A Continuously Variable Digital Delay Element.

C. IC. F a r r o w
ATkT
Lliddletown, New Jersey 07748

.4 BSTRA CT

T o transfer digital samples of analog d a t a from a system

running under one clock t o a system running under
another clock requires an interpolator t o compensate for
delay between clocks.

This paper describes a Finite Impulse Response (FIR)

Filter which synthesizes a controllable delay. By
changing the delay the filter has the ability to
interpolate between samples in the d a t a stream of a
band-limited signal. The original motivation for this
work was t o provide a digital interpolator capable of
compensating for the delay between echo canceller
output and the receiver input for echo cancellation
based modems, for example CCITT's V.32 modem.
Another possible application is the transfer of d a t a
between quasi-synchronous T-carrier systems. T h e
variable delay filter can also be used a s a more general
computational element. Because high sampling rates are
not required, this filter is especially suited for
implementation on a DSP (Digital Signal Processor).
The interpolator has been implemented, in real time, on
a WE@DSPBO. This interpolator can also be used as a
practical way t o reconstruct an original band-limited
signal from samples taken at the Nyquist rate.

One way t o do this is t o generate an analog signal by

outputting the digital samples through a Digital-te
Analog Converter (DAC): pass the DAC o u t p u t through
a Low-Pass Filter and then re-sample the LPF o u t p u t
using an Analog-teDigital Converter (ADC) (Figure 2).
F i y r a 2.

Fx I t e r
C o n v e r t er

C o n v e r te r

Input

Output

Sanip l e
Clock

Samp I e
CI ock

R N R C O G

I N T E R P O L R T O R

1. INTRODUCTION.

A Continuously Variable Digital Delay (CCDD) Element

has been built using a n FLR filter whose tap coefficients
are a function of the desired delay. This is represented
in Figure 1. where T = t a p spacing, aT = Delay.

Fiqure

-4 new approach, presented here. uses a DSP as a

Continuously Variable Digital Delay ( C l D D ) (Figure 3).
The "input" samples are sampled into the "input'* buffer
by the "input" clock. The DSP is synchronous with the
"output" clock, which also measures the delay between
the output and input clocks t o give Q = D / T . The DSP
has a delay line into which d a t a samples are shifted.
Normally just one new sample is shifted into the delay
line from the input buffer.

Figure 3 .

Input
Samples

4A

[
f

Digital \
iprocessori p f

Input
Samp I e
Clock

V e c t o r F o r m of C o n t i n u o u s l y V a r i a b l e
D i g i t a l Delay.

ISCAS'88

D I G T R L

gutpyt
amp e s

output
Sample

Clock
I N T E R P O L R T O R .

2641
CH2458-818810000-2641$1 .OO (C 1988 IEEE

If t h e delay parameter a overflows two n e u * samples are

shifted i n t o t h e delay line; if a underflows t h e n no new
samples are shifted in. T h e interpolated samples
generated by the D S P are t h e n transferred t o t h e
"out p ut " b u ffe r.

-64

Fiqure 5 .

Serl al - t o

Para1 lel-

Paralell
Converter

to-Ser I a1
Converter

2. APPLICATIONS.

and

Following are some interpolation applications of t h e

CVDD:

De-Coder

Echo Canceller.

INTERPOLRTOR

and

YLaw
Coder

YLaw

-64

The application t h a t stimulated this work was t h e Echo

Canceller (EC). used t o cancel echoes of t h e local
transmit signal in modems where t h e transmit and
receive signals occupy t h e same spectrum [ l ] . .4 highlevel architecture of t h e modem is shown in Figure 1.
Without a n EC, t h e receiver has no way t o distinguish
between the far-end signal and t h e echoes of t h e nearend signal.

Kb/s
Ser7al B > t
Stream I n

Figure 4.

Kb/s

Serral B l t
Stream Out

Ser i a I-toParalell
Convcrte r
and

+ INTERPOLRTOR

_j

Data
Modem

?-Law

De-Coder

64

9600 Hz C l o c k

Kb/s

Serial Bit
Stream I n

Hybr i d

.
l
,

INTERPOLATOR

....... ....

...... .

flrchltecture

Recerver

Clock

of

V o i c e - B a n d Modem,

The E C operates under the transmit clock, because the

signals from the transmitter are used as a n "Ideal
Reference" t o generate a signal t o cancel echo signals
generated synchronously from the transmitted signal.
The signals from the E C now have t o be re-sampled a t
t h e Receive Clock rate, for processing in t h e receiver.
Although t h e far transmitter operates a t nominally t h e
same rate as the near transmitter. their two signals are
generated using separate crystal oscillators so t h a t their
phase difference may be constantly changing i.e. there
is a slight frequency offset.
PCM-PCM Interface.

Figure 5 shows a method of converting P C M (Pulse Code

hiodulation) d a t a samples from one T1 carrier system t o
d a t a samples for another T1 carrier system, each
operating under independent "8OOO Hz" clocks (see 12)).
PCM-"Analog" Modem Interface.

