Echo Canceler with Two Echo Path Models by yjg19349


									IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-25, NO. 6 , JUNE 1 9 7 7                                                                  589

                        Echo Canceler with Two Echo Path Models

    Abstract-An adaptive echo canceler with two echo path models is        originate solely from its input signal. During double-talking,
proposed to overcome the false adaptation problem for double-talking.      however, the output signal contains not only the echo of the
The echo canceler possesses two separate echo path models (EPMs), one
                                                                           input signal, but thenear-end talker’s speech signal as well.
(background EPM) for adaptively identifying echo path transfer charac-
teristics and the other(foreground        EPM) forsynthesizing an echo     Under this condition, the EPM transfer characteristics may be
replica to cancel outecho.The         parameter values of theforeground    greatly disturbed unless adaptation is disabled prior to initia-
EPM  are refreshed by those of thebackground EPM, according to a           tion of double-talking [lo] .
transfer control logic, when the logic determines that the background          A simple andmost direct approachfor this would be to
EPM is giving a better approximation of echo path transfer characteris-
                                                                           use a double-talking detector and to enable or disable adapta-
tics than the foreground EPM. Completely digital hardware implemen-
tation is described. Using the hardware, it is shown that virtually com-   tion according to thedetectoroutput.          As yet no detector
plete             protectionis
                                    realizable                by the
                                                                   new     adequate  for      present
                                                                                            the      purpose has         been designed. For
method.                                                                    example, using the conventional method of comparing receive-
                                                                           in and send-in signal levels, it is well known that a considerable
                                                                           portion (sometimes a whole syllable) of the initial part of
                        1. INTRODUCTION
                                                                           break-in near-end talker’s speech is lost. If such a method were
                                                                           used for an echo canceler, adaptation would proceed for a con-
E     CHO in telecommunicationnetwork is inducedmostly
      by impedance mismatch at the hybrid coil connecting the
two wire and the four wire systems. This echo problemhas
                                                                           siderable amount of time in the presence of double-talking,
                                                                           before the  adaptation      is finally disabled bythe    detector.
become increasingly more important with the sharp             increase         The purpose of this paper is to propose an adaptive echo
of large loop delay telecommunication channels, particularly               canceler with a new structure, overcoming the double-talking
with the    advent of satellite communication. In a channel                problem. The    proposal      is to have two separate echopath
with large loop delay, it is clear that smooth conversation is             models (EPMs), one (background EPM) for adaptively identify-
difficult without proper echo control[ l ] , [ l l ] , [ 121.              ing echo   path           characteristics
                                                                                              transfer              and     theother(fore-
    Although an echo suppressor [14] , [15] has been utilized              ground EPM) for synthesizing the echoreplica to cancel out the
as an echo   control     device, it
                                  inherently includes the well             echo. Theparameter values of the foreground EPM are re-
known shortcoming of speech clipping. It also has a problem                freshed by those of the background EPM, only when the latter
in thatecho is littlecontrolledorlittle        suppressed during           is determined, bysome criteria,to give abetterechopath
double-talking.                                                             transfer characteristics approximation than the former.
    To overcome the shortcomings of an echo suppressor, the                    In this paper:thestructureand          main functions of the
echo canceler has been investigated [ 11 , [3] - [9] . An echo             proposed echo canceler are described; its performance charac-
canceler has some means of measuring echo path character-                  teristics are illustrated through computer simulations: a com-
istics and of synthesizing a replica of theecho using the                  pletely digital hardware implementation is described; and ex-
measured  characteristics.    In other words, it has an echo               perimental results are given for double-talking protection
path model (EPM). The echo replica generated is utilized for               capability, convergence characteristicsand a subjective eval-
canceling out the actual echo. The measurements of echo path               uation.
characteristics are carried out byadjusting parameter values
within the EPM so that EPM transfer characteristic conforms                                  11. SYSTEM DESCRIPTION
to that of the echo path. Among the measurement methods,
the one that adaptivelymeasures or identifies the echo path                   A basic structure of the proposed echo canceler i s depicted
characteristic from the input and output     speech signals [3] -          in Fig. 1. Thesystemoperates      on 8-kHzsampled data. De-
 [9] is considered most favorable, because an actual echo                  noting discrete time by k, the incoming signal z ( k ) to the echo
path is not perfectly constant [ S I .                                     canceler send-in terminal is the sum of the echo y ( k ) and near-
    The adaptive method, however, raises a new problem. The                end talker’s speech n(k). Thecircuit noise considered is in-
problem derives from the fact that, for correctly identifying              cluded innear-end talker’s speech signal n(k). Estimation of
the characteristics, theoutput signal of an echopathmust                   the echo path transfer characteristic is carried out, as a back-
                                                                           ground operation, by the background EPM and ADAPTATION
                                                                           PROCESSOR, using the background error signal ecb)(k)and
   Paper approved bytheEditorforSpaceCommunication          of the         receive-in signal x(k). Specifically, the background EPM
IEEE CommunicationsSociety for publicationwithout oral presenta-
tion. Manuscript receivedNovember 2, 1975; revised May 10, 1976.           possesses aparameterset      hi(b)(k)(i = 0 , 1, -, N - 1) to
This work was initiated based experience gained through studyunder
                             on                                            approximate the echo path transfer characteristic and a means
INTELSAT   ContractCSC-SA-141.            was constructed under
                                   Hardware                                for synthesizing an echo replica (’y(b)(k))    corresponding to
INTELSAT Contract CSC-IS-351and modified afterwards.
   The authors are with Nippon
                         the     Electric
                                        Company, Kawasaki,                 this set of parameter values. The backgroundecho replica
Japan.                                                                     (y(b)(k)) subtractedfromthe
                                                                                      is                       send-in signal z(k) andthe
590                                                                                                                IEEE TRANSACTIONS ON COMMUNICATIONS, JUNE 1 9 7 7

                                SENDIN                                                SEND OU1           that the transfer    is prohibitedfor the clear double-talking
                                                                                                         condition, easily detectableby the conventional signal level
                                                                                        To far end
                                                                                                         comparison method.
                                                                                                             It appears that condition (i) alone might work well for the
                                                                                                         present purpose. This is because it is generally considered that
                                                                                                         the smaller the error power is, the better the model character-
                                                                                                         istics are. Actually, condition (i) was found to work well, as
                                                  U                                                      long as the break-in near end talker’s speech signal (n(k))is
                                XIK           AOAPTATION PROCESSOR                     Speech tmrn tar
                                                                                       sod talker        assumed to be randomnoise. Consequently,condition (i) is
                             REC OUT                                                  REC IN             effective fora break-in of such noise-likespeech sounds as
                           FG     EPM : twegromd         echo path modal                                 voiceless consonants,for example Is/, /h/,etc.For         voiced
                          BG    EPM : background echo path m0d.l
                                                                                                         sounds, in particular for vowels, however, condition (i) alone
                                      k   :   Indlcalel dlsumte   tlrne.
                                                                                                         was not satisfactory   and thetwoadditionalconditions        (ii)
  Fig. 1.    Functional block diagram showing basic                           proposed adaptive
                               echo canceler.
                                                                                                         and (iii) were introduced.
                                                                                                             Finally 0 = 0.875, = r 3 , M = 128 (16 ms), D = 3 and
                                                                                                         T = 128 ms were selected by preliminary experiments for the
error signal               is fedbackcorrect set
                                    to     the              of parameter                                 values of 0,y. M , D and T. It might appear that 18 dB echo
values h i ( b ) ( k ) .Parameter values h i ( f ) ( k of theforeground
                                                       )                                                 cancelation (y = T 3 ) is too large fora transfer condition.
EPM are supplied by those of the background EPM, by means                                                However, experiments showed that smooth transfer is actually
of the TRANSFER CONTROL LOGIC, when the logic deter-                                                     possible, asis shown later by an example through a computer
mines that the background EPM is better than the foreground                                              simulation.
   In the following, a transfer control logic andadaptation                                              B. Adaptation Algorithm
algorithm, which are the two main functional elements of the                                                The learning identification method of Ref. [2] was adopted
echo canceler, are described.                                                                            as theadaptationalgorithm.        Using. the algorithm, theecho
                                                                                                         path transfercharacteristics are identifiedin the form of an
A. Transfer Control Logic                                                                                impulse response. Thus, the echo replica is synthesizedby a
   As criterions determining transfer actuation, the following                                           transversal digital filter. The coefficients of the filter or the tap
three conditions were experimentally  selected    to be simul-                                           gains correspond, in this case, to the aforementioned param-
taneously satisfied:                                                                                     eters of an echo path model.
                                                                                                            Denote the i-th tap gain or i-th coefficient of the filter at
    (i) Lj(e(b)) OLj(e(f))                                                                       (1)                             )
                                                                                                         time k by h i ( f ) ( k and                       and
                                                                                                                                            for foreground background
                                                                                                         EPM, respectively. Then, the adaptation algorithm is expressed
    (ii) L j ( e ( b )< -yLyLi(z)
                      )                                                                          (2)
                                                                                                         exclusively in terms of h i ( b ) ( k ) ,as describedabove,   and is
   (iii) L j ( z )   <L j ( x )                                                                  (3)     given by.

