Академический Документы
Профессиональный Документы
Культура Документы
Invited Paper
-I { : -
led to development of the echo candler. Aduprive echo cancellation
is accomplished by automaticallysynthesizingareplica of the echo
path response,convolving the incoming signal with the replica andsub-
TALKER
A ti-_ __-
TALKER
B
tmcting the resulting echo estimate from the echo path output. Trans- LZ ECHO OF AS SPEECH
mission is not thereby interruptedin any way and two-way information
transnission proceeds normally with little or no echo. Detailed mathe Fig. 1 . A typical long distance telephone circuit. Each line represents a
maticalformulations for theconvergenceproperties of suchdevices two-wire circuit. The boxes marked N are balancing impedances. In
allow understanding of the effects of various hardware configwtions a local call, the two-wire circuit of one customer is connected diiectly
and input signal properties. Echo cancellers may be applied to the net- t o the two-wire circuitof the other.
work in a number of ways and the development of a VLSI canceller
chip
promisesinexpensiveimplementation of thew applications.
Worldwide tests have shown the effectiveness of candlation methods the normal acoustic coupling which is blocked by covering the
..
for satellitetransmLIslon. ear with the earpiece and makes the telephone circuit seem
alive. At the other extreme, if an echo is delayed by hun-
dreds of milliseconds, it might be objectionable to some
I. INTRODUCTION
! I I
I 1
OUT- - I ECHOSUPPRESSOR
r----1
cluding the terrestrial portions, therefore, the round-trip delay
in a phone conversation relayed via satellite can be as long as
\I ; a
1 II
600 ms. If, as is quite conceivable, twosatellite relays are used
for a call, the delay can be twice this value.
When a conversationis conducted over a path with such long
delays, the interruption rate increases [ 11 as does the sensi-
tivity to improperoperation of echocontrol devices(see
Section 11). This degrades the performance of echo suppres-
sors onsuch circuits. Sincethe early1960s therefore, a L----J
number of investigators have considered new approaches to ECHO CANCELLER
controlling echo on circuits with long delays [4]. This effort Fig. 4. How an echo canceller might replace a suppressor.
has received further impetus from the launching of the domes-
tic satellite program in 1976.
One of the new ideas that has been proposedis that of selec- We turn now to a discussion of the subjective effects of
tive adaptive attenuation of different frequency components delay and echo.
in the echo with the intent of minimizing mutilation of the
interrupters speech. The most sophisticated of these devices II. EFFECTSOF ECHO AND DELAY
is themultibandcenter clipper, proposed byMitchell and Transmission delay, circuit echo,andecho delay are all
Berkley [ 51, which we will discuss in Section 111-C. parameters of a normal telephone circuit. They produce vary-
The proposal that we believe has the greatest potential is ing effects on telephone users depending on their absolute and
that of echo cancellation as opposed to echo suppression. The relative magnitudes.
idea is illustrated in Fig. 4, which is similar to Fig. 3 with the In general, transmission delay alone is difficult for partici-
suppressor replacedby a canceller. Instead of breaking the pants in a two-way conversation, even in a laboratory situa-
path L2 a synthetic echois generated from As speech and sub- tion, to detect. Even so, experiments have shown that some
tracted from the signal going out on Lz. If the synthetic echo impairment in communication quality does occur even when
is an exact replica of the actual echo no echo is returned to A , only transmissiondelay is present.
y e t the path L2 is available f o r transmitting Bs speech t o A Circuit echo plays a more complex role in circuit quality. As
even during interruptions. mentioned in the introduction, echo returning with negligible
In Section IV, wewill explore this idea in some detail. In delay(i.e.,less than the inverse circuit bandwidth), at levels
particular we will describe an adaptive canceller which is ca- comparable to thosenormally receivedby the listener, are
pable of synthesizing a filter to match the transfer character- called sidetone and are essential to prevent the circuit from
istics of the echo path on which it happens to be operating. soundingdead. Echoes withlongdelays (e.g., greater than
Further, it is capable of tracking variations in the echo path 35 ms) produce annoyance for the talker which generally in-
during the course of a conversation. The descriptions of two creases with increasing echodelay and echolevel.?
such adaptive devices were given in 1966 in two simultaneous Another form of echo degradation occurs when the talkers
publications; one by Sondhi and Presti [6], and the other by speech reaches the listener with two, or more, different delay
Becker andRudin [ 71. These deviceswere built along the times. This listener echo is generally ignored since proper con-
lines suggested by J. L. Kelly and B. F. Logan. The first full trol of talker echo usually controls listener echo as well. How-
exposition of the principles of adaptive echo cancellation was ever, if the listener echo is due to local reflections (e.g., within
given by Sondhi [8]. Since then much work has been done at a central office), with delays less than a few milliseconds, the
Bell Laboratories [ 91 -[ 121, as well as by investigators at many resulting spectral distortion gives the received speech a hollow-
other laboratories, notably at the Australian Post Office[ 131- nesi or coloration. Listener echo of this type is not controlled
[ 151, at COMSAT Laboratories [ 161-[ 181, and at Nippon by conventional echo suppressors but the echocanceller is po-
Electric Company [ 19I , [ 201 . tentially a control technique.
The equations governing the adaptive echo canceller arise in In the following two sections, degradation produced by echo
various applications of system identification and have many and delay, separately and in combination, is examined in more
interesting mathematical properties. We will devote part of detail.
Section IV to these mathematical properties, in particular to
some of the recent work on estimating the rate at which the A . Effects ofTransmission Delay with No Echo
adaptation process converges. Pure transmission delay cannot be detected by one talker.
Finally, in Section IV-Ewewill briefly consider Koll and Rather, its effects areseen in the overall conversational dy-
Weinsteins recentproposal [ 211 to use echo cancellers for namics. An easy way to demonstrate one form of this effect is
data transmission. During the past decade telephonelines have alternate counting by talkers A and B. The rate of the count
been increasingly used for this purpose, and this use is likely to is controlled by the reaction time of each participant plus the
grow in thefuture. At present, high-speed full duplex (Le., total two-way transmission delay.
simultaneoustwo-way)datatransmission is done on wholly This effect occurs in a more natural, and more serious, way
four-wire circuits. A clever use of an adaptive echo canceller when one talker attempts to interrupt the other. If A is talk-
to cancel the echoes generated by a hybrid located at the data
terminal offers the intriguing possibility of using the ordinary Echo delay is equal t o the sum of transmission delays in both direc-
dialed-uplines of the telephone network for high-speed full tions. However, without an auxiliary transmission path, there is n o way
duplexdata transmission. This could be quite desirable to t o tell how delay is divided between directions. In fact, artificial delay
experiments(those not using actual satellitecircuits) are generally
customers whose data transmission needs are not large enough carried out with all delay lumped into one transmission direction for
to warrant leasing specialfour-wire circuits. simplicity and economy.
SONDHI AND BERKLEY: SILENCINGECHOES ON TELEPHONE 95 1
with split suppressors but the general principles apply equally r------------1
to both types.
An operating echo suppressor must accomplish one major
goal-removal of perceptible echo with minimum degradation I I I
of desired speech. The desired speech may be from talker A
to B (See Fig. 1) from B to A or simultaneous transmission in
both directions. As in Fig. 3, we w li discuss the operation of
the split echo suppressor at B's end with the understanding
that similar operations also occur at A's end of the transmis-
sion system. Interactions between the echo suppressors at the
two ends will not be considered, except to mention that gross
imbalance in end conditions (e.g., talker volume andloop
losses) can result in difficulties during simultaneous two-way
speech (double talking). Very early suppressors sometimes al- a
I
-
S
lowed lockout conditions wherein neither end could hear the I SUPPRESSION
other for significant periods of time. This rarely occurs with
LOSS
L - - - - - - - - - -J
I
modem properly adjusted systems. Fig. 5. Logical structure of a modem split echo suppressor.
'dBmO is the signal power in dB relative t o 1 m W at a point in the .' Centerclipping Echo Suppressor
designated as a zero
circuit transmission
level point (OTLP). Although mena conventional suppressor, as described above, is
the transmission level may be widely different at different points in the
network, t o avoidconfusion all dB values inthis paper are givenat a applied to terrestrial circuits with delays U P to 100 ms, the
OTLP. echo the
effects
andof echo suppressor action itself are gen-
SONDHI AND BERKLEY: SILENCINGECHOES ON TELEPHONE 953
OUT , r----l
I
I I
t t I
0)
Fig. 6 . (a) Instantaneousresponse of one form of centerclipper.
(b) Effect of center-clipping on a speech signal. t
I
erally subjectively undetectable. However, even with the best
modem designs, the combined effects of echo and the echo
suppressor are audible t o customers when round-trip satellite
delays are encountered. Fig. 8 . CCES using n bandpass Nters (BP), center clippers (CC) and
A number of experiments, both laboratory and field tests peak detectors.
using actual satellites or simulated delays, have been carried
out. Depending on the exact methodology and conditions of
echo suppressors and, on the other.side, the remarkable oper-
the experiments considerable variation in resultshavebeen
ating characteristics of the echo canceller.
obtained. Percent difficulty at 600-ms round trip has ranged
One unique design, called the center-clipping echo suppres-
from about 33percent (interviews of actual users of Early-Bird
sor (CCES),was proposed in1970[SI. Discussion ofthis
satellite circuits [ 11) to as much as 58 percent in laboratory
method of echo control illustrates some of the flexibility avail-
experiments [23]. At 300-ms delay, comparable figures ranged
able to the designer faced with the speech degradation prob-
from 17 percent for real circuits with simulated delay to 42
lems discussed above.
percent in the laboratory.
The CCES takes two partially independent routes towards
The most general conclusions which can be drawn from
reducing the effects of echo suppression on speech.
these results are the following:
Rather than switching the speech path on and off (or, at
1) the best conventional echo suppressors produce signifi- least, inserting large losses) a center clipper affects only por-
cant improvements of transmission quality for all delay tions of a signal which instantaneously fall below a threshold
conditions where notable degradation exists without such level p . A signalbelow this level is reduced t o zero while
suppressors; values above p are passed unchanged, as shown in Fig. 6(a).
2) when delays of 600 ms or longer are encountered the re- Fig. 6(b) shows an example of a center-clipped speech wave.
sulting circuit, with conventional echo suppressor, is A simplifieddiagram of a CCES is shown in Fig.7. The
significantly worse than a comparable circuit with pure threshold is set so as t o equal or exceed the return echo peak
delay and no source of echo. amplitude. If A alone talks the echo at point b is completely
Theseresultssuggest a need for improved methods of echo removed. However, if B breaks in, signals exceeding threshold
control, especially on long-delay circuits. The above tests also p will be heard by A. This is a unique feature of the CCES; no
indicated that modemconventional echo suppessors did a very double-talking decision is required sincebreak-in speech is
good job removing echo; the major cause of degradation on automatically transmitted, albeit somewhat distorted by the
long-delay circuits appears, in interview reports, to be cutting, operating center clipper. This distortion, although it is par-
fading, and other categories primarily indicating speech tially masked during double talking, leads t o another aspect of
mutilation effects. the CCES design, division of the spectrum into subbands.
Thus motivated, numerous ideas for improved echo control Previous to the CCES invention several proposals had been
have appeared in the last decade. Few of these methods have made for comb-fiiter control of echo [30]-[32]. The CCES
beenextensivelyinvestigated primarily because theymust also takes advantage of the reduced spectral requirements for
complete, on the one side, with the simplicity and (especially speech during certain periods (e&, whenprimarily one for-
in the most recent digital systems) low cost of conventional mant is excited)-allowing break in with no distortion-and, in
954 PROCEEDINGS OF THE IEEE, VOL. 68, NO. 8, AUGUST 1980
IMPULSE
APPLIED
AT 0
- -<
< 15 rnsec
5 rn?.ec-
-
Fig. 9 . A typical echo path impulse response observed at point b in Fig.
4 when an impulse is applied at point a.
Fig. 10. A transversal Nter. The boxes represent delays, each of dura-
addition, evenwhen significant clipping is required, removes tion A.
most of the distortion energy. (The only harmonic distortion
produced by a symmetrical center clipper are odd harmonics concerned, although of course it produces the long delay with
so thesecomponents are completely removed bybandpass which A hears the echo.)
Titers whose lower and upper cutoff frequencies differ by less One convenient type of filter which can be made to match
than a factor of three.) Narrow-band filters also minimize the any such echo path by a suitable choice of parameters, is a
effects of intermodulation distortion. A full CCES configura- transversal fiter (i.e., a tapped delay line). As shown in Fig.
tion is shown in Fig. 8. As expected, even the banded con- 10, the output y ( t ) of such a filter is related to its input x ( t )
figuration withouttheoutput, distortion reducing, fiitering by the relation
shows significant improvement over baseband operation.
Optimization of filtering parameters has not been studied N-1
since the rapidly decreasing costs have made the echocanceller y(t)= gnx:(t- nA)
n =o
seem apreferable solution. However,even if theecho can-
celler is eventually fully accepted, it may be that a simplified where g,s are the tap weights. If x ( t ) is band limited to the
CCESwillbeused to remove residual echo in the echo can- frequency range ( I f I < W), and A is chosen to be the corre-
celler systems discussed below [ 161, [ 171. spondingNyquist interval ( A = (2W)-) we obtainaband-
limited filter for which the gns are Nyquist rate samples of the
lV. ECHO CANCELLATION impulse response. Thus N A must be chosen larger than the
The methods of echo control that we have discussed so far longestechopath that needs to be accommodated.Speech
have one featurein common: they tamper with the pathcarry- signals onthetelephonenetwork have little energyabove
ing the echo and therefore, under adverse conditions, multilate about 3600 Hz, so A is conservatively chosen as 125 ps. Thus
the interrupters speech. This is avoided in themethod of 120 taps are required for the typical duration of 15 ms men-
echo control now known as echo cancellation to distinguish it tioned above.
from conventional echosuppression. All echo cancellers that have been built so f a r haveused
The basic idea of echo cancellation is, as mentioned in the transversal filters (with as many as 256 taps) for generating the
introduction, to synthesize a replica of the echo and subtract synthetic echo. However, as we will discuss in Section IV-DS,
it from the returned signal. We will assume that the echo path it is quite feasible, and in some situations advantageous, to use
is linear, and therefore completely specified by its impulse re- other types of adjustable filters.
sponse (or, equivalently its transfer function). All echo can- The simplest way to use such a filter would be as follows:
cellers designed so far make this assumption and accept non- 1) after the call has been established apply a test pulse at point
linearity of the path asan uncorrectable perturbation. (The a; 2) measure the responseat point b; 3)set up the tapweights
only suggestion that has been made so far for handling non- of the transversal filter proportional to the samples of the mea-
linear echopaths [ 101 is to consider the general Volterra suredimpulseresponse; 4) generate thesyntheticechoby
kernel expansion of the echo path. The feasibility and advan- fiitering As speech through this fiiter. Early proposals [33],
tages of this approach have not yet been demonstrated.) [ 341 for echo cancellation did justthis.
Let us refer again to Fig. 4 which illustrates the cancellation Forsuch an open-loopprocedure to work in practice it
of echoes generated by hybrid B. The synthetic echo is gen- would seem necessary to make frequent measurements of the
erated by passing As speech through a filter that matches the echo path during a conversation. This is because a single mea-
transfer function of theechopath.Theechopath is, of surement is not reliable and, more importantly, because the
course, qulte variable dependinguponthedistance tothe echo path is not completely constant for the duration of a
hybrid,the characteristics of the two-wire end circuit, etc. call.Besides obvious step changes such as connection or dis-
However, if an impulse is applied to the point ain Fig. 4, then connection of extensionphonesduringaconversationand
the response at point b will have the generic form shown in transfer of calls via key telephones or private exchanges, there
Fig. 9. There is an initial interval called flat delay for which are slow changes in gain and other fluctuations of the echo
there is no response at b. This flat delay is the round-trip de- path. The transmission of test pulses required for such adjust-
lay in the four-wire circuit between the canceller and hybrid ments is quite undesirable.
and is typically less than10 ms. The ringing afterthe flat These difficulties are avoided in the self-adaptive echo can-
delay can last a further 5 ms or so: (It is important to note celler. The speech itself is used in place of test pulses and a
that the satellite portion of the circuit is to the leftof point a. control loop continuously adapts the transversal filter to take
It is not included in the echo path with which the canceller is care of fluctuations in the echo path.
ential equations
p-7 I I
In vector notation
d &,
-
dt
=KF(e(t))x(tn
- A), n = 0 , l , - - * , N -1 . (4)
I +-- l I
dhh
4 '- 'I
EPATH
CHO\
I
-
dt
=KF(e)x
h
where hh is a column vector with components h , ( t ) ,and x is a
column vector with components x , = x ( t - nA). (For brevity
we will generally not explicitly show the dependence of quan-
tities on t . )
To heuristically justify this equation as a reasonable adapta-
Fig. 11. Block diagram of an echo canceller. As in Fig. 10, D denotes a tionalgorithm,considerafunction C ( e ) such that F ( e ) =
delay of duration A. dC/de. Then from (2) and (3), iifollows that the gradient of
C with respect to the tap weights h is
carry over to digital realizations although there are some signif- vi c = - -
dC X = - F ( e ) x.
icant differences which we will point out later. de
A block diagram of the adaptive canceller is shown in Fig.
1 1 . The signal x ( t ) is thespeech signal of speaker A. The Thus dhh/dt in ( 5 ) is a vector pointed in the direction in which
signal y ( t ) is the signal at point b of Fig. 4 and is the sum of C ( e ) decreases most rapidly. If C(e) is anyreasonablecost
echo z ( t ) of A's speech and a component n ( t ) . In the case function (e.g., e', I e I, -
* ) then decreasing C decreases I e 1.
when A alone is talking, n ( t ) is just low-level circuit noise. The proof that I e I approaches zero underideal conditions is
During periods of double talking' n ( t ) w i be predominantly
l as follows: let n ( t ) = 0 and let the echo path be exactly repre-
B's speech, which may be large compared to the echo. sentable as a transversal fiiter with fiied tap weights h . Then
The transversal filter generates an estimate z^(t)of the echo, in vector notation
given by
z ( t ) = h'x (7)
N-1 where ' indicates matrix transposition. Likewise
z^(t)= h, ( t )x ( t - nA).
n=o z^(t)= $'x (8)
Thus in these ideal conditions
(As mentioned earlier the tap delay A is typically 125 ps and
the number of taps N may be as large as 256.) e ( t ) = z ( t ) - z^(t)
The control loopuses the error
= (h - $)' X
A '
=rx (9)
where we have defied the misalignment vector r as
to continuously improve the estimate ?(t).
h
Ideally the system should drive itself to the conditione ( t ) = r = h - h. (10)
n ( t ) (not necessarily e ( t ) = 0 because n ( t ) may contain B's
speech which must be left as undistorted as possible). Such Undertheassumption that h is independent of t , dhh/dt =
ideal echocancellation is possible withthissystemonly if - ( d r / d t ) ,so that ( 5 ) becomes
n ( t ) = 0 and z ( t ) is exactly representable as the output of a dr
fixed-tap transversal filter with x ( t ) as its input. Under these - = -KF(r'x) x.
dt
conditions we will show in thenextsectionthattheecho
powerdoeseventually go to zero. Convergence of echo to If we premultiply both sides by 2r' and note that 2r'(dr/dt) =
zero, however, is not an adequate criterion of performance for d / d t (r'r) we get
a system of this sort. What is needed is the convergence of the
impulse response to the response of the echo path. This is a d
- (r'r) = - 2 ~ e ~ ( e ) . (1 2)
much stronger requirement which cannot be satisfied without dt
demanding certain properties of the input signal x ( t ) . We will
consider the convergence of the impulse response, the rate of Suppose nowthat F ( e ) has the same sign as e and that F ( e ) is
convergence as well as performance under nonideal conditions zero only if e is zero. Then (12) shows that r'r is nonincreas-
in Section IV-C. ing. It is strictly decreasingwhenever e # 0, i.e., whenever
there is an uncancelled echo. To show that e eventually goes
B. The Adaptation Algorithm to zero we integrate both sides of ( 1 2 ) between 0 and some
time 7. Thus
Observing that the input of the nth integrator in Fig. 11 is
d & / d t , it is clear that the tap weights satisfy the set of differ-
-
and < rrlo. Thus as 7 -+ the integrand on the right-hand
side mustapproach zero.6 Since the integrand is zeroonly
stay perpendicular to x(t) without decreasing in length at all.
This is more or less what happens when the loop gain K is
when e = 0 this proves that the uncancelled echo eventually made large. As shown in [35] when K is large, the vector r
goes to zero. rapidly becomes perpendicular to x(t) thus making e(t) small.
It is important to note that even under these idealized con- As long as the level of x(t) stays large the control loop keepsr
ditions the above argument does nor prove that the misalign- almost perpendicular to x thus keeping e(t) small. Hence from
ment vector r goes to zero. One simple situation in which r (1 1) becomes small even though r may be quite large.
cannot possibly go to zero is if the vector x(t) stays within a This situation is clearly quite undesirable, because slight de-
subspace of dimension less than N. Then since dr/dt is in the partures from the idealized assumptions will lead to significant
direction of x there is no adjustment of r in adirection perpen- errors. Thus even though the objective is to control echo the
dicular to x. formulation of theproblemshould be as an identification
Thus evenwhen the uncancelled echo becomes identically problem, with the emphasis on the rapid convergence of the
zero, equation (11) onlyguarantees that r stops changing. misalignment vectorr.
Bur it could stall a t any magnitude and orientation. It can be The classical method of estimating convergence rate for
forced to go to zero only by imposing conditionson the input equations such as (12) is the method of averaging, f i t pro-
process x. One necessary and sufficient condition is the mix- posed hi the 1940s [36], and often rediscovered since then.
ing condition((21) discussed in Section IV-Cl). Loosely The basic argument for (1 2) is as follows: suppose x is a sta-
speaking, this condition guarantees that the vector x does not tionaryrandom vector, and let us taketheexpectation,or
stay confined to a subspaceof dimension <N for toolong. ensemble average, of both sides of (12). If K is small enough
If n ( t ) is not zero, then (9) changes to it may be assumed that r on the right-hand side is statistically
independent of x . With this assumption the expectation can
e(t) = ri: +n(t) (14)
be calculated for avariety of different functions F and random
where n ( t ) is assumed to include the uncancellable echo if processes x. For example, let the components of x be zero
the echo path is not exactly representable by an N-tap trans- mean Gaussianprocesses and let (i) F ( x ) = x or (ii) F ( x ) =
versal fiiter. Proceeding as above we get sgn ( x ) (i.e., F ( x ) = 1 if x 2 0, and - 1 otherwise.) Then if
( ) denotes ensembleaverage we have [ 81,
dr
- = -KF[rx + n] x (15)
dt
and
in case (i) and
d
- rr = -KF(e)
2 (e - n ( t ) ) (16)
dt
in place of (1 1) and (1 2). The behavior of these equations is
much more complicated that that of (1 1)and (12) and will be +
in case (ii). Here is the correlation matrix of the xis. In
discussed in Section IV-C2.However, some qualitative ob- terms of the maximum and minimum eigenvalues A,, and
servations canbe made here. As long as n ( t ) is a low-level Amin of Cg it is easy to.derive upper and lower bounds. Thus
signal compared to e(t), theright-hand side of (15) is still with the usual notation,rr = Ir 1 we have
negative almost all the time except near zerocrossings of e(t).
In this situation the length of r decreases as before. Once e(t)
I r l r = o exp(-KAm,t)G Irl G Irlt=o exp(-KA,int) (19)
is.down to the level of n ( t ) , r fluctuates randomly. If x satis- for case (i) and
fies the mixing condition then we wil see in Section IV-C2
that r eventually converges to a sphere aroundthe origin and is
guaranteed to staywithinit.Theradius of that sphere de-
pends on the loopgain K, the noise level, etc. for case (ii).
Mostanalysesofconvergence rate, even themoreerudite
C. Convergence of Impuke Response ones [ 371 are based on the averaging principle. For the con-
We turn now to questions concerning theconvergence of the tinuous algorithms being discussed here, the method is valid
misalignment vector r to zero. As mentioned above this is the [38] only for a time interval of the order O(l/K). For the
more basic problem even though ostensibly the objective here time-discretized version (see (29)) it is valid only if successive
is the control of the echo z ( t ) . To appreciate the importance input vectorsx are statistically independent.Unfortunately,
of this assertion, let us first consider the ideal situation with for echo cancellation the input is a highly correlated speech
n ( t ) = 0 (i.e., z ( t ) = e(t)). signal. Therefore, the application of the method, particularly
1 ) Convergence under Noiseless Conditions: It is clear in this for large K, can lead to quite erroneous conclusions. For ex-
casethatifr=Othene(t)=O. Ontheotherhandife(t)=O ample, equations(1 9)and (20) imply that in thenoiseless situa-
we can only conclude that r is perpendicular to x(i.e., rx = 0). tion, increasing K always results in faster convergence. This
We can conclude nothing about the length of r. In fact for a conclusion is false.
given input process x(t), it is possible for the vector r ( t ) to Following Sondhi and Mitra 1351, rigorous bounds on com-
vergence rate valid for all values of K can be derived for inputs
that satisfy a mixing constraintand anaverage powercon-
Pathological situations where, for example, the integrand might a p
proach a series of zero-area spikes, can be ruled out by imposing weak straint. The mixing condition is satisfied by x@), if positive
conditions on the function F [ 8 1. numbers a and T exist such thatfor any time-invariant
SONDHI AND BERKLEY: SILENCINGECHOESONTELEPHONE 951
vector d ,
. rr+T
10 12
-
2 4 6 8 14 16 18 i
a KTL~
In addition, let us demand that the function F of Fig. 11 sat-
isfy a sector condition,i.e., Fig. 12. Upper bound for convergence rateas a functionof (normalized)
loop gain for various degrees of mixing. (See Section IV-C.)
a l e < eF(e) Q az e (23)
with some positive numbers a l and ( 1 2 . With these definitions
used t o bound the residualmisalignment due t o any pertur-
it can be shown that [351
bation which can be represented by (25). We mention two
IrlQ lrlt=oeb(T-r). (24) exampleswhich are of considerable importance to echo
cancellation.
Thus mixing inputs guarantee that the misalignment vector r 1)If, contrary tothe assumption so far, the echo path
converges exponentially to zero. impulse response h is not Tonstant, then the bound (27) a p
The dependence of the exponential decay constant b on the plies with U such that 1 h ( Q U . In this case the bound is
parameters a,a l , a z , K , T , L is summarized in Fig. 12. Here smallestwhen the convergence is fastest. This is reasonable,
the dimensionless quantity bT is plotted against the dimen- for one would expect the tracking ability of the adaptive loop
sionless quantity azKTLZ for various values of a third dimen- to be a maximum when its convergence rate is a maximum.
sionless quantity ala1 L z . For a given input and a given 2) If n ( t ) is not zero (due to thepresence of double talking,
function F , the only adjustable design parameter is K . Then due to inadequacy of the echo path model or due to circuit
Fig. 12 shows that as K increases the upper bound firstim- noise) let S be a bound on its r m s value, i.e.,
proves and then becomes worse and worse beyond an opti-
mum value of K , with bT becoming proportional to 1/K for
large K .
What is more important is that this behavior of the upper
bound qualitatively reflects the behavior of the actual conver-
gence process. This was shown in [ 3 5 1, for a very general class
of inputs. Also in [35], it was shown that under the assump- Then the bound (27) applies with U = a z K L S .
tions that is finite and Ix 1 2 I 2 0, one can derive a lower Note from Fig. 12, that bT is proportional t o K as K + 0,
bound on Ir I whose qualitative behavior is the same as that of and to 1/K as K + 00. Thus the bound obtained here becomes
the upperboundfor large K , for most initial conditions. independent of K for small K and increases as K Z for large K .
Finally, the existence of an optimum loop gain has been ob- This quantifies the intuitive notion that faster convergence can
served experimentally (see, e.g., [ 391 ). be achieved only at the expense of an increased sensitivity to
2 ) Convergence in the Presence of AdditivePerturbation: noise. (Of course, beyond the optimum value of K , increasing
The effects of many different kinds of departurefrom the K reduces the convergence rate and increases the sensitivity t o
ideal conditions assumed in the previous section can be esti- noise.)
mated by considering the modified equation At first glance the bound appears to be unduly conservative
for K 0. Intuitively, as K + 0 one might expect the residual
-f
I ) Discrete-TimeVersion of theAlgorithm: The continu- Thus, if the echo path does not change appreciably d h g the
ous algorithm of (5) can be converted to a discrete-time algo- double-talking interval, the echo stays cancelled.
rithm in a variety of ways. The simplest (and to our knowledge Althoughthere is roomforimprovement, the following
the only one tried so far) is to replace the time derivative by a simple double-talk detector [ 121 has been shown to be ade-
forward difference. Thus quate for the present purpose: if the signal z ( t ) > half the
A A largest magnitude of x ( t ) over the preceding Z seconds then
hm + 1 = hm + KTF(e, X, (29) declare double talk. The factor half is based on the assump-
where m is the sample number and T, is the sampling interval. tion that the return loss is at least 6 dB, and the time in-
(When the filter is a tapped delay line, T, is chosen equal t o terval Z is chosen to be the largestallowable length of im-
the tapdelay A, or some integral multiple thereof.) pulse response (including the flat delay).
There is one basic difference between the continuous and Instead of limiting the loop gain to one of two values (0
discrete algorithms. In the absence of noise, the continuous during double talking and K otherwise) Hoge [41 I proposes
algorithm never increases the length of the misalignment vector an algorithm to vary it continuously with the noise level. His
r. In contrast, thediscrete algorithm may increaser (and hence algorithm is based on thefollowing ideas: let e , = rk x, + n , ,
diverge) forsufficiently large K . Toshowthis, note that if where the noise samples nm are independent and identically
h is subtracted from both sides of (29) we get distributed (i.i.d.) and also independent of the input x . Then
with F ( . ) = (.) the averageof (3 1) over the noise ensemble
r,+, = r , - KT,F(e,)x,. (30) gives
Therefore with noise assumed to be zero we get
I r,+l I 2 = I rm I 2 - 2KTs e, F(e,1 + [KTF(e, I X , I 11.
(34)
(3 1) where R , is the average of I r, I and u2 is the noise variance.
Then the length of r decreases iff the sum of the last two terms Clearly R , + is minimized for K = K , where
is negative. This is true iff
XU)
*
I 1 I
& b-----& \ ECHO
PATH
r
I1
I 1
I1
- I1
I ,
+ 1 1
FOREGROUND
Fig. 13. A generalized echo canceller. The boxes w may be linear or non-
linear fdter networks. e ! -A+!
Fig. 15. An echo canceller withtwoecho path models. The back-
groundfilter is adaptedessentially as in Fii. 11. The coefficients
of the foreground tilter are set equal to those of the background
whenever the latteris judged to be better.
A STATION B
Fig. 16. Illustrating the placement of an echo canceller for data trans-
mission.
I
wc I
I
VARIABLE
BULK
TAPS
DELAY VARIABLE
Fig. 17. Two variable transversal fdters separated by a bulk delay equal
M T A SYMBOLS I
I
I
I
4 FILTER O D U L o R b
I
II F ( t 1
I
t o the round-trip delay in the four-wire circuit. Such a configuration k I I
I
is necessary when a cancelleris located as in Fig. 16. I
I BASEBAND
I
i _ _ _ _ _ _SIGNAL
_ _ _ _ _ _ _ _J
criteria) is now formally identical to theproblem for the trans- TRANSMITTER
versal filter.
RCVR
Gitlin and Thompson [44] have proposed to use this idea y (t)
for echo cancellation with the configura>on shown in Fig. 14.
Fig. 18. Generation of data signals.
During double-talking intervals, 6 and b are frozen and 9 is
obtained recursively by means of (41). Note that this adapta-
tion algorithm does not stop automatically when speaker A is expensive than leasing atwo-wire line. However, cancellers
silent (i.e., when x is zero). This condition must also be de- could be quite useful for data transmission over circuits estab-
tected and the adaptationfrozen. lished in the usual manner bydialing through the switched
The main drawback of this proposal is that the estimated telephone network. For in this situation, full duplex operation
recursive filter of (41) is often unstable. However, in simula- without echo cancellers is possible only by making two calls-
tionexperimentswith q as large as 16,the filter [441 is oneforeach direction of transmission. And two calls cost
claimed to stay stable. If this is substantiated by moreex- twice as much as a single call.
haustive experiments then the proposal has the potential of Although basically the canceller in this application is similar
significantly reducing the number of adapted coefficients. to those discussedabove (transversal fdter with tap weights
The amount of savings that might be realized in practice wil updated along the gradient of some cost function of the error),
depend upon the ensemble of echo paths. there are several significant differences.
6) TwoEcho Path Models: In t,he algorithmsdescribed The first difference is in theplacement of the canceller.
above, the most recent estimate of h is used to generate the Unlike voice echo cancellers which are envisioned as part of
synthetic echo. Ochiai e t al. [201 propose an interesting varia- the telephone network, a data echo canceller would be located
tion in which two impulse response estimates are generated as at the dafa set as part of the equipment on the customers
shown in Fig. 15. One, called the background estimate, is premises. To understand why this is necessary, let us examine
adapted essentially as in (29). The other, called the fore- Fig. 16. Note that a hybrid is needed at each data set in order
ground estimate, is used to synthesize the echo which is sub- to separatethetransmitand receive directions. Suppose A
tracted from the signal returned to speaker A. The foreground transmits data; then the first echo received is that due to the
estimate is changed only if a decision mechanism indicates that mismatch at the local hybrid, followed by the ringing in the
thebackgroundestimate hasprovided a sufficiently better two wire circuit between As local hybrid and the hybrid at
cancellation than the foregroundestimatefora sufficiently As central office. For voice circuits, this ringing, which
longtime. In such case theforeground is set equal to the merges into the artificially provided sidetone, is quite harm-
background. less (see Section I). For data transmission, on the other hand,
The decision algorithm used by Ochiai e t al. is too compli- this early echo is just as damaging as the longer delayed echo
cated to describe here. Thereader is referred to[20]for from the mismatch and ringing at Bs hybrid. And it is clear
details. that the early echo would persist even if the four-wire portion
of the circuit were absent, or completely anechoic. The only
E. Echo Cancellation for Data Transmission way to deal with this echo is to place the cancellers at the
We close our account of echo cancellation by briefly men- data sets as shown.
tioning the recently proposed application of the idea to full The necessity of placing the canceller at the data set gives
duplex data transmission. rise to the second major difference between thevoice and data
For data transmission over a private line, echo cancellers do cancellers. The span of time over which echoes arrive is much
not at present appear to have much of an application. This is larger for the latter. This is because the longdelayed part of
because leasing a four-wire private line is only slightly more the echo in this case arrives after the round-trip delay in the
SONDHI
TELEPHONE
ANDON
BERKLEY:
ECHOES
SILENCING 96 1
TRANSMITTED
SYMBOLS
where Tis the intersymbol interval andg(t) is an appropriately
ak - FILTER
designed pulse shape. The passband signal F ( t ) actually trans-
mitted over the line is obtained by modulating a sinusoidal
carrier with f ( t ) . Thus
F ( t ) = f ( t )cos 2nfct + f ( t )sin 2sf,t (44)
RECEIVED
SYMBOLS where 7 is a linearly filtered version of fp). (For example, in
the case of upper sideband modulation f is the negative ofthe
Hilbert transform o f f ) . If fcT is rational (= n 1 Inz ), then [45 1
it is possible to write
F(t)= akghkmodn,(t - kT) (45)
k
(a)
TRANSMITTED
SYMBOLS
-
where gl, . , gn2 are derivable in a straightforward manner
ak from g.
Assuming a linear echo path, the signal y ( t ) in Fig. 18 con-
sists of the received signal plus an echo which is a linear trans-
formation of F ( t ) . A direct application of the echo cancella-
tionprocedure would be to generateasyntheticecho by
RECEIVED
SYMBOLS passing F ( t ) through a transversal filter, and update the filter
tap weights to minimize error. However, this would not take
I 'I I u -
SAMPLER advantage of the specific structureof the signal F ( t ) .
It was Mueller [46] who first pointed out how the structure
of the data signal could be usefully exploited.The idea is
based on the simple observation that f ( t ) , F ( t ) and hence the
TRANSMITTED
echo in y ( t ) are linear combinations of the data symbols 'Ik.
SYMBOLS Therefore, the synthetic echo can be generated directly as a
linear combination of the a i s . The cancellation itself may be
done in baseband or passband.
Three different canceller configurations based on this idea
have been proposed. These are illustrated schematically inFig.
19 (a)-(c), the data set and canceller at station A being shown
in each case.
RECEIVER
Configuration (a) (Mueller [461) is a symbol rate canceller.
NYOUIST Assuming cancellation at baseband, thesynthetic echo is
RATE
SAMPLER generated by the transversal filter T. This syntheticecho is
(c) subtracted from the receivedsignal at baseband and the re-
Fig. 19. Three possible configurations for a data-driven echo canceller. ceiver output sampled at symbol rate is used to update the
In each case the synthetic echo is generated from the data symbols. transversal filter with agradient algorithm. This configura-
(a) Symbol rate cancellation in baseband. (b) Cancellation combined tion requires synchronization of received and transmitted
with adaptive equalization. (c) Nyquist rate cancellation.
sequences.
Configuration (b) (Falconer et nl. [471) is similar to configu-
four-wire circuit; i.e., as much as 600 ms for a satellite circuit. ration (a), except that an adaptive equalizer is inserted before
This problem is dealt with by splitting the transversal filter the point where cancellation occurs. The tap weights of the
into two adjustable transversal filters separated by a bulk delay canceller andequalizerarejointlyupdated by agradient
as shown in Fig. 17. The bulk delay is measured by an inter- algorithm.
rogating pulse during the start-up procedure. In configuration(c) (Weinstein [45]),the cancellation is
Finally, there is a considerable difference in the properties done in the passband at a sampling rate which is greater than
of the signalsused inthetwo applications.Thestatistical Nyquist rate (twice the highest frequency) of the passband.
properties of speech signals are very hard to quantifyex- Although this is much higher than symbol rate this configura-
cept in very broad terms, e.g, their bandwidth, the range of tion has the advantage that synchronizationbetweentrans-
fundamental frequencies that might be encountered, etc. Data mitted and received sequences is not necessary. If the sampling
signals, on the other hand, have much simpler statistical prop- rate is chosen to be an integral multiple I of the symbol rate,
erties and are much more structured. Thus sequences of data then each symbol multiplies I consecutive taps of the trans-
symbols ak (selected from some alphabet, e.g., f l for binary versal filter, once eachduring a symbol interval.
sequences) may quite accurately be assumed to be sequences A discussion of the details of these configurations is beyond
of independent and identically distributed variables. Further, the scope of the present paper.
the transmitted analog signal corresponding to such a sequence
is generated in a precisely defiied manner. To take the most V. CONCLUDINGREMARKS
usual c&e of pulse amplitude modulation, the baseband signal A . Implementation
corresponding to thesequence ak is
All adaptive echo cancellers that have been proposed in the
f(t) = a k g ( t - kT) (43) literature use one form or another of the basic gradient search
k algorithm. And all cancellers that havebeen actuallycon-
962 IEEE,OF THE
PROCEEDINGS VOL. 68, NO. 8 , AUGUST 1980
structed have used the transversal filter with variable gain coef- ties or phase roll, and because double-talk detection would be
ficients as the adaptive structure. more accurate in the absence of flat delay. And of c o m e with
When echo cancellers were first proposed in the mid-1 960s,a the echo cancelled at itspoint of origin, there wouldno longer
filter of this type with say 50 taps, appeared to be prohibitively be any need for echo suppressors.
expensive. The rapid advances in digital circuitry during the On the other hand,if echo cancellers are intended forsatellite
past decadehave dramatically changed thesituation. As we circuits alone, then locating them at thesatellite ground stations
have noted above,several cancellers (e.g.,[ 121 ,[ 161 ,[ 181)have wouldrequire the least number.Thenumber of cancellers
been constructed with hundreds of taps. Recently Duttweiler requiredbecomes progressivelylarger as one locates them
and Chen [48] at Bell Laboratories havedeveloped a single- farther away from the ground stations. The number would be
channelintegrated circuit echo canceller chip. Eachsuch many orders of magnitude larger if the cancellers were to be
VLSI chip has approximately 35 000 gates and is based on the located at the hybrids.
128-tap design described earlier by Duttweiler [ 121. It is not clear which placement will ultimately yield the best
It is to be expected that as digital circuits become even compromise. Here too the picture is changing rapidly. Only
cheaper and faster, more sophisticated systems will be imple- a few years ago the ground station appeared to be the only
mented. We have already mentioned the possibility (Section feasible location. With cancellers available on inexpensive IC
IV-D5) ofusing other, possibly nonlinear, filter structures, chips, it may well become feasible in the future to install a
although it is not yet clear what structure might give the most canceller at each hybrid. It may also befeasible to usecan-
economical design, or most accurate cancellation. As to cellers for eliminating listener echo in digital local offices.
alternative algorithms, a variety of proposals for systemidenti- Further ahead, data echo cancellers similarto the onesdiscussed
fication (see e.g., [49]) have been made in contexts other than in Section IV-Emaybeused at the telephone set in digital
echo cancellation. Forexample, as is well known,the cost subscriber loops [541.
function of (36) with c ( e k ) = e: can be minimized byinverting
a single N X N covariance matrix. The same result can also be C. Field Trials
achieved recursively by fast updating of gain matrices [ 501 . We close our account of echo cancellers by briefly mentioning
It is also possible to estimate the echo path in the frequency some recent studies aimed at evaluating cancellers under nor-
domain. Such procedures [ 5 1 I might become attractive when mal operating conditionsin a telephone network.
inexpensive hardware for fast Fourier transformation becomes A comprehensive seriesof trials of this type was conduc-
readily available. tedduring1973and1974 byCOMSAT Laboratories [ 171
Even for the simple transversal filter adapted by a gradient in collaborationwith seven membercountries (Australia,
algorithm, there is a variety of possible implementations. 1) Brazil, France, Japan, West Germany, U.K., and the U.S.) of
Conventional analogcircuits (e.g, [6]); 2)circuits using charge- the InternationalTelecommunications Satellite Organization
coupled devices (CCD) [521; and 3) digital circuits, which in (INTELSAT). Some of the trials used cancellers designed by
turn maybe implementedwith discrete components [ 121 , COMSAT [ 161 and the others used cancellers with two echo
[ 161, [ 181 or with integrated circuits before, and may be de- pathmodels (see Section IV-D6)designedby the Nippon
signed to operate on a single channel or with circuitry shared Electric Company [ 201. In each case a single wide-band center
with many channels. clipper wasused to eliminate residual echo. The objective
The current trend, of course, is toward digital implementa- evaluationconsisted of monitoringthetimecourse of the
tions. However, the others, specially the CCD implementations estimatedimpulseresponse.The subjective evaluation was
merit further investigation. Oneof the problemswith CCD done by means of interviews withthe customers.
technology is the difficulty of manufacturing multipliers which It is not easy to summarize the findings of these trials be-
are linear in both inputs. It has been shown [ 1 11 , [ 53 1 that cause of the wide variety of circuit conditions (signal, noise,
the linearity requirement can be considerably relaxed. Other andecho levels) and languagesinvolved.Nevertheless the
problems such as charge transfer efficiency, variations in gain general conclusion [ 171was that the echo cancellers . . .
and dc offset, etc., [52], need further attention. On the posi- significantly improvetheperformance of single-hop satellite
tive side, CCD implementation eliminates the need for analog/ circuits when echo is the principal cause of the difficulty. A
digital and digital/analog conversion. This consideration alone large percentage of the terrestrial portions of the circuits used
might make itthe least expensive alternative in situations inthestudy had highloss and/or highnoiselevel. Under-
where echo cancellation is required on single analog channels. standably, on such circuits the echo cancellersyielded only
marginal improvement.
B. Placement More recently (1976 and 19771, Bell Laboratories and Long
The impulse responseto be simulated is clearly of least dura- Lines conducted a series of trials [ 241 to evaluate performance
tion right at the hybrid which gives rise to the echo. (Farther of domestic satellite circuits equipped with echo cancellers or
away the flat delay increases and the active portion of the im- one of two types of echo suppressors-the Western Electric 4A
pulse response spreads by convolution with the responses of or the new digital echo suppressor terminal (EST)designed for
intervening transformers, filters, etc.) From the point of view electronic switching offices. The cancellers used in these trials
of simplicity of design, therefore, the ideal location for a can- were supplied by COMSAT Laboratories. They were adapted
celler is close to the hybrid. Such placement would improve for insertion intothe Bell Systemtelephonenetwork,and
performance because the echo path would have no nonlineari- their double talk detectors were replaced by ones designed at
Bell Laboratories.
The trials were conducted on calls between Pittsburgh and
9Another consequence of suchplacement is worthnoting.Atthe
hybrid, instead of adaptively synthesizing an echo path, it is possible to Atlanta andbetweenPittsburghand San Francisco. As con-
adapt the balancing network. When the network matches the impedance trols, the study included terrestrial circuits (with simple loss
of the two-wirecircuit,thehybridbecomesanechoic. It is not clear on the Pittsburgh-Atlanta circuits, and with 4A echosup-
whether this procedure would be an improvement over the canceller,
but it is an alternative worth some study. pressors on the longer Pittsburgh-San Francisco circuits). The
SONDHI AND BERKLEY: SILENCING ECHOES ON TELEPHONE 963
trial procedure was quite similar to the one used by COMSAT Proc. Inst. Elec. Eng.,vol. 118, pp.401-408,1971.
[ 2 4 ] G.K. Helder and P.C. Lopiparo, Improving transmissionon
Laboratories [ 171 in the international study. The conclusion domestic satellite circuits, Bell Lab. Rec., vol. 55, pp. 202-207,
was that one-hop satellite circuits equipped with cancellers Sept. 1977.
performed significantly better than those equipped with echo [ 2 5 ] E. W. Holman and V. P. Suhocki, A new echo suppressor,
Bell Lab. Rec.,vol. 44, pp. 139-142, 1966.
suppressors, and almost as well as the terrestrial circuits. An 1261 J . E. Unrue,Echosuppressor design considerations, IEEE
interesting related finding was that half-hop circuits (i.e., Trans. Commun.,vol. COM-16, pp. 616-624,1968.
I271 M. B. Gardner and J. R. Nelson,Combating echo in speech
circuits in which one direction is via satellite and the other via circuitswithlong delay, I. Acoust. SOC.Amer. vol. 35, pp.
a terrestrial circuit) equipped with the EST also performed as 1762-1772,1963.
wellas the terrestrial circuits.Presumably theround-trip [ 2 8 ] D.L. Richards and J. Hutter, Echosuppressors for telephone
connections having longpropagationtimes, Roc. Inst. Elec.
delayof 300 ms on such a circuit is short enough for echo Eng.,vol. 116, pp. 955-964, 1969.
suppressors to work satisfactorily. [ 2 9 ] R. W.Hatch and A. E. Ruppel, New rules for echo suppressors
inthe DDDnetwork,Bell Lab. Rec., vol. 52, pp. 351-357,1974.
DEDICATION [ 3 0 ] C. C. Cutler, Echo suppressor using comb filters, U. S. Patent
3 175 051, 1965.
We dedicate this article to thememory of J. L. Kelly, Jr. [ 3 1 ] B. Peroni,Echosuppression and noise eliminationsystem for
telephone circuits, U. S. Patent 3 567 873, 1971.
[ 3 2 ] D. A. Berkley and J. S. Courtney-Pratt, Telephonic transmission
REFERENCES usingcomplementary comb filters,U.S. Patent 3 622 714, 1971.
[ 3 3 ] J. L. Flanagan and D. W. Hagelbarger, Self-adjusting echo can-
R. G. Gould and G. K. Helder,Transmissiondelay and echo celler,U.S. Patent 3 535 473, 1970.
suppression, IEEE Spectrum, pp.47-54, Apr. 1970. [ 3 4 ] A. Miura, S. Kobayashi, R. Sato, and K. Nagata, Blockless echo
A. B. Clark and R. C. Mathes, Echo suppressors for long tele- suppressor,EZectron. Commun. Jap.,vol. 51-A,pp. 10-18,1968.
phone circuits, Roc. AIEE, vol. 44. pp. 481-490, Am. 1925. [ 351 M. M. Sondhi and D. Mitra, New results on the performance of a
P. T. Brady, A statistical analysis of on-off patterns in sixteen well-known class of adaptivefilters, Proc. IEEE, vol. 6 4 , pp.
conversations, Bell Syst. Tech. J.,vol. 47, pp. 73-92, 1968. 1583-1597,1976.
S. B. Weinstein, Echo cancellation in the telephone network, [ 3 6 ] N.N. Bogoliubov and J. A. Mitropolski, Asymptotic Methods in
IEEE Commun. SOC.Mag., pp. 9-15, Jan. 1977. the Theory of Nonlinear Oscillations. New York: Gordon and
0.M. M. Mitchell and D. A. Berkley, A full-duplex echo sup- Breach, 1961.
pressor using center clipping, Bell Syst. Tech. J., vol. 4 0 , pp. [ 37 1 B. Widrow e t al., Stationary and nonstationary learning char-
1619-1630, 1971. acteristics of the LMS adaptive filter, Proc. IEEE, vol. 64, pp.
M. M. Sondhi and A. J. Presti, A self-adaptive echo canceller, 1151-1162,1976.
Bell Syst. Tech. J.,vol. 4 5 , pp. 1851-1854, 1966. [ 38 ] R. 2. Khasminskii, On stochastic processes defined by differential
F. K. Becker and H. R. Rudin, Application of automatic trans- equations with a small parameter, Theory Rob. Appl. (USSR),
versal filters t o the problem of echo suppression, Bell Syst. Vol. 1 1 , Pp. 21 1-228,1966.
Tech. J.,vol. 45, pp. 1847-1850, 1966. [ 391 B. J. Miller, A general method for computing system parameters
M. M. Sondhi, An adaptive echo canceller, Bell Syst. Tech. J., withanapplication to adaptive control, presented at AIEE
Vol. 46, PP. 4 8 7 4 1 1 , 1 9 6 7 . Winter General Meet. (New York), Jan. 1962.
J. R. Rosenberger and E. J. Thomas, Performance of an adaptive [ 4 0 ] A. Weiss and D. Mitra, Digital adaptive filters: Conditionsfor
echo canceller operating in a noisy, linear, time-invariant environ- convergence, rates of convergence, effectsof noise and errors
ment,Bell Syst. Tech. J., vol. 50, pp. 785-813, 1971. arising from the implementation, IEEE Trans. Inform. Theory
E. J. Thomas, An adaptive echo canceller in a nonideal environ- Val. IT-25, PP. 637-652,1979.
ment(nonlinear or timevariant), BelZ Syst. Tech. J., vol. 5 0 , [ 4 1 ] H. Hoge, Analysis of an adaptive echo canceller with optimized
PP. 2779-2795,1971. gradient gain,Siemens Res. Dev. Rep.,vol.4, pp. 127-131, 1975.
[ 11 1 D. Mitra and M. M. Sondhi, Adaptive fdte-ring with nonideal [ 4 2 ] K. Steiglitz, On the simultaneous estimation of poles and zeros
multipliers:Application to echo cancellation, in h o c . 1975 in speech analysis,ZEEE Trans. Acoust. Speech, SignalProcessing,
In?. Conf. Communications, pp. 30-11-30-15, 1975. vol. ASSP-25, pp. 229-234, June 1977.
[12] D. L. Duttweiler, A twelve-channel digital echo canceler, [431 R. E. Kalman, Design of aself-optimizing control system,
IEEE Trans. Commun., vol. COM-26, pp. 647-653, 1978. Trans, ASME, vol. 8 0 , pp. 468-478, Feb. 1958.
[ 131 A. W. Thies and R. B. Zmood, New ways of echo suppression, [ 4 4 ] R. D. Gitlin and J. S. Thompson, A new structure for adaptive
Austral. Telecommun. Rev., vol. 1, pp. 14-19, 1967. digital echo cancellation,inProc. Nat. Telecommunication Confi,
[ 141 N. Demytko and L. K. Mackechnie, A high speed digital adaptive (Dallas, TX) PP. 8.2, 1-7, 1976.
echo canceller, Austral. Telecommun. Rev., vol. 7 , pp. 20-28, [ 4 5 ] S. B. Weinstein, A passband data-driven echo canceller for full-
1973. duplex transmission on two-wire circuits, IEEE Trans. Commun.,
[ 151 N. Demytkoand K. S. English, Echocancellationon time- V O ~ .COM-25, pp. 654-666,1977.
variant circuits,Roc. IEEE, vol. 65, pp. 444-453, Mar. 1977. [ 4 6 ] K. H. Mueller, A new digital echo canceller for two-wire full-
[16] S. J. Campanella, H. G. Suyderhoud, and M. Onufry, Analysis duplex data transmission, IEEE Trans. Commun., vol. COM-24,
of an adaptive impulse response echo canceller, COMSAT Tech. PP. 956-962,1976.
Rev.,vol. 2 , pp. 1-36, Spring 1972. [ 4 7 ] D.D. Falconer, K.H. Mueller, and S. B. Weinstein, Echo can-
[17] H. G. Suyderhoud, S. J. Campanella, and M. Onufry, Results cellation technique for full-duplex data transmission on two wire
and analysis of a worldwide echo canceller field trial, COMSAT lines, in Con5 R e c Not. Telecommunication Con5 (Dallas TX),
Tech. Rev., vol. 5 , pp. 253-274, Fall 1975. PP. 8.3 1-7, 1976.
181 0. A. Horna,Echo canceller with adaptive transversal fdter [481 D. L. Duttweiler and Y. S. Chen, A single chip VLSI echo can-
utilizing pseudo-logarithmic coding, COMSAT Tech. Rev., vol. celler,BellSyst. Tech. J.,vol. 59, pp. 149-160, 1980.
7 , PP. 393-428, Fall 1977. 1491 P. Eykhoff, System Identification. New York: Wdey, 1974.
19 ] Y Kato, S. Chiba, T. Ishiguro, Y. Sato, M. Tajima, T. Ogihara, 1501 L. Ljung, M. Morf, and D. Falconer, Fast calculation of gain
S. J. Campanella, H. G. Suyderhoud, and M. Onufry, A digital matrices for recursive estimation schemes, Znr. J. Contr., vol.
adaptive echo canceller, NEC Res. Dev., no. 31, pp. 32-41, 27, PP. 1-19,1978.
Oct. 1973. 151 1 J . B. Allen and L. R. Rabiner, Unbiased spectral estimation and
201 K. Ochiai, T. Araseki, and T. Ogihara, Echo canceler with two system identification using short-time spectral analysis methods,
echopath models, IEEE Trans. Commun., vol. COM-25, PP. ~. Bell Syst. Tech. J., vol. 5 8 , pp. 1743-1763, 1979.
589-595, June 1977. [ 521 B. K. Ahuja et al.,in Con5 Rec. Not. Telecommwucation Con5
(21 ] V. G. Koll and S. B. Weinstein, Simultaneoustwo-way data (Birmingham, AL), k c . 3-6, paper 30.3.
transmission over a two-wire line, IEEE Trans. Commun., vol. 1531 D. hiitra and M. M. Sondhi, Qualitative behavior of nonlinear
COM-21, pp. 143-147, Feb. 1973. differential equations describing adaptivefilters using nonideal
[22] R. R. Riesz and E. T. Klemmer, Subjective evaluatiEn of delay multipliers, IEEE Trans. Automat. Conrr., vol. AC-24, pp. 276-
andecho suppressors in telephone communications, Bell Syst. 283,1979.
Tech. J.,vol. 42, pp. 2919-2943, 1963. I541 P. T. Nielsen and M. W. Gram, A digital hybrid for two-wire
[23] G. Williams and L. S. Moye, Subjective evaluation of unsup- digital subscriber loops, in Con$ Rec. Nat. Telecommunication
pressed echo in simulated long-delay telephone communications, Conf., 1978, paper 21.2.