3. IMPLEMENTATION.

If we make each of t h e t a p coefficients of t h e system of

Figure 1 an n t h order polynomial in a , then we can rearrange the structure t o have n + l FIRS with fixed
coefficients, and t h e delayed o u t p u t is a polynomial in
delay whose coefficients are t h e o u t p u t s of each FIR.
This structure is represented in Figure 7.
4. COEFFICIENT COMPUTATION.

T h e transfer function of a filter with a flat delay r is

given by:

G(&t,r)= e l d T

(1)

and the transfer function of a n FIR with coefficients

* * *
Ck . . . and sampling interval T is:

using a delay parameter (I such t h a t r=aT and t a p

coefficients which are a polynomials in a then:

Figure 6 show a way of converting T 1 carrier PCXi

samples a t 8000 Hz directly to samples a t 9600 Hz into
t h e front end of an "analog" modem, thus obviating t h e
need for a n ADC or a D-IC.

2642

If we t o want t o determine the coefficients for a CLDD,

we minimize:

from (6) we see that

Having optimized [ C ] over one sample interval, we can

find the derivative of t h e interpolated function of s at
any point in the range - % 5 a 5 %. Indeed we can
make a good polynomial approximation of the
interpolated function of x over the above range.

with respect t o [ C ] .

For a filter of length N t h e filter coefficients are

numbered from C o ( a ) t o C V - I ( N ) . Using a C \ D D of
length N, where S is even, we subject (4) to the
constraints:

n # (-V-1)/2
12 = (-V-1)/2

Cn(Y2)= 0;
= 1;
C,(

- %) = 0:

(5)

# .\/2
n = -lj/2
12

= 1:

6. PERFORMANCE.

With the above constraints a n d a delay of + % - t a p

about the center of the filter the sample just before or
just after the center is transmitted exactly (with a n even
length filter. there is no center tap). This Delay element
is not intended t o serve any smoothing (or band
limiting) function and no a t t e m p t is made t o control its
out-of-hand response, as will be seen later.

A laboratory model of the CI'DD used as a n interpolator

DSP2O. In this set-up.
has been implemented o n a \.\EmB
the line signal was sampled into an ADC at a constant
clock rate, asynchronously with t h e transmitter clock.
These samples were then fed into a n interpolator using
the C \ D D a n d interpolated by the receiver clock, into a
\.*.32 receiver. The receiver clock was derived by the
receiver "timing recovery" from the interpolated
samples. Tests run over this system using a V.32 signal
showed no significant degradation in performance.
LVhat w a s used was a n 8-Tap, 3rd-Order CVDD; the tap
coefficients are given in Table I.

-.013824

5. COMPUTATIONAL APPLICATIONS.

Referring t

-. 157959

. I008

.6 16394

- I ,226364

- 616394
-. 157359
.054062

I .226364

- ,1008

-.013824

-. 003 I 4 3

.01s287

.63 I836
-.465576
-. 465576
.631836
-.21624a
.055298

3
-.012573
.077148

- ,483198
,905457
-.905457
,403198

-.E77148
.012573

C o e f f i c i e n t s , C(N(MI, f a r % T a p , 3 r d O r d e r ,
D i t i a l V a r i a b l e D e l a y w i t h Tap S p a c i n g 1/9600 s e c s .
and U p p e r U s e f u l F r e q u e n c y o f 3150 Hz.

E
I

2
.mm
-. 2 1 6 2 4 8

-. 0 19287

.054062

Figure '7.

1
.003 143

This CVDD was designed t o operate over a frequency

band from 0 t o 3150 Hz. Figures 8 and 9 are plots of
Delay a n d Amplitude error over this frequency range for
-%
a 5 VZ. These figures show relatively level
plateaus for the range of interest. As was stated in
section 5 . no attempt has been made to constrain the
out-of-band response. The CVDD is a linear FIR, for
any given setting of a , and if the timing drift between
transmitter and receiver is very small. i.e d a / d t is very
small, then a receiver using an adaptive equalizer will
have no trouble tracking any slight variations of delay or
amplitude introduced by the Ck'DD.

<

K(Delay
Parameter)

n
m
S t r u c t u r e o f , 3 r d Order,
Figure 7 .
Continuously V a r i a b l e D i g i t a l D e l a y .

Figure 10. shows the behavior of Cn(a)as a function of

a. It can he seen that, as a consequence of the
constraints (5) t h a t C,(-%) = 1 and and C3(%) = 1. A
further consequence of ( 5 ) is that:
Cn o

(V,).Cn,

(%)'.Cn, 2

4- (%)3'Cn,
3 .= O

c, 0 + (-%).C, 1 + (-%)?.Cn,2+ (-%)3.Cn,3

=0

this agrees with the fact t h a t the numbers in column 2

2643

of Table I are -4 times t h e numbers in column 0. a n d t h e

numbers in column 3 are -4 times t h e numbers in
column 1: except for rous 3 a n d 1. This suggests t h a t
this CLDD can be implemented n i t h only 2 instead of 4
FIRs, thus reducing t h e a m o u n t of computation. In fact
for any polynomial of degree "N" t h e CCDD can be
implemented with N-1 FIRs. This type of CCDD has
three parameters:
1. Upper Useful Frequency.
2. Number of T a p s (Nt).
a n d 3. Degree of Polynomial (Dp).