            with hangover time T for the condition                                                          hi(b)(k    + 1) = hicb)(k)+ A h i ( b ) ( k )                 (5)

                                                                                                            A h i ( b ) ( k ) 01
                                                                                                                                   x(k - i)    - e(b)(k)
                                                                                                                                            (x(k - i))2
where L j ( A ) indicates ashort timepower of a signal A ( k )                                                                       i=
for j-th time interval, containing consecutive M samples, and
is approximated by                                                                                       for i = 0, 1, 2, -, N - 1 where N is the number of the digital
                                                                                                         filter taps, a is a constant and

                                                                             -).                 (4)                  =z(k)-~             ( ~ ) ( k )                     (7)
0 and y are constant positive values less than                                 1(b   < 1, y < l),          ~ ( ~ )=k )
                                                                                                                  (                            - x(k   -   i).
 respectively. Another condition to further restrict the transfer                                                          i=O
was also provided. Conditions (i)                          -
                                        (iii) were required to be
                                                                                                         Convergence (more correctly, nondivergence) is guaranteed
simultaneously satisfied over D consecutive time intervals, for
                                                                                                         for a value of 0 < 01 < 2 [2] .
the transfer actuation.
    Of the three conditions, (i) indicates that the error signal
level from the background EPMis smaller than that from the                                                                   111. COMPUTER SIMULATION
foreground EPM, by more than -20 log 0 dB. Condition (ii)                                                   The following computer simulations show examples of
indicates that the degree of send-in signal (identical with ‘echo’                                       detailed operations of the echo canceler, for both convergence
fornondouble-talking condition)       cancelationby     theback-                                         and double-talking conditions. Simulations were conducted for
ground EPM is larger than -20 log y dB. Condition (iii) means                                            the echo canceler with 324 filter taps, 15 bits filter coefficient
OCHIAI e t al.: ECHO CANCELER WITH TWO ECHO PATH MODELS                                                                                                            59 1

                                             Given echo path impulse response        w        5
 W        5                                                                           i       4
 -I                                                                                           .3
 >    4                                                                              W
          3                                                                          G        2
 :    2
 W                   X-x-pp.-X+X-X-
 =        I

                                                                                             30     -
 @    20
                                                                            l 5 dB
                                                                                             20     -                                                        -5dB

      IO                                                                   i o   I            10    -                                                        I‘
                                                                                                                                                             - 2

          2                                                                                                                                                    I

              0                       200                         400 ms                       0-                                                            -0
                                       / Hello /                                             -4     -                                                        --I

                           Receive-inlevel         = - IOdBmO
                                                                                                                        200                     400 ms
                           Echo return loss ( E R L ) - l O d B
                           IH(”(0)- IH‘bl(0) = 0                                                                         level
                                                                                                                 Recelve-in        = -10 dBmO
                                                                                                                  ERL    10dB
Fig. 2. A computersimulationexampleforconvergence.Thecurves
   were visually fittedtodata.Characterswithsidebarsareboldface                           Fig. 3.       A computer simulation for a double-talking condition. The
   in text.                                                                                                   curves were visually fitted to data.

