A/D & D/A

Conversion
LEARNING OBJECTIVES

$Develop the sample rate conversion process as the process for

mutlirate systems

þ Perform sample rate conversion

þ Use cascade equivalences for sample (rational) rate conversion

J Tutorial:
!Solve problems related to the above.
<MATLAB Exercises

&Chapters 10,’Digital Signal Processing’, Sanjit K. Mitra

AFMH Created: Last update: DSP_L6_T1

June 1999 July 2005

A/D & D/A Multirate Systems –

Conversion Sample Rate Conversion
• In many practical applications of DSP, there may be a need to change the
sampling rate of a signal. Examples:
– Digital audio: 32KHz in broadcasting, 44.1KHz in digital compact disk
(CD), and 48KHz in digital audio tape (DAT)
– Video:
Composite video: NTSC – 14.3181818MHz, PAL – 17.734475MHz.
Component video: 13.5MHz (luminance), 6.75MHz (chrominance)

• Two methods for sample rate conversion:

ΠThe sampled signal may be converted back into an analog signal and
then resampled.
• The signal may be resampled in the digital domain.

• The second approach will be discussed. This approach offer the advantage
of not introducing additional distortion in passing the signal through an
additional D/A and A/D conversion.
AFMH Created: Last update: DSP_L6_T2
June 1999 July 2005

1
 A/D & D/A Sample Rate Reduction
Conversion (by an integer factor, M)
Reducing the sampling rate by an integer factor, M:

With a new sampling period, Ts' = MTs , the resampled signal is:

xd [ n ] = xa [nTs' ] = x a [ nMTs ] = x[ nM ]

This can be accomplished by taking every Mth sample of x[n] - down-sampler.

x[n ] x d [n ]
↓M
M=2

Ts 2 Ts

AFMH Created: Last update: DSP_L6_T3

June 1999 July 2005

A/D & D/A MATLAB Example

Conversion Sample Rate Reduction (by an integer factor, M)

% Illustration of Down-Sampling by an Integer Factor, M=3

Input Sequence
clf;
1
n = 0: 49;
m = 0: 50*3 - 1; 0.5
Amplitude

x = sin(2*pi*0.042*m); 0

y = x([1 : 3 : length(x)]); -0.5

subplot(2,1,1) -1

stem(n, x(1:50)); 0 5 10 15 20 25 30 35 40 45 50
Time index n
axis([0 50 -1.2 1.2]);
Output Sequence
title('Input Sequence');
1
xlabel('Time index n');
0.5
ylabel('Amplitude');
Amplitude

subplot(2,1,2) 0

stem(n, y); axis([0 50 -1.2 1.2]); -0.5

title('Output Sequence'); -1

ylabel('Amplitude'); 0 5 10 15 20 25 30 35 40 45 50
Time index n
xlabel('Time index n');
Ref: Program P10_2, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T4
June 1999 July 2005

2
 x[n] xd [n ]
A/D & D/A Sample Rate Reduction ↓M
Conversion (by integer factor, M) M =2
Ts 2T s

Down-sampling generally results in aliasing.

Sampling period, T s is increased i.e. sampling frequency, Fs is decreased resulting in
spectral overlapping of the replicated spectrum.
∞
Recall the DTFT of x[n]=xa[nT s ] is: X (e j ω ) = X s ( jΩ) Ω=ω = T1s
Ts
∑X
k= −∞
a ( j Tωs − j 2Tπsk )

X (e jω )
2π
Ts
... ...
− 4π − 2π 0 2π 4π
ω
∞
Similarly the DTFT of xd[n]=x[nM]=xa[nMT s ] is: X d (e jω ) = 1
M Ts ∑X
r =−∞
a ( j MωTs − j M2πTrs )
X (e jω )
2π

... MTs
...
− 4π − 2π 0 2π 4π
ω
Spectrum stretches by factor M (amplitude scaled down by M) resulting in aliasing.