Dc =

-0.5

r
i
l
,
=

Figures 11. and 12. are plots of t h e upper useful

frequency for -40 d B and -60 d B rms error for CVDDs of
various numbers of taps a n d degrees of polynomial.
T h u s it is possible t o select Dp a n d N t t o meet t h e
requirements for any particular CCQD.

I II

0.25

ac= 0.5

Plot of Tap Coefficients (C(Alpha,N)) for

various values of Alpha for the Digital
Variable Delay with the Coefficients given
in table I.
32

Fiqure
N

28

11.

Po I y n o m i a l I n t e r p o 1 a t o r ,
Upper U s e f u l Frequency
v s . Number o f T a p s
f o r RMS E r r o r = - 4 0 d B .

24
r
20
0

16

P
Do
eg
l yr n
ee
o moi fa l

T
12 a

Plot of Normalized Delay Error vs. Delay

(Alpha) and Frequency for the
Digital
Variable Delay with coefficients given in
table I.
U
a, a,

.%

N 7 J .

L a
O S
Z E

26

.
.-*

(,

0 2400

Fiqure

.2

E-L

Dp ) =3,'

Frequency,

00 6

- @ U
I d - L

Dp>3

.,..

Dp=l

32

.-,L

/,

3000

3600

Hz

4200

12.

Polynomial Interpolator,
Upper U s e f u l Frequency
v s . Number o f T a p s
f o r RMS E r r o r = -60 dB.

24

Degree o f
Po 1 y n o m i a

r
30

(Dp)=
o

16

/p4

Dp,=2

12 a

Upper
Usefu1
Frequency, Hr

!/

Plot of Normalized Amplitude Error vs.

Delay Setting (Alpha) and Frequency for
the
Digital
Variable
Delay
with
coefficients given in table I.

2644

00

1200

1800 2 4 0 0

3000

3600

4200

T h e plots of Figures 8,9,11 a n d 12 were all made for a

sampling rate of 9600 samples/second.

7 . FEATURES.

REFERENCES

\\-hat has been described is a Continuously \-ariable

Digital Delay Element which can be used to generate a
delayed version of a sampled signal from the samples.

Jean-Jacques LC'erner. "An Echo-CancellationBased 4800 Bit/s Full-Dupleu DDD hIodem'*.

in
IEEE
Journal
on
Selected
Areas
Communication. i - 0 1 SAC-2. S o 3 . September
1984.

The CLDD can be incorporated in Digital Signal

Processing applications as interpolator for applications
such as modems with Echo Cancellers. transfer of PChI
d a t a from one clocked system t o another, or a more
general computational element for generating functions,
i.e. the derivative, of the sampled signal.

J. C. Candy and 0 . J. Benjamin. "A Circuit T h a t

Changes the \Vord Rate of Pulse Code h[odulated
Signals", The Bell System Technical Journal, I7oI.
62, No. 4. April 1983, pp 1181-1188.

When a band-limited signal is sampled at the Nyquist

rate it is theoretically possible t o reconstruct the original
signal from the samples. One might ask the question
"how?". The Continuously Variable Digital Delay
described here gives an answer: at least for a limited
length CLDD. we can come up with a very good
ap p rosimation.
8. COMPARISON.

In this approach t o interpolation. an interpolated sample

is generated only as needed. Other approaches [2]
generate a continuum of closely spaced samples and
select the desired outputs from them. Thus, the amount
of computation required here is much less than other
approaches. .%lso. because "Delay" is a parameter in the
computation, the resolution of the delay is limited only
by the precision of the arithmetic. not by the
computation rate.

APPENDIX
Figure A1 shows two different representations of the
same FIR (Finite Impulse Response Filter). Figure A1 is
included here t o explation the symbology used in the
rest of the paper.
M-Stage

Delay L l n e .

9. EXTENSIONS.

It should be apparent t h a t , within reason, we can

generate filters with arbitrary transfer functions of the
form C(w,z). One such application might be t o control
the formant in music applications. Extending the same
concept t o more dimensions we could simultaneously
control the delay and amplitude distortions in a
transmission line simulator.

Outut
( 2 )

- .
L

i'l-Stage

F i n i t e

Impulse

Response

Coefficient
Vector.

S a m ro l s
Vcct o r .

o u t ut
( 2 )

V e c t o r Form o f

F i g u r e R1.

2645

'

FIR Filter.

F i l t e r .

'