accuracy and 13 bitssampled data accuracy, comforming to                                   the condition of “ ’ ( 0 ) = M b ) ( 0 )= 0. The figure indicates
the one actually implemented (described later in details). For                             that the transfer, from the background to foreground EPM, of
an echo path impulse response, a typicalonewith about 15-                                  the parameter (tap gain) values is very smooth, except for a
ms time duration (shown in Fig. 2) was utilized. To indicate                               very short period immediately following initiation of conver-
the degree of convergence the following parameters c ( f ) ( k )                           gence. As a reason for this smooth transfer, it should be noted
and d b ) ( k )were used for foreground and background EPM,                                that the degree of echo cancelation by the background EPM,
respectively.                                                                                                                   ,
                                                                                           20 log { L j ( z ) / L j ( d b ) ) }is much larger than a corresponding
                                                                                           d b ) ( k )value. This phenomenon was observed throughout all
                                                                                           thecomputer simulations.It also appeared thatthe same is
                                                                                           true for the hardware echo canceler described later. Although
                                                                                           the reason for this is not clear yet, one explanation may be as
                                                                                           follows. A short time frequency spectrum of a speech sound,
                                                                                           especially of a vowel, is usually concentrated on some fre-
                                                                                           quency regions and an EPM can get large cancelation, for that
Here, with N = 324, IIA                      =               ai2 for vector A = particular speech, by its transfer characteristics becoming good
{ao, 0 1 , .-, a N - l } , H = {ho, h l , -., h N - l } indicates 8-kHz only for the frequency regions, while discussed above, d b ) ( k )
sampledsequence of the given impulse response, H ( f ) ( k )                          = exhibits cancelation          capability      for signals with various fre-
{ h O t f ) ( k ) , h l ( f ) (-, ) h N - l ( f ) ( k ) } andHcb)(k) = { h o ( b ) ( k ) , quency spectrum shapes.
                                k ,
h,cb)(k), ..., h N - l ( b ) ( k ) } .                                                         Figure 3 shows an example of double-talking. The receive-
    In relation to echo cancelation,d b ) ( k )and c ( f ) ( kindicate in signal (x@)) is the same speech “Hello” as in Fig. 2 and the
average degrees of echo cancelation                      for theconditionthat              break-in speech (n(k)),another person’s “Hello,” was supplied
various receive-in speech sounds with a flat average spectrum at the levels c ( f ) ( k )= 23.6 dB and d b ) ( k > 23.9 dB. As is                   =
are used with     fixed values of                                            .
                                                           and H ( f ) ( k ) This condi- seen from curves 1 and 2 , double-talking is detected, by the
tion is roughly met duringdouble-talking, assuming that the                                signal levels comparison(condition              (iii)), at time instant a.
parameter values are properly frozen to H ( b ) ( k )and H ( f ) ( k ) . Note thatthebackground EPM parameter values (curve 6 )
Therefore d b ) ( k )and c ( f ) ( k give estimates of average degrees have already been contaminated heavily beforetime a. The
of echo cancelation during double-talking.                                                 foreground EPM ( 5 ), however, is shown to retain, undisturbed,
    Figure 2 shows typical convergence characteristics fora                                the parameter values immediately preceding the initiation of
receive-in speech “Hello”. The convergence was initiated with double-talking.
592                                                                               IEEE TRANSACTIONS ON COMMUNICATIONS, JUNE 1 9 7 7

          SEND   IN

                 1I      REGISTER


          0                                                                                                                  0
          REC OUT                                                                                                      REC I N
                                           Fig. 4. Functional designed echo canceler block diagram.

                      IV. IMPLEMENTATION                                  324). Other system parameters, such as bit lengths of X and
                                                                          H registers, are shown in Table 1. The ADAPTATION PROC-
    Average return loss of an echo path is 15 dB with standard            ESSOR  calculates      the
                                                                                                   correction     (Ahi(b)(k))  according to
deviation of 3 dB [13] and a minimum of about 6 dB [ 113.                 relation (6) and generates a new   coefficient value          + 1)
On the other hand, required over-all echo loss is about 31 dB             according to (5). Forthe gain factor a in (6), a = 1 was
ormore [ l ] , [ l l ] . Consequently, 25 dB was selected as a            selected. For the calculation of (6), exponent arithmetic was
target for the echo return loss enhancement (ERLE) attainable             used due to its simplicity of implementation. This means that
by the echo canceler.                                                     only the information on the exponent in a floating point ex-
    A completely digital implementation was used because of               pression of a number is utilized for approximating the mag-
its high reliability and because of the accuracy required in the          nitude of thenumber,orthat          only theinformation on the
internal processing, with- analog to digital and digital to analog        position of the  most                binary
                                                                                                     significant         bit is retained.A
converter attheinterface.324taps                ( N ) were used on the    minimum value was set for the calculation of Z ( x ( k - i ) ) 2 ,
digital filters. This value corresponds to about 40 ms at the             so that correction
                                                                                the               value                  smaller for weak
8-kHz sampling frequency used. The 40-ms value covers echo                receive-in signal.. This is because weak receive-in signal induces
path end'delay up to 25 ms, so as to meet CCITT recommen-                 small echo signal which is most likely contaminated by noise
dation (G.16'1) for an echo suppressor. Another 15 ms is used             on an echo path.
for the main part of an impulse response which appears after
an elapse of time corresponding to the enddelay.                                                V. EXPERIMENTS
    A functional block diagram of the designed echo canceler
is depicted'in Fig. 4. The figure also indicates signal relations            Experiments were conducted on the hardware echocanceler
within theecho canceler forthe timeintervalwhere                  error   implemented as above. An echo      return   loss enhancement
signals e ( f ) ( k )
                    and e ( b ) ( k )
                                    have been generated and new echo      (ERLE) value was measured as follows: at the time instant at
replicas y ( f ) ( k+ 1) and y ( b ) ( k+ 1) are being generated. The     which the value was measured, the foreground EPM parameter
H(f),            and X registers all consist of recirculated shift        values were frozenby disabling the transfer fromtheback-
registers. There are 324 register stages for each ofthe H ( f )           ground EPM. Receive-in speech (x@)) was replaced by a band
and H ( 6 ) registers, corresponding to the number of the digital         limited (300  -  3,400 Hz) white noise. The ERLE value was
filter taps. The number of the X register stages is 323 so that           obtained as send-in (z(k)) to send-out (ecf)(k))signal power
the relative position of x ( k - i>' samples and hi(k 1) values           ratio for thecondition. This correspondsto measuring the
shifts by onesamplewhen               the oldest sample x ( k - 323) is   c(f)(k) value defined for the computer simulation. As an echo
replaced by the new value x ( k + 1). The system clock, or the                                             -
                                                                          path, a bandpass fdter of 300 3,400 Hz was used in com-
clock frequency for the shiftregister, is 2.592 MHz (= 8 kHz X            bination with an attenuator giving echo return loss values.
OCHIAI er al.: ECHO CANCELER WITH TWO ECHO PATH MODELS                                                                                                                                                   593
                         TABLE 1                                                                         z
  1     Sampling "rate                           8 kHz (125