AFMH Created: Last update: DSP_L6_T5

June 1999 July 2005

A/D & D/A MATLAB Example

Conversion Effect of Down-Sampling in the Frequency Domain
% Effect of Down-sampling in the Frequency Domain, M=3
% Use fir2 to create a bandlimited input sequence
clf; Input Spectrum
freq = [0 0.42 0.48 1]; mag = [0 1 0 0]; 1

x = fir2(101, freq, mag); 0.8

Magnitude

0.6
% Evaluate and plot the input spectrum
0.4
[Xz, w] = freqz(x, 1, 512); subplot(2,1,1); 0.2
plot(w/pi, abs(Xz)); grid 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
xlabel('\omega/¹'); ylabel('Magnitude'); ω/¹

title('Input Spectrum');
% Generate the down-sampled sequence 0.5 Output Spectrum

M = input('Type in the down-sampling factor0.4= ');

y = x([1: M: length(x)]);
Magnitude

0.3

% Evaluate and plot the output spectrum 0.2

[Yz, w] = freqz(y, 1, 512); subplot(2,1,2); 0.1
plot(w/pi, abs(Yz)); grid 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ω/¹
xlabel('\omega/¹'); ylabel('Magnitude');
title('Output Spectrum');
Ref: Program P10_4, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T6
June 1999 July 2005

3
 A/D & D/A Sample Rate Reduction (by an integer factor, M)
Conversion Decimator Design

Thus, in order to avoid aliasing, x[n] should be filtered prior to down-sampling

with a lowpass filter having a cut-off ωc = π/M rad. (or π/MT s rad/sec)

The cascade of a lowpass filter with a down-sampler is called a decimator.

ω
− π 0 π
M
ω = Ω Ts
M

x[n ] Lowpass filter x d [n ] discrete continuous

gain=1 ↓M frequency
variable
frequency
variable
cutoff=π/M

Decimator
jω
X (e jω ) 2π
MTs
X d (e )
2π
Ts
... ... ...
... 4π 2π 2π 4π
ω
ω − − 0
Ts
4π 2π
− − 2π 4π
0 MT s MT s MTs MTs
Ts Ts Ts Ts Ts π
MTs
AFMH Created: Last update: DSP_L6_T7
June 1999 July 2005

A/D & D/A MATLAB Example

Conversion Decimator Operation

% Illustration of Decimation Process, M=5

% decimation process filters the input data with a
lowpass filter and then resamples the resultant Input Sequence
smoothed signal at a lower rate. 2

clf; 1
Amplitude

M = input('Down-sampling factor = '); 0

n = 0:99; -1

x = sin(2*pi*0.043*n) + sin(2*pi*0.031*n); -2
0 10 20 30 40 50 60 70 80 90 100
Time index n
y = decimate(x,M,'fir'); Output Sequence
subplot(2,1,1); stem(n,x(1:100)); 2

title('Input Sequence'); 1
Amplitude

xlabel('Time index n'); ylabel('Amplitude'); 0

subplot(2,1,2); -1

m = 0:(100/M)-1; stem(m,y(1:100/M)); -2
0 2 4 6 8 10 12 14 16 18 20

title('Output Sequence'); Time index n

xlabel('Time index n');ylabel('Amplitude');

Ref: Program P10_5, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T8
June 1999 July 2005

4
 A/D & D/A Sample Rate Increase
Conversion (by an integer factor, L)
Increasing the sampling rate by an integer factor, L
With a new sampling period, Ts = Ts / L ,the resampled signal is:
'

~
x d [ n ] = x a[ nTs' ] = xa [ nTs / L] = x[ n / L] n = 0,± L ,±2 L,...

This is done by inserting (L-1) zeroes between

each sample of x[n] - up-sampler.
L= 2 x[ 0 ] x[1 ]
x[n ] ~
xd [ n] x[ 2 ]
x[ 3 ] x[ 4 ]

n ↑L
0 1 0 L 2L 3L 4L n
Ts

The up-sampler expands the time scale by a factor of L.

