S290000, C379S406080, C379S406090, C708S322000
Reexamination Certificate
active
06654412
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention generally relates to a learning method of an adaptive filter, and an adaptive filter provided with that learning method. More specifically, the present invention relates to a learning method of an adaptive filter with a predictive updating scheme for acoustic echo cancellation, and to an adaptive filter provided with that learning method.
2. Description of the Related Art
Multi-path echoes of an operating environment result in uncomfortableness to the far-end user in hands-free communications. For example, while a person using a cellular phone without holding it in hand, the receiver (or microphone) receives sounds generated from the loudspeaker and conducted/reflected by the environment. Finally, such sound will transmit to the far-end user and, after a period of time, the far-end user hears the sound previously spoken by the user. Such a phenomenon is called the acoustic echo effect. The effect should be reduced or eliminated by an apparatus, e.g., an acoustic echo canceller.
The most popular way to cancel the acoustic echo effect is to set up an acoustic echo canceller between two signal lines, as shown in FIG.
1
.
FIG. 1
is the diagram illustrating the acoustic echo effect and the installation of an acoustic echo canceller
24
in the art. The input signals x(t) generated from a far-end input signal source
10
are sent to a loudspeaker
14
via a first signal line
12
for generating sound. The sound is reflected or conducted by the environment
16
and enters the microphone
18
to generate real output signals d(t), that is, an acoustic echo. The real output signals d(t) are sent to the loudspeaker
22
of the far-end user via a second signal line
20
, creating the acoustic echo effect. The acoustic echo canceller
24
is installed between the first and the second signal lines (
12
and
20
, respectively) to learn the room response of the environment
16
. The acoustic echo canceller
24
generates estimated signals y(t) according to the input signals x(t) at the first signal line
12
and sends the estimated signals y(t) to the second signal line
20
for canceling the real output signals d(t). Error signals e(t) are defined as the error between the real output signals d(t) and the estimated signals y(t). If the real output signals d(t) and the estimated signals y(t) are very close, then the error signals e (t) that the far-end user hears are small, thus the acoustic echo effect is eliminated. However, the environment
16
is time variant, the acoustic echo canceller
24
needs to automatically trace the changes of the environment. As shown in
FIG. 1
, the acoustic echo canceller
24
updates its own signal response according to the error signals e(t). Such an acoustic echo canceller
24
with a self-learning function is called an adaptive filter. In particular, an adaptive filter with finite impulse response (FIR) is called an adaptive FIR filter.
FIG. 2
is a functional diagram of an adaptive FIR filter according to the prior art. For illustrative purpose, the operation of the system according to
FIG. 2
is a digital system where the continuous time variable t is substituted with a discrete time variable k. x(k)
30
represents the input signals, while H represents the impulse response of the environment. Since the environment is not predictable, the environment is defined as an unknown system
40
. H
k
, the estimated result of H, represents the impulse response of the adaptive FIR filter
34
at the time of k. y(k)
32
represents the estimated output signals generated by the adaptive FIR filter
34
, and d(k)
38
represents the real output signals generated by the unknown system
40
. For the adaptive FIR filter
34
, at least
3
steps are needed to update H
k
per iteration, as follows:
Step 1: generating the estimated output signals y(k)
y
(
k
)
=
H
k
T
conv
X
(
k
)
≡
∑
l
=
0
L
-
1
h
(
k
)
*
x
(
k
-
l
)
(
1
)
X(k)≡[x(k−L+1), . . . , x(k−1), x(k)]
T
H
k
≡[h
k
(0), h
k
(1), . . . , h
k
(L−1)]
T
Wherein the superscribed “T” denotes the transport operation, and L denotes the length of the adaptive FIR filter.
Step 2: calculating the error signals e(k)
36
e
(
k
)≡
d
(
k
)−
y
(
k
) (2)
and;
Step 3: updating the coefficients of H
k
H
k+1
=&PHgr;[H
k
] (3)
Wherein, &PHgr;[.] denotes the updating algorithm.
If the normalized least-mean-square (NLMS) algorithm is applied, for example, then
H
k+1
=H
k
+&mgr;/|X|
2
*e
(
k
)*
X
(
k
) (3a)
Wherein &mgr; denotes a fixed step size for updating. From the Eq.(1), Eq(2) and Eq.(3), (2L+1) operations of multiplication and 2L operations of addition are needed for learning per iteration for the adaptive FIR filter
34
with such an NLMS algorithm. However, only one time of learning and one updating the coefficients of H
k
is achieved after the adaptive FIR filter receives one of the input signals x(k) when such an NLMS algorithm is used. Thus, it is well-known that the efficiency, or more particularly, the convergence rate of such an NLMS algorithm is very slow when the input signals are color.
In addition to the learning method using the NLMS algorithm as mentioned above, the prior art provides a multiple learning method performing multiple updating using the input signals x(k) received. For the purpose of multiple updating, the input signals x(k)
30
and the real output signal d(k)
38
must be stored. Given that M times of revisions are desired, the stored input signal x(k)
30
can be represented as
X
(
k
),
X
(
k
−1),
X
(
k
−2), . . . ,
X
(
k−M
+1).
The stored real output signal d(k)
38
can be represented as:
d
(
k
),
d
(
k
−1),
d
(
k
−2), . . . ,
d
(
k−M
+1).
Thus the algorithm of the multiple learning method becomes:
For m=1 to M
Update
H
k
using
X
(
k−m
+1) and
d
(
k−m
+1)
End
The adaptive FIR filter
34
experiences M revisions after receiving each of the input signals x(k)
30
. These additional revisions can significantly speed up the convergence rate of the NLMS algorithm. However, M revisions require M*(2L+1) operations of multiplication and 2*M*L operations of addition. Such an algorithm leads to an huge number of calculations and results in a complex system.
Therefor, there is a general need for an adaptive filter and a learning method thereof that overcome the aforementioned problems in the art. In particular, an adaptive filter and learning method thereof are needed to improve the adaptation of the adaptive filter with a relatively low number of operations, and to achieve an optimized convergence rate.
SUMMARY OF THE INVENTION
The present invention provides a learning method for canceling real output signals generated by input signals and an environment signal response. The adaptive filter receives the input signals from an input signal source and generates estimated output signals according to an filter coefficients and the input signals. The first step of the learning method according to the present invention is storing the input signals and the real output signals within a past time period. Then a predictive input signal is generated according to the input signals to predict one of the input signals at the next sampling duration. A predictive output signal is also generated according to the real output signals to predict one of the real output signals at the next sampling duration. Finally, using the predictive input signal, the predictive output signal, the input signals and the output signals as data, the filter coefficients is updated, causing the estimated signals to approximate the real output signals.
The present invention also provides an adaptive filter for canceling real output signals generated by input signals and an environment signal response. The adaptive filter receives the input signals
Chen Heng Chou
Chen Oscal Tzyh-Chiang
Industrial Technology Research Institute
Pham Chi
Tran Khanh Cong
LandOfFree
Adaptive filter and learning method thereof does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Adaptive filter and learning method thereof, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Adaptive filter and learning method thereof will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3180598