  2     S i z eo f   X and H r e g i s t e r s   324 s t a g e s (40 ms)
  3     Clock                                           H
                                                 2.592 M z (38611s)                                      v)
  5     A t o D and D t o A c o n v e r t e r    1 3b i t s( l i n e a r )                               3
        accuracy                                                                                         z
                                                 13bits                                                  3
  5     X registeraccuracy
        H registeraccuracy

                                                 15 b i t s                                              a
                                                                                                                   2o   t
                M u l t i p l yo u t p u t       19 bits

                 Accumulator                     24 b i t s                                                                             E cho       ( Send-in)          level
  -     CalculationinAdaptationCircuit           E x p o n e n tA r i t h m e t i c        Fig. 5.            Measured ultimate ERLE values for speech signals. The line
                                                                                                                         was visually fitted to data.

A . Convergence Characteristics
   As is expected from the computer simulations, convergence
characteristics similar to those of a cancelerwith one EPM
[ 161 are actually obtained.
    Figure 5 shows ultimately  obtainable    or ultimateERLE
                                                                                                                                                                                   TRACK1         TREK 2

values. It also shows that the 25-dBtarget is almost always
realized for the typical echo level of -20     -30 dBmO. The
ultimate ERLE.value was measured as the ERLE value after                                          I                                             C    L       O         W
a sufficiently long speech (15 words or more) was supplied as
the receive-in signal (x(k)). Convergence was initiated with
both theH registers cleared.
                                                                                           TC : t r m s t r c a n t m l         dgml
    A 5 word 2 second longtelephonespeech           was used to
                                                                                           CI : double talklng
measure convergence rate, or convergence speed, the  and                                        counter
ERLE value immediately after cessation of the speech was                                   C2 : transfer            control signal
measured. The     results showed that about20-dB ERLE         is                                counter

realized in this case for the same conditions shown in Fig. 5.                                          Fig. 6.         Experimental setup for measuring false transfer.
Tests were also conducted with end delays up to 25ms inserted
in the echo path in addition to the filter and attenuator. The
                                                                                                                 TABLE 2
results were almost identical to those described above.                                     ERROR RATE MEASUREMENTS FOR TRANSFER CONTROL
                                                                                                    FOR SPEECH. ERL=10 dB. RECEIVE-IN
B. Double-TalkingProtection Capability                                                                   SPEECH LEVEL=-10 dBmO
   To evaluate performance for the double-talking protection                                N-X         (dB)                               0         -10                   -20                 -30
capability, the      shown Fig.
               circuit      in              6 was prepared. This
                                                                                                                                        -10              0                  10                   20

configurationenables automaticdetection of double-talking,                                  Y-N         (dB)

thus making it possible to count transfer actuations during                                 FALSE                           A              0             o         2.5 x                   7 x    lo-*
double-tallung, or false transfers.                                                         TRYSFER
                                                                                                                            B              0             o         7       x               2      LO-^
    Transfer of parameter values frombackground EPM was
suspended after completion of convergence. The number of
control signal occurrences actuating the transfer under double-                              X:       R e c e i v e - i ns p e e c hl e v e l

talking conditions was counted. Break-in speech signals were                                 N : B r e a k - i ns p e e c hl e v e l
added after completion of convergence and suspension of the                                  Y: Echolevel
transfer. The total double-talking length was counted by                                     A : Same t a l k e r ' s s p e e c h f o r b o t h t h e r e c e i v e - i n a n d t h e
counter c1 in 16-ms units. The counter counted the number                                             b r e a k - i ns i g n a l s .       About 40 minuteseachforonemaleand
of transfer control signals (which were issued synchronously                                          o n e e m a l e p e a k e rT o t a l ,
                                                                                                          f         s            .                       80 minutes.
withthe 16-ms           clock
                 interval           pulse) duringdouble-talking.                             B: D i f f e r e n t t a l k e r ' s s p e e c h f o r t h e r e c e i v e - i n           and t h eb r e a k -
False transfer rate was measuredby a 2 / a l , where al and a2                                        i ns i g n a l ,T o t a l ,          40 minutes.
indicate, respectively, the final output of c1 and c2 after all
test speeches have been             The
                           supplied. speech          signals used
included dictation of stories and were arranged so that double-                           The false transfer was also zero when the break-in signallevel
talking occurs frequently.                                                                was 10 dB smaller thanthat of the receive-in signal. False
   Table 2 shows results of measurements. The table indicates                             transfer occurs when break-in signal level is more than 20 dB
that, for the most frequently encountered condition of equal                              lower than the receive-in signal level. However this would be
send-in and receive-in speech level, the false transfer was zero.                         noproblem, since such    conditions are rarely encountered.
594                                                                             IEEE TRANSACTIONS ON COMMUNICATIONS, JUNE 1 9 7 7

Further, the degree of disturbance fortheechopath         model
will be smaller for such smaller break-in signal conditions, even
if a false transfer happens to be actuated.

C. Subjective Evaluation
     A subjectiveevaluation experiment was conductedunder
laboratory           The
           conditions.         newly developed echo canceler
with twoechopath models (two EPMs E.C.) was evaluated                                                ERL = 15dB
together with an echo canceler with one echo path model (one                                                          @
EPM E.C.). In the latter, the direct method of      enabling and
disabling adaptation was used as a double-talking protection                         0
strategy. The control was actuated according to the output of                        E5
                                                                                           2         U @
a double-talking detector, whose logic is identical to condition
(iii) of thetransfercontrol logic. Adaptation algorithm, the
                                                                                      -    1 1

number of the digital filter taps and accuracies which charac-                                       ERL = 24 dB
terize the one EPM E.C. were the same as those of the two
EPMs E.C..
    Figure 7 shows the results and clearly indicates that the two                                                              Q
                                                                                     w     3
EPMs E.C. is by far better rated than the one EPM E.C.. Each                         a
                                                                                     a    2
point represents the average value of 20 responses. The prin-
cipal conditions of the experiment were as follows. Model 600                        z
                                                                                     -     1
                                                                                                     300    600    900           n
                                                                                                                           1200 r s
telephonehandsets (generally usedincommercial           service in
Japan) were used. Sixty (60) phons (A scale) Hoth room noise,
which was measured at the handsets of subjects in telephone
booths, was added. Two types of random circuit noise (-8                                  2 : Good            Q Two EPMs E.C.
                                                                                          3: Fair
dB/oct, 0.5 mV;flat, 1 mV), were added to thetwo wire                                     4 : Bad             0 No canceller.
terminal at the two-fourwire junction.                                                    5 : Poor                 delay only.
                                                                      Fig. 7.    Subjective evaluation results. The lines were visually fitted   to
                       VI. CONCLUSION

    Anew adaptive echo canceling method, using twoecho               Mr. M. Tajima, consultations by Mr. T. Matsushima, Dr. Y.
path models, was proposed to overcome the false adaptation           Ohashi, Dr. H. Kaneko   and Mr. R. Mimori. Valuable sug-
problem for double-talking. Using a control logic for param-         gestions and discussions by Mr. S. Chiba and Mr. T. lshiguro
eter values transfer,fromthe background EPM tothefore-               are deeply appreciated. The authors also express many thanks
ground EPM, it was shown that the method enables virtually           to Mr. H. Kaneko for supplying them with excellent codecs,
complete double-talking protection, retaining convergence            to Mr. S. Tanaka for his help in experiments and tomany
characteristics similar to that in case of using one echo path       others who were quite helpful in implementing the hardware.
    Simpler and more effective transfer control logic should be
sought further to simplify theecho canceler structure. Sub-
jective evaluations in comparison with echo cancelers utilizing            A.Miura, S. Kobayashi, R. Satoand K. Nagata,“ABlockless
                                                                           Echo Suppressor,” IEEE Trans. Commun. Technol., Vol. COM-
other double-talking protection strategies    should      also be          17, NO. 4, pp 487-495, Aug. 1969.
conductedfurthertojustifyactual          effectiveness, since the          J. Nagumo and A. Noda, “A Learning Method of System Identi-
echo canceler possesses two costly convolution circuits. How-              fication”, IEEE Trans. On Automatic      Control,    Vol.AC-12,
                                                                           No. 3, pp 282-287, June 1967.
ever, it is possible to consider total cost reduction on a multi-          Y. Kato, S. Chiba, T. Ishiguro, Y . Sato, M. Tajima, T. Ogihara,
channel base. In this case, the present structure is very natu-            S. J. Campanella, H. G. Suyderhoud and M. Onufry, “A digital
rally, as as
         well       effectively,
                               extensible tothe multichannel               Echo Canceller”, NEC Researchand Development, No. 31, pp
                                                                           32-41, Oct. 1973.
systemin which an adaptive echo canceling circuit,thatis,                  S. J. Campanella, H. G. Suyderhoud and M. Onufry, “AndlySiS
thebackground EPM withadaptation processor, is provided                    of Adaptive
                                                                              An         Impulse
                                                                                               ResponseEcho Canceller”,                Comsat
for a plural number of foreground EPMs.                                    TechnicaIReview, Vol. 2, No. 1, pp 1-38, Spring 1972.
                                                                           M. M. Sondhi, “An Adaptive Echo Canceller”,       Bell Syst. Tech.
                                                                           J, Vol. 46, No. 3, pp 497-511, March 1967.
                    ACKNOWLEDGMENT                                         F. K. Becker and H. R. Rudin, “Application of Automatic Trans-
                                                                           versal Filters to the Problem of Echo Suppression”, Bell Syst.
                                                                           Tech. J., Vol. 45,No. 12, pp 1847-1850, Dec. 1966.
   The authors would like to thank Dr. S . J. Campanella, Mr.              M. M. Sondhi and A. J. Persti, “A Self-AdaptiveEchoCancel-
H.G. Suyderhoudand Mr. M. Onufry,who were in charge                        ler”, Bell Syst. ,Tech. J., Vol. 45, No. 12, pp 1851-1854, Dec.
of the INTELSAT contracts.They also appreciateconstant                     1966.
                                                                           H. Brueggemann, G. P. Kidd, D. K. Machechaie and A. J. Segler,
encouragement inthis workby Dr. T. Sekimoto, leadership                    “Experiments on AdaptiveEchoCancellation”,           A . T.R., Vol.
and pertinent advice by Mr. K. Fujimoto, Mr. Y . Kat0 and                  4, No. 1, pp 10-15, May 1970.
IEEE TRANSACTIONS O N COMMUNICATIONS, VOL. COM-25, NO. 6 , JUNE 1 9 7 7                                                                              595

      N. Demytko,L. K. Mackechnie,“A High SpeedDigitalAdap-                   tory, Linguistics Department, University of California, Berkeley, CA.
      tive Echo Canceller”, A.T.R., Vol. 7, No. 1 , pp 20-28, 1973.              Mr. Ochiai is a member of the Institute of Electronics and Communi-
      J . R. Rosenberger and E. J. Thomas, “Performance      of an Adap-      cation Engineers of Japan and the Acoustical Society of Japan.
      tive Echo Canceller Operating in a Noisy, Linear, Time-Invariant
      Environment”, Bell Syst.   Tech. J., Vol. 50, No. 3 , pp 785-
      812, March 1971.
      G. Williams, and L. S. Moye, “Subjective Evaluation of Unsup-
      pressed EchoinSimulatedLong-delayTelephoneCommunica-
      tions”, Proc. IEE, Vol. 118, No. 3/4, pp 4 0 1 4 0 8 , March/April
      1971.                                                                                       Takashi Araseki was born in Hokkaido, Japan,
      G . K. Helder, “Customer Evaluation of Telephone Circuits with
                                                                                                  on Oct. 14, 1946. Hereceived the
      Delay”, Bell Syst. Tech.      J., Vol. 45, No. 7, pp 1157-1191,
                                                                                                   from Hokkaido University, Hokkaido, Japan, in
      Sept. 1966.
      R. G. Gould and G. K. Helder, “Transmission Delay and Echo                                   1969.
                                                                                                      He joined, in 1969, Central Research Labora-
       Supression”, IEEE Spectrum, Vol. 7, No. 4 , pp 47-54, April
                                                                                                       Nippon Company
                                                                                                   tories, Electric     Limited,
                                                                                                   Kawasaki, Japan,wherehe hasengagedinre-
       P. T. Brady and G. K. Helder, “Echo Suppressor Design in Tele-
                                                                                                   searchanddevelopment      on speechsignalpro-
       phonecommunications”, Bell Syst. Tech. J., Vol. 42, No. 6 ,
                                                                                                   cessing,inparticular   onechocontrol    devices,
      pp 2893-2917, NOV.1963.
                                                                                                  ipeech band compression and speech synthesis.
       J. Nakamichiand A. Sato,“Echo SupressorDesign forSpace
                                                                                                      Mr. Araseki is a member of the Institute of
       Communication” Proc. ofIECE, Vol. 49, No. 1 , pp 53-59, Jan.
                                                                              Electronics and Communications Engineers of Japan and the Acoustical
       1966 (in Japanese).
       H. G. Suyderhoud and M. Onufry,“Performance of A Digital               Society of Japan.
       Adaptive Echo Canceller in a simulated Satellite Circuit Environ-
       ment”, Proc.of AIAA 4th Communications Satellite Systems
      Conference, Washington, D.C., April 24-26, 1972.                                                           *
                                   *                                                                  TakashiOgihara    graduated      the
                                                                                                                                  from Nippon
                        Kazuo Ochiai received the B.S. degree from the                                       Technical
                                                                                                      Electric       College,Kawasaki, Japan,
                        Universityof     Electrocommunications,  Chofu,                               in 1966.
                        Tokyo, Japan, in 1963.                                                           In 1960, he joined Nippon Electric Company
                          Sincehe     joinedCentral     Research
                                                               Labora-                                Limited,Kawasaki, Japan, wherehehasbeen
                        tories, Electric
                            Nippon Company      Limited,                                              engaged in the circuit designs for speech signal
                        Kawasaki, Japan, in 1963, he has been engaged                                 processing equipments, such as speech band
                        inresearchanddevelopment         on speechsignal                              compressionsystem,echo       cancelersandvoice
                        processings,inparticular      on speechsynthesis,                             operated equipments.
                        speech band compression and echo control. For                                    Mr. Ogihara is a member of the Institute       of
                        one year from 1968, he was a Guest Researcher                                           and
                                                                                                      Electronics Communication            Engineers of
                        on speech synthesis, at the Phonology Labora-         Japan.

             400 Mbit/s Digital Repeater Circuitry Provided a
             New Design Method and Beam Lead GHz Transistors

   Absrruct-A       compact, high-performance
              stable,        and                 regenerative                 new computer-aided design method, and has been successfully applied
repeater circuitry, suitable for digital transmission systems up to several   to the 4 0 0 Mbit/s experimental coaxial cablePCM system.
hundred Mbit/s, has been provided through utilization of      new devices,         Major features of this repeater circuitry are:
such as 7 GHz beam lead shielded junction transistors, and through a               (1) anequalizingamplifierwithlownoisefigure         (7.6 dB),small
                                                                              intersymbolinterference(12%),andautomatic           Line equalizationof
    Paper         by Editor Transmission
         approved the      for                  of
                                          Systems the                         21 dB tracking    range at 200MHz;
                                                                                   (2)aregenerative     outputcircuitwithbipolar   pulses of 2.4 Vop
the International Conference on Communications,      San Francisco, CA,
June 17-19, 1975. Manuscript     received
                                        January        8, 1976; revised       amplitude and 700 ps rise time; and
February 21, 1977.                                                                 (3) total performance with sufficient noise margin (10 dB for error
    M. Kawashima and I. Fudemoto are with Fujitsu Limited, Kawasaki,          rate 10-l’ over a line of 56 dB loss at 200 MHz), small static pattern
Japan.                                                                        jitter(20°pp), smallersize (270 X 160 X 5 2 mm),andlowerpower
    Y. Mochida and T. Uyehara are with Fujitsu Laboratories Limited,          (5.8 W).
Kawasaki, Japan.                                                                   These have been achieved by use of:

To top