In the frequency domain, the spectrum is simply scaled (compressed by factor L) in
frequency: ∞ ∞

∑ ~x [n]e ∑ x[n]e
~
X d ( e jω ) = d
− jnω
= − jnL ω
= X ( e jLω )
n= −∞ n =−∞

AFMH Created: Last update: DSP_L6_T9

June 1999 July 2005

A/D & D/A MATLAB Example

Conversion Sample Rate Increase (by an integer factor, L)

% Illustration of Up-Sampling by an Integer Factor

% Input Sequence
1
clf;
n = 0:50; 0.5
Amplitude

x = sin(2*pi*0.12*n); 0

y = zeros(1, 3*length(x)); -0.5

y([1: 3: length(y)]) = x;
-1
subplot(2,1,1) 0 5 10 15 20 25 30 35 40 45 50
Time index n
stem(n,x);
title('Input Sequence'); Output Sequence
1
xlabel('Time index n');
ylabel('Amplitude'); 0.5
Amplitude

subplot(2,1,2) 0

stem(n,y(1:length(x))); -0.5
title('Output Sequence');
-1
xlabel('Time index n'); 0 5 10 15 20 25 30 35 40 45 50
Time index n
ylabel('Amplitude');
Ref: Program P10_1, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T10
June 1999 July 2005

5
 A/D & D/A Sample Rate Increase
Conversion (by an integer factor, L)
X a ( jΩ )

In the frequency domain: − Ω0 0 Ω0 Ω

∞ ∞
jω
X (e )
∑ ~x [n]e ∑ x[n]e
~
X d ( e jω ) = d
− jnω
= − jnL ω
= X ( e jLω ) 1 / Ts
n= −∞ n =−∞ ... ...
− 2π − π 0 π 2π ω
Spectrum is scaled ~
X d ( e jω ) L=2
or compressed 1 / Ts
... ...
After up-sampling, it is necessary to ω
− 2π
− π 0 π 2π 4π
= 2π
remove the frequency scaled images of L L L L L

X a(jΩ), except those that are at integer H (e jω )

L
multiples of 2π. ... ...
2π ω
− 2π
L
− π
L
0 π
L L
4π
L
= 2π
~
This is accomplished by filtering xd [ n] X d ( e jω )
with a lowpass filter that has cutoff L/ Ts

frequency of π/L and a gain of L. In the ... ...

ω
time domain, the in-between zero- − 2π
L
− π
L
0 π
L
2π
L
4π
L
= 2π
samples are effectively interpolated. AFMH Created:
June 1999
Last update:
July 2005 DSP_L6_T11

A/D & D/A MATLAB Example

Conversion Effect of Up-Sampling in the Frequency Domain
% Effect of Up-sampling in the Frequency Domain
% Use fir2 to create a bandlimited input sequence
freq = [0 0.45 0.5 1]; mag = [0 1 0 0]; 1
Input Spectrum

x = fir2(99, freq, mag); 0.8

% Evaluate and plot the input spectrum

Magnitude

0.6

[Xz, w] = freqz(x, 1, 512); subplot(2,1,1); 0.4

plot(w/pi, abs(Xz)); axis([0 1 0 1]); grid 0.2

xlabel('\omega/¹'); ylabel('Magnitude'); 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ω/¹
title('Input Spectrum'); subplot(2,1,2);
% Generate the up-sampled sequence
Output Spectrum
L = input('Type in the up-sampling factor = 1');
y = zeros(1, L*length(x)); 0.8
Magnitude

y([1: L: length(y)]) = x; 0.6

% Evaluate and plot the output spectrum 0.4

[Yz, w] = freqz(y, 1, 512); 0.2

0
plot(w/pi, abs(Yz)); axis([0 1 0 1]); grid 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ω/¹
xlabel('\omega/¹'); ylabel('Magnitude');
title('Output Spectrum');
Ref: Program P10_3, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T12
June 1999 July 2005

6
 A/D & D/A Sample Rate Increase (by an integer factor, L)
Conversion Interpolator Design

As explained in the previous slide, the up-sampling process consists of an up-

sampler which inserts (L-1) zeros between each sample of x[n], followed by a
lowpass filter.

