Scribd
Upload
159 views9 pages

Sampling Rate Conversion

Upsampling and downsampling are processes to convert the sampling rate of a digital signal. Upsampling increases the rate by inserting zeros or interpolating between samples, while downsampling decreases the rate by selecting every other sample. To avoid aliasing when changing the rate, a low-pass filter is used to remove high frequency content above the new Nyquist frequency. Practically, linear phase filters provide real-time processing but less sharp outputs, while ARMA filters offer sharper results but require offline processing.

Uploaded by

Tahmid Hassan Talukdar
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
159 views9 pages

Sampling Rate Conversion

Upsampling and downsampling are processes to convert the sampling rate of a digital signal. Upsampling increases the rate by inserting zeros or interpolating between samples, while downsampling decreases the rate by selecting every other sample. To avoid aliasing when changing the rate, a low-pass filter is used to remove high frequency content above the new Nyquist frequency. Practically, linear phase filters provide real-time processing but less sharp outputs, while ARMA filters offer sharper results but require offline processing.

Uploaded by

Tahmid Hassan Talukdar
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

INTRODUCTION TO UPSAMPLING &

DOWNSAMPLING
What is Sampling Rate Conversion?
As the name suggests, the process of converting the sampling rate of a
digital signal from one rate to another is Sampling Rate Conversion.
Increasing the rate of already sampled signal is Upsampling whereas
decreasing the rate is called downsampling.
Why to do it?
Many practical applications require to transmit and receive digital
signals with different sampling rates. So at many stages in application we
need to do sampling rate conversion.
How to do it?
There are two ways in which we can achieve sampling rate
conversion:
1. First approach is to do D/A conversion to recover back original
analog signal. Then we can do A/D conversion with desired
sampling rate. The advantage of this technique is that the second
sampling rate need not hold any special relationship with old one.
But the problem is signal distortion introduced by D/A and
quantization effects of A/D.
2. Second approach is to work in digital domain only. So well have
to predict the digital signal with desired rate using the already
sampled signal in hand.

Now for this lecture, well look at the second choice of sampling rate
conversion.

UPSAMPLING
Lets consider, simplest case of upsampling. We want to double the
sampling rate of signal. So what we do is insert 0s in between two
successive samples. As shown:

Obviously this is a bad approach. As we dont have data for


intermediate samples, lets generate it.
Method-1: Repetition
Repeat the current sample.
The corresponding filter kernel will be

The output waveform will be:

Method-2: Interpolation
The intermediate sample value will be the average of its
neighboring values.
The filter kernel will be:

The output waveform will be:

Now this a very good method but it produces, two aliases of each
frequency in frequency domain . So we should cut the high frequency
contents to avoid aliasing.
Why to cut?

Because they dont contain any new information. They are just
repeating information. We also know that maximum frequency
content in original signal cannot be greater than Fs/2 so there is
definitely no more information in baseband. So we can afford to cut
high frequency contents.
How to cut it?
Use a perfect low pass filter. That is we will keep slow moving
spirals and reject fast moving spirals in frequency domain.

so filter kernel in frequency domain is as shown. Its amplitude is 1 for


frequencies in the range - /2 to +/2 and zero for rest all frequencies. So we
cut the high frequency aliases. So filter kernel in frequency domain is set of
slow moving spirals having amplitude 1.
Lets figure out, what will be the corresponding filter kernel time
domain.
/2

e0 d = /2 (/2)=

/ 2

/2

ej d = 1/j(ej/2 - e-j/2 )= 2

/ 2

/2

e2j d = 0

/ 2

/2
/ 2

e3j d = 2/3

Thus the perfect interpolator is an Infinite Impulse Response (IIR) filter. The
filter is not causal hence cannot model using ARMA. We cannot implement
the filter because its infinite.
There is a class of functions called Analytic functions. According to
Taylor, all the information in the analytical function is at zero, so you dont
have to go far. We can express these functions as Lauriant series and model
them. But for this, the function should be continuous at all points. But our
perfect interpolator filter kernel is discontinuous at - /2 and /2.
One interesting observation is that, we are getting zeros at previously
sampled points in the filter kernel.
So to do upsampling faster we can use Multi-channel Polyphase
Multi-sampler with two filter banks.
1. 1st filter bank doing nothing (corresponding to already sampled
points).
2. 2nd filter bank with half sample delay.

UPSAMPLING BY 4
Here the filter kernel will look like this:

In this case, well have to predict three new samples, between already
present pair of samples. We now will have 4-filter banks.
1. Bank with 0 sample delay.
2. Bank with 1/4 sample delay.
3. Bank with 1/2 sample delay.
4. Bank with 3/4 sample delay.

PRACTICAL DESIGN OF FILