 |
Description  |
|
|
BACKGROUND
This invention pertains to adaptive hybrids which serve to, for example,
eliminate or reduce "talker" echoes on telephone channels.
Prior art adaptive hybrids and their uses are described, for example, in
Messerschmitt, David G., "An Electronic Hybrid for Adaptive Balance In
Telephony," IEEE Transactions on Comm., August 1980, pp. 1399; Dotter, et
al., "Implementation of an Adaptive Balancing Hybrid," COM-28, p. 1408,
August 1980; and White, Stanley A., "An Adaptive Electronic Hybrid
Transformer," IEEE Transactions on Comm.," December 1972. However, the
previous schemes have two major shortcomings:
1. They are not sufficiently versatile and effective against the widely
differing echo path characteristics encountered on telephone channels.
2. The adaptive algorithms, in general, tend to be costly to implement on a
per-stage basis; and the effectiveness is usually limited by using only
one or two adaptive stages.
FIG. 1 shows a block diagram of a widely known prior art adaptive hybrid
100. Adaptive hybrid 100 is trained on the signal from the local
transmitter received via port 103. This is usually done while the signal,
normally received from the far end, is absent.
Filters 101 and 102 approximate the characteristics of nominal short and
nominal long "near-end" echo paths, respectively. Interpolator 104 is
automatically adjusted in response to the operation of adaptive algorithm
105 in an attempt to provide an echo estimate to summer 106 that closely
approximates the echo. Circuit 106 subtracts the estimated echo from one
output of the fixed hybrid, and this output normally includes echo plus
received signal.
A major limitation of the prior art adaptive hybrid 100 of FIG. 1 arises
from the fact that a substantial percentage of echo paths (even near-end
echo paths) differ widely from any characteristic that is capable of being
provided as an interpolation between the two filters 101 and 102. This
fact is indicated by the scatter diagrams in Gresh, P. A. "Physical and
Transmission Characteristics of the Customer Loop Plant" BSTJ, Vol. 48,
pp. 3337-3385, December 1969; and Manhire, L. M. "Physical and
Transmission Characteristics of the Customer Loop Plan," BSTJ, Vol. 57, pp
33-59, January 1978. Therefore, the prior art approach of FIG. 1 provides
little, if any, attenuation of some echoes.
SUMMARY
Echo attenuation facilitates achievement of efficient, full-duplex data
communications on two-wire channels, for example in a V.32 modem. While
echo cancellation can be achieved without an adaptive hybrid, some major
advantages can be achieved by using an adaptive hybrid in conjunction with
echo cancellation. When needed, the adaptive hybrid reduces the amplitude
range requirements on the echo canceller and on analog-to-digital
conversion, thereby reducing overall cost. This reduction in range
requirements reduces the digital word-size required for high-performance
echo cancellation and reduces the bit-accuracy needed in the
analog-to-digital converter. These reductions in needed word-size and bit
accuracy substantially reduce implementation cost. In a V.32 modem, for
example, the adaptive hybrid can make it feasible to use a readily
available, 16-bit digital signal processor to implement a high-performance
echo canceller.
Normally, the adaptive hybrid is used to reduce near-end echoes, which are
usually much larger than far-end echoes. The echo canceller attenuates the
remaining near-end echo and the far-end echo. Two major objectives are:
(1) Cost effectiveness and (2) versatile, effective correction of echoes
with various, realistic characteristics. Both objective are achieved in
accordance with the teachings of this invention by employing simple but
multiple stages. A special training algorithm is also used in accordance
with the teachings of this invention.
BRIEF DESCRIPTION OF DRAWINGS
FIG. 1 is a block diagram of a typical prior art adaptive hybrid;
FIG. 2 is a block diagram one embodiment of an adaptive hybrid constructed
in accordance with the teachings of this invention, shown within a modem
structure as a means of example;
FIG. 3 is a block diagram depicting one embodiment of a correction network
suitable for use in the embodiment of FIG. 2;
FIG. 3a is a diagram depicting an alternative connection of the correction
network of FIG. 3;
FIG. 4 is a graph depicting one quadrature component of the frequency
response of one stage of this invention;
FIG. 5 is graph depicting the other quadrature component of the frequency
response of the same stage of this invention; and
FIG. 6 is a block diagram depicting another embodiment of an adaptive
hybrid constructed in accordance with the teachings of this invention.
DETAILED DESCRIPTION
Overview
In this specification, the following nomenclature is used.
U=Echo before correction
U=Estimate of U
E=U -U=Error in echo correction provided by the adaptive hybrid
r=Signal from remote transmitter
Sgn=Polarity of
FIG. 2 is a block diagram depicting one embodiment of an adaptive hybrid
201 used within a V.32 modem 200, which will serve as but one example of a
system constructed in accordance with the teachings of this invention.
The number of stages, N+1, can be selected for a particular application. In
the V.32 modem, N=3 is a cost-effective choice. The choice of N involves a
tradeoff of cost versus ability to highly and consistently attenuate
echoes with various characteristics.
In the embodiment of FIG. 2, the echo path is from node 212 through analog
features 210 and hybrid 254 to one or more reflection features along the
transmission channel, from which the echo returns through hybrid 254 to
analog features 211 and to differencer 251. The echo correction path
provided by adaptive hybrid 201 is from node 212 through N+1 stage
correction network 252, and then to differencer 251.
In one training procedure, adaptive hybrid 201 is trained before echo
canceller 203 is trained. Thereafter, adaptive hybrid 201 remains fixed
until a new training is called for. Alternatively, training of adaptive
hybrid 201 can partly overlap in time with the training of echo canceller
203.
The learning feature of adaptive hybrid 201 of FIG. 2 automatically adjusts
correction network 252 (which includes N+1 DACs subcircuit 254) so that
its characteristics approximate those of the near-end, "talker" echo path.
Then, since the same signal goes through both the echo path and correction
network 252, a large part of the echo is cancelled in differencer 251. In
one embodiment, adaptive hybrid learning subcircuit 253 is digital and is
implemented in a digital signal processor (DSP), which can concurrently
implement various other modem functions, if desired.
Correction Network
FIG. 3 depicts one embodiment of correction network 252 suitable for use in
the embodiment of FIG. 2. The embodiment of FIG. 3 is an N+1 stage
correction network shown with learning subcircuit 253 of FIG. 2 receiving
digital information pertaining to voltage levels V.sub.O through V.sub.N
and, in response thereto and the resulting signal from differencer 251,
providing adjustable gain factors .alpha..sub.O through .alpha..sub.N. The
adjustable gain factors .alpha..sub.O through .alpha..sub.N are learned
during training. In one embodiment, .alpha..sub.O through .alpha..sub.N
then remain fixed during reception of subsequent signals, including data
signals. In an alternative embodiment, one or more of .alpha..sub.O
through .alpha..sub.N continually adapt during the reception of subsequent
signals. Impedances Z.sub.n-1, Z.sub.n-2 associated with gain factor
.alpha..sub.n may be any suitable impedances, for example a resistor and a
capacitor, respectively. The best selection of impedance values depends
upon the application, but the following explains a general approach. An
echo path characteristic can be represented in the frequency domain by
H(.omega.)=R(.omega.)+jI(.omega.), .omega.=2.pi.f (1)
where j=-1 while R(.omega.) and I(.omega.) represent two quadrature
components, each a function of frequency. On some echo paths, which are to
be corrected, R(.omega.) and I(.omega.) are curves that have large bends
within the bandwidth of the transmitted signal. Then, the RC values in
correction network 252 need to be relatively large. With large RC values,
correction network 252 can be automatically adjusted, by adjusting the
.alpha.'s so that this network coarsely approximates any desired
R(.omega.) and I(.omega.) within wide limits. However, on more typical
echo paths, R(.omega.) and I(.omega.) have lesser bends. Then, the RC
values can be smaller in order to produce better accuracy of approximating
the needed correction characteristic.
There is no generally optimum set of RC values and no definitely best RC
selection procedure. However, RC values that lead to reasonably accurate
correction of various echo path characteristics can be selected by
trial-and-error, as explained below. Since the RC values need be selected
only once for a particular application, trial-and-error is tolerable.
The characteristic that the adaptive hybrid corrects is the transfer
function of the echo path, FIG. 2, which was given by Eq. (1). The
transfer function of the n-th RC network of correction network 252 is
##EQU1##
The transfer function of the nth stage of correction network 252 is
.alpha..sub.n H.sub.n (.omega.).
In FIG. 3, the transfer functions of the individual stages (exclusive of
the .alpha.'s) are given in Table 2.
TABLE 2
______________________________________
Stage R.sub.n (.omega.)
I.sub.n (.omega.)
______________________________________
0 1 0
##STR1##
##STR2##
2
##STR3##
##STR4##
______________________________________
Stage 3 is the same as Stage 1 except Stage 3 uses R.sub.3 C.sub.3 instead
of R.sub.1 C.sub.1 and uses .alpha..sub.3 instead of .alpha..sub.1.
In the embodiment described in Table 2, the zeroth stage includes zero
impedance Z.sub.0,1 and an infinite impedance Z.sub.0,2, thereby providing
R.sub.n (.omega.)=1 and I.sub.n (.omega.)=0.
FIGS. 4 and 5 show, as examples, representative plots R.sub.n (.omega.) and
I.sub.n (.omega.) for a stage of the type of stages 1 and 3.
The overall transfer function of correction network 252, with N+1 stages,
odd N, is
H.sub.c (.omega.)=R.sub.c (.omega.)+j I.sub.c (.omega.) (3)
where
R.sub.c (.omega.)=.alpha..sub.0 +.alpha..sub.1 R.sub.1 (.omega.)
+.alpha..sub.2 R.sub.2 (.omega.)+. . . +.alpha..sub.n R.sub.N (.omega.)
(4)
I.sub.c (.omega.) =.alpha..sub.1 I.sub.1 (.omega.)+.alpha..sub.2 I.sub.2
(.omega.)+. . . +.alpha..sub.n I.sub.N (.omega.) (5)
Now, before proceeding to describe a suitable RC selection procedure,
understanding can be enhanced by observing that we could do the following:
1. Select any reasonable set of N of the RC values, where the values within
this set differ widely from each other.
2. Select any realistic echo path for R(.omega.) and I(.omega.), for
present illustrative purposes.
3. Select
##EQU2##
frequencies within the signal (and echo) bandwidth, assuming N is odd, for
example. Good choices for a voice-band channel and for N=3, for example,
are f=800 hertz and f=2800 hertz.
4. At each of the selected frequencies set R.sub.c (.omega.)=R(.omega.) and
I.sub.c (.omega.)=I(.omega.). Thereby, obtain
##EQU3##
equations from Eq. (4) and
##EQU4##
equations from Eq. (5).
5. Solve the N+1 equations simultaneously for the .alpha.'s. This procedure
gives correction characteristics, R.sub.c (.omega.) and I.sub.c (.omega.)
that exactly equal the desired correction characteristics R(.omega.) and
I(.omega.), respectively, at the two selected frequencies. Investigation
of various cases has shown that, in general, at the other frequencies
within the echo bandwidth, the echo will then be substantially reduced
relative to no echo correction.
However, it is better to refine this procedure as follows, where N=3, for
example. First, let us consider how to select three good RC values to use
at the start of a trial-and-error refinement procedure. To do this, select
about 3 or 4 echo path characteristics, H(.omega.), that differ widely
from each other, but are all exceptionally severe. Start with one of these
paths and estimate Mean [R(.omega.)], where the mean is taken across the
echo bandwidth. Plot R(.omega.)Mean [R(.omega.)] versus .omega.. Compare
the results with the curves of FIG. 4. Apply visual, coarse interpolation
between the curves and select an associated RC value that, when used for
stage 3, enables that stage alone to approximate
R(.omega.)Mean[R(.omega.)]. Use this value for R.sub.3 C.sub.3 in the
first RC set trial.
Next, set R.sub.1 C.sub.1 =0.33R.sub.3 C.sub.3 and R.sub.2 C.sub.2 =0.133
R.sub.3 C.sub.3. Use equations such as those in Table 2 for R.sub.n
(.omega.) and I.sub.n (.omega.), for n=1, 2 and 3. Substitute these
equations into equations (4) and (5). Substitute two values of .omega.
(say .omega.=1600.pi. and .omega.=5600.pi.) into these equations and, at
each of these values, set R.sub.C (.omega.)=R(.omega.) and I.sub.C
(.omega.)=I(.omega.). Solve the resulting four equations simultaneously
for the .alpha.'s. Using these .alpha.'s in Equations (4) and (5), solve
for R.sub.C (.omega.) and I.sub.C (.omega.). Plot these characteristics
and compare with the R(.omega.) and I(.omega.) of the selected severe echo
path.
Repeat the above procedure with each of the other selected, severe echo
paths. Observe the corrections of the severe paths and determine which of
these paths is most poorly corrected, considering both R(.omega.) and
I(.omega.). From observation of this most poorly corrected echo path,
choose a new set of RC values, then try this set against all of the severe
echo paths.
Try different sets of RC values; and, for each set, see how well correction
network 252 corrects each of the severe echo paths. After a little
practice, the results obtained from each set of RC values can be used to
improve the next set, except when the RC set is near optimum. In addition
to observing the accuracy of the correction, it is important to avoid any
.vertline..alpha..vertline.<1 in order to avoid implementation difficulty.
At each point in the procedure, a major objective should be to reduce the
error in correction of that echo path which is most poorly corrected. This
procedure should be continued until further trials result in little
further improvement.
One alternative procedure is similar to the above except as follows: (1) It
uses an adaptive hybrid and simulates each of several severe echo paths,
then uses the adaptive algorithm described below to select the .alpha. set
associated with each RC set. (2) It uses a spectrum analyzer to observe
the error in the echo correction as a function of frequency. The
characteristics of the individual stages preferably differ substantially
from each other, in order to facilitate adaptivity.
In one embodiment of this invention suitable for use in a V.32 modem, the
selection of RC values is depicted in Table 1.
TABLE 1
______________________________________
R.sub.1 C.sub.1 = 5 .times. 10.sup.-5 ohm-farads
R.sub.2 C.sub.2 = 2 .times. 10.sup.-5 ohm-farads
R.sub.3 C.sub.3 = 1.5 .times. 10.sup.-4 ohm-farads
______________________________________
In applying any procedure for selecting the RC values, it should be noted
that the adaptive hybrid might be used in conjunction with a fixed part of
the hybrid, where the fixed part corrects some nominal curves of
R(.omega.) and I(.omega.). The RC values in Table 1 are for use when the
fixed part of the hybrid consists essentially of a 600 ohm resistor in
series with the common hybrid transformer.
With the RC values in Table 1, correction network 252 has been applied
analytically to various echo paths, some severe. The .alpha.'s were
obtained by solving simultaneous equations of the form of Equations (4)
and (5), as described earlier. It has been found that correction network
252 consistently provides excellent accuracy of correction. A comparison
of frequency-domain plots of characteristics of echo paths and echo
correction indicates that correction network 252 can reduce the overall
echo by at least 20 db on severe paths, and can reduce it by much more on
most echo paths. These results have been verified by testing an
experimental adaptive hybrid (using the adaptation algorithm to obtain the
.alpha.'s) on a few selected echo paths.
In an alternative embodiment, the structure of FIG. 3a is used except
subsequent stages received their input signal from the common node of the
impedances of the previous stage. Thus, for example, FIG. 3a depicts an
alternative embodiment of FIG. 3, showing just the impedances and their
interconnections.
Alternative Embodiment
FIG. 6 is a block diagram which depicts another embodiment of this
invention. Circuit 400 of FIG. 6 includes a number of subcircuits which
are numbered similarly to certain subcircuits in FIG. 2.
Analog features subcircuits 210 and 211 are not a part of the adaptive
hybrid, but are included, if desired, for purposes such as filtering and
scaling. During adaptive hybrid training, the signal r is ordinarily
absent, so the output from fixed hybrid 254 consists mainly of echo U,
although some noise is also present.
N+1 stages are shown, where each stage, except one, consists of a simple RC
network and a multiplying DAC. Then, Summer 234 combines the outputs of
the N+1 multiplying DACs 254 to provide an error correction signal
##EQU5##
Uis subtracted from U+r by differencer #1. The training algorithm
implemented by training algorithm subcircuit 253 adjusts the values of
.alpha..sub.1 through .alpha..sub.n in order that
U.apprxeq.U, approximately, (7)
so the bulk of the near-end echo is corrected. In one embodiment,
differencer 251 is implemented utilizing switched capacitor techniques, as
is well known in the art.
The voltages V.sub.O through V.sub.N and the input signal to ADC 205 are
typically sampled at the same rate. Typically, the sampling rate is 7200
or 9600 sps for voice-band signals, although any suitable sampling rate
can be used.
The subcircuits labeled Sgn.sub.O through Sgn.sub.N provide polarities of
samples of voltages V.sub.O through V.sub.N, respectively, to training
algorithm subcircuit 253. A major advantage of the algorithm of this
invention is that it can make these one-bit numbers SgnV.sub.O through
Sgn.sub.N suffice.
In an alternative embodiment of this invention, a typical prior art
learning algorithm for echo cancellation and adaptive equalization is used
in an adaptive hybrid in conjunction with the correction network
constructed in accordance with the teachings of this invention. However,
this alternative embodiment requires analog-to-digital conversion of each
sample of V.sub.O through V.sub.N, with several bits of accuracy per
conversion. The novel algorithm of this invention which allows simple
polarity decisions to be used instead of multi-bit analog-to-digital
conversions is now described.
The Algorithm For Learning
One embodiment of a novel algorithm for learning (or "training") of the
adaptive hybrid of this invention is now described. The novel algorithm of
this invention is suitable for implementation in dedicated circuitry or in
a digital signal processor (DSP). If desired, the adaptive hybrid can
time-share the DSP with various other modem functions.
For most purposes, the adaptive hybrid should be trained with a random, or
nearly random, signal. Then, after its training, it will work with various
types of signals (such as data signals, tones, and repetitive patterns)
when transmitted by the local transmitter. However, if and when the
Adaptive Hybrid is to be used with only one type of signal, it will
perform better after being trained on that type of signal. The following
embodiments of the algorithms of this invention apply generally to various
types of locally transmitted signals, and are also highly suitable for
training on the signal specified by CCITT for echo canceller and/or
equalizer training in the V.32 modem. The first part of the time span of
the CCITT echo canceller training signal is used for the adaptive hybrid
training. The adaptive hybrid adaptive training time varies, but is about
50 to 100 milliseconds in the V.32 modem.
In a simple embodiment of the learning algorithm, each gain coefficient
.alpha..sub.O through .alpha..sub.N is updated once each sample time, as
follows:
(.DELTA..alpha.).sub.m,n =kE.sub.m SgnV.sub.m,n
.+-.k.vertline.E.sub.m.vertline. (8)
.alpha..sub.m,n =.alpha..sub.m-1,n +(.DELTA..alpha.).sub.m,n (9)
where
is a small constant
m refers to the m.sup.th sample
n refers to the n.sup.th stage
V.sub.n =n.sup.th stage signal at one input to a multiplying DAC
E.sub.m =Output error=U.sub.m -U.sub.m +N.sub.m
U.sub.m =m.sup.th echo sample
U.sub.m =m.sup.th echo correction sample
N.sub.m =Instantaneous value of noise component of the m.sup.th sample
Initially, .alpha..sub.O =0.5, .alpha..sub.n =0 for n.noteq.0 in the
simplest case.
In one embodiment, a refined version of the algorithm described above is
used:
##EQU6##
Updating of gain coefficients occurs once every four sample times and each
update is based upon the last 32 values of E.sub.p SgnV.sub.p,n. This
embodiment improves each updating loop characteristic. Also, it reduces
the digital word size requirement, compared to the simple embodiment
described above, because it allows the error E to become smaller before
.DELTA..alpha. becomes smaller than the least significant bit.
Also, instead of a constant, k.sub.q is reduced once every 32 sample times,
as follows:
##EQU7##
and m=0 at the start of training. Typical values are I=0.05 and a=0.96.
This procedure allows the training to proceed rapidly at first, then slow
down for higher precision.
Various alternative embodiments of this algorithm are contemplated. For
example, in one embodiment, the individual stages employ different values
of I and/or q in Equation (11). In an alternative embodiment, the number
32 and/or the frequency of updating is altered.
The following shows the basis of the correction. Let the echo path
characteristic be R(.omega.)+jI(.omega.), j=-1. Each (nth) corrector stage
has a characteristic R.sub.n (.omega.)+j.sub.I.sub.n (.omega.). The
overall corrector has the characteristic R.sub.c (.omega.)+jI.sub.c
(.omega.), given by Equations 3, 4 and 5.
For a given, realistic echo path, the .alpha.'s can be adjusted so that
R.sub.c (.omega.) approximates R(.omega.) and I.sub.c (.omega.)
approximates I(.omega.). Then, the correction network characteristic
approximates the echo path characteristic, as desired.
The basis of the learning algorithm is easier to understand if we start
first with the simple version, Equations 8 and 9.
Let g.sub.p be the optimum of .alpha..sub.p, then, neglecting noise
##EQU8##
For the case of p=n, V.sub.n SgnV.sub.n is always positive. Therefore,
from Eq(12), the term (-.alpha..sub.m,n +g.sub.n)V.sub.n SgnV.sub.n
consistently has the correct polarity to reduce the error, .alpha..sub.n,
in the n.sup.th stage. In the case of a random data signal, for example,
the effects of the terms where p.noteq.n tend to average to relatively
small values, except when some of the other stages error, some of the
{g.sub.p -.alpha..sub.p }, .noteq.n, predominate. In the case where all of
the individual stage errors are nearly equal, since the increments (the
{.DELTA.a}) are very small, each .alpha. is driven generally in the
correct direction; i.e., considerably more than half of the increments
will be in the correct direction until the errors become very small. When
some of the individual stage errors predominate, these errors will be
driven relatively strongly toward zero until they cease to predominate.
The updating of each stage, each .alpha., can be visualized as an
acquisition loop. The refined algorithm based upon Eq(10) improves each
updating loop characteristic. Also, it reduces the digital word size
requirement, compared to the simple version, because it allows the error E
to become smaller before .DELTA..alpha. becomes smaller than the least
significant bit.
All publications and patent applications mentioned in this specification
are indicative of the level of skill of those skilled in the art to which
this invention pertains. All publications and patent applications are
herein incorporated by reference to the same extent as if each individual
publication or patent application was specifically and individually
indicated to be incorporated by reference.
The invention now being fully described, it will be apparent to one of
ordinary skill in the art that many changes and modifications can be made
thereto without departing from the spirit or scope of the appended claims.
For example, in the embodiment of FIG. 2, the fine echo correction can be
eliminated, if desired, simply by eliminating digital to analog converter
271, differencer 234, and associated error correction learning. In
addition or alternatively, the programmable gain stage 272 can be
eliminated, if desired, as can AGC stage 204.
* * * * *
|
|
 |
|
 |
Description |
|