The cascade of an up-sampler and a lowpass filter is called an interpolator.

x[ 0 ] x[1 ] x[ 3 ] x[ 4 ]
x[ 2 ]

0 L 2L 3L 4L n

x[n ] x~d [ n ] Lowpass filter x d [n ]

↑L gain=L
cutoff=π/L

x[ 0 ] x[1 ] x[ 3 ] x[ 4 ]
x[ 2 ]
n
0 1
Ts
0 L 2L 3L 4L n

AFMH Created: Last update: DSP_L6_T13

June 1999 July 2005

A/D & D/A MATLAB Example

Conversion Interpolator Operation
% Illustration of Interpolation Process
clf;
L = input('Up-sampling factor = '); Input Sequence
2
% Generate the input sequence
n = 0:49; 1
Amplitude

x = sin(2*pi*0.043*n) + sin(2*pi*0.031*n); 0
% Generate the interpolated output sequence -1

y = interp(x,L); -2
0 5 10 15 20 25 30 35 40 45 50
% Plot the input and the output sequences Time index n

subplot(2,1,1);
stem(n,x(1:50)); 2
Output Sequence

title('Input Sequence'); 1
xlabel('Time index n'); ylabel('Amplitude');
Amplitude

0
subplot(2,1,2);
m = 0:(50*L)-1; -1

stem(m,y(1:50*L)); -2
0 50 100 150 200 250

title('Output Sequence'); Time index n

xlabel('Time index n'); ylabel('Amplitude');

Ref: Program P10_6, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T14
June 1999 July 2005

7
 A/D & D/A Fractional-Rate Sampling Rate Alteration
Conversion (by a rational factor)
The cascade of a decimator that reduces the sampling rate by a factor M with
an interpolator that increases the sampling rate by a factor L results in a
system that changes the sampling rate by a rational factor of L/M.

x[n ] Lowpass filter Lowpass filter x d [n ]

↑L gain=L gain=1 ↓M
cutoff=π/L cutoff=π/M

Because the cascade of two lowpass filters with roll-off frequencies π/M and
π/L is equivalent to a single lowpass filter with cut-off frequency

ωc = min {Mπ , πL }

The sample rate convertor may be simplified:

x[n ] Lowpass filter x d [n ]

↑L gain=L ↓M
cutoff=min{π/L , π/M }

AFMH Created: Last update: DSP_L6_T15

June 1999 July 2005

A/D & D/A

Conversion
Examples

Application Example #6

Digital audio tape (DAT) drives have a sampling frequency of 48 kHz, whereas
a compact disk (CD) player operates at a rate of 44.1 kHz. In order to record
directly from a CD onto a DAT, it is necessary to convert the sampling rate
from 44.1 to 48 kHz. Therefore, consider the following system for performing
this sample rate conversion

xCD [n] x DAT [n ]

↑L H (e jω ) ↓M

Find the smallest possible values for L and M and find the appropriate filter to
perform the conversion.

AFMH Created: Last update: DSP_L6_T16

June 1999 July 2005

8
 A/D & D/A
Conversion
Examples

Solution App. Example #6

Given that 48000=27.3.53 and 44100=22.32.52.72, to change the sampling rate

we require 7 3 5
L
= 2 .3.5 = 2 .5 = 160
M 22 .32 .52.72 3 .7 2 147

Therefore if we up-sample by L=160 and then down-sample by M=147, we

achieve the desired sample rate conversion.

The lowpass filter we require must have a cut-off frequency

ωc = min{Mπ , πL } = 160
π

and the gain of the filter, L=160.

AFMH Created: Last update: DSP_L6_T17

June 1999 July 2005

A/D & D/A MATLAB Example

Conversion Fractional-Rate Sampling Rate Alteration

% Illustration of Sampling Rate Alteration by a Ratio of Two Integers

clf; Input Sequence

L = input('Up-sampling factor = '); 2

M = input('Down-sampling factor = '); 1

Amplitude

n = 0:29; 0

-1
x = sin(2*pi*0.43*n) + sin(2*pi*0.31*n);
-2
y = resample(x,L,M); subplot(2,1,1); 0 5 10 15 20 25
Time index n
stem(n,x(1:30));axis([0 29 -2.2 2.2]);
title('Input Sequence'); Output Sequence
2
xlabel('Time index n'); 1
Amplitude

ylabel('Amplitude'); 0

subplot(2,1,2); -1

m = 0:(30*L/M)-1; -2
0 5 10 15 20 25 30 35 40 45
stem(m,y(1:30*L/M));axis([0 (30*L/M)-1 -2.2 2.2]); Time index n

title('Output Sequence');
xlabel('Time index n'); ylabel('Amplitude');
Ref: Program P10_7, DSP Lab using Matlab, Mitra AFMH Created: Last update: DSP_L6_T18
June 1999 July 2005

9
 A/D & D/A
Conversion
Examples

Application Example #7
Suppose that we are required to slow a segment of speech to one-half its
normal speed. The speech signal sa(t) is assumed to have no energy outside of
5 kHz, and its sampled at a rate of 10 kHz, yielding the sequence
s[n] = sa[nT s].
The following system is proposed to create the slowed-down speech signal.
s a (t) s[n] w[n] Interpolation y[n] ya (t )
A/D ↑2 Filter
D/A

S a ( j Ω)
Ts Ts
A
Assume that S a(jΩ) is as shown:
− 10000π Ω
10000π

(a) Determine the spectrum of w[n].

(b) Suppose that the d-t filter is described by the difference equation
y[n] = w[n] + 1/2{w[n-1] + w[n+1]}.
Find the frequency response of the filter and describe its effect on w[n].
(c) What is Ya(jΩ) in terms of S a(jΩ)? Does ya(t) correspond to slow-down
speech? AFMH Created: Last update: DSP_L6_T19
June 1999 July 2005

A/D & D/A

Conversion
Examples
S( e j ω )
Solution App. Example #7
A / Ts
(a) Since sa(t) is sampled at the
... ...
Nyquist rate, the DTFT of the ω
− 2π −π π 2π
sampled speech signal, s[n] is: W(e )jω

Up-sampling by a factor of 2 scales the A / Ts

frequency axis of S(e jw) by a factor two: ... ...
ω
−π −π π π
(b) The unit sample response of the d-t filter is: 2 2

h[n]=1/2 δ[n+1] + δ[n] + 1/2 δ[n-1]

which has a frequency H(e jω ) = 1 + cos ω

To see the effect of this filter on w[n], note that due to the up-sampling,
w[n] = 0 for n odd. Therefore with y[n] = w[n] + 1/2 w[n-1] + 1/2 w[n+1]
it follows that  w[n] n even
y[n ] =  1
 2 w[ n − 1] + 2 w[n + 1]
1
n odd
Thus, the even-index values of w[n] are unchanged, and the odd-index values are the
average of the two neighbouring values, i.e. h[n] performs a linear interpolation
between the values of w[n]. AFMH Created: Last update: DSP_L6_T20
June 1999 July 2005

10
 A/D & D/A
Conversion
Examples

Solution App. Example #7

(c) The output of the D/A convertor, ya(t) has a Fourier transform W(e jω )

T Y (e ) | Ω |< π / Ts
jΩTs A/ Ts
Ya ( jΩ) =  s ... ...
ω
 0 otherwise −π −π π
2
π
2
Since Y ( e ) = H (e jω )W (e jω ) = (1 + cos ω )W (e j ω )
jω (1 + cosω )

and W (e j ω ) = S (e j 2 ω )
−π −π π π ω
then T (1 + cos ΩTs )S ( e j 2 ΩTs ) | Ω |< 10000 π 2 2

Ya ( jΩ) =  s
 0 otherwise
which is the product of (1+cos ΩT s) and T s S(ej2ΩTs ) as illustrated:
Thus, ya(t) does not correspond to slowed-down Ya ( j Ω)
speech due to the images of sa(t) that occur 2A
in the frequency range 5000π<|W|<10000π and
the nonideal linear interpolator. Ω
−10000 π − 5000 π 5000 π 10000 π
Note a better approximation - use a DC convertor
with sampling rate of 2T s to eliminate images. AFMH Created: Last update: DSP_L6_T21
June 1999 July 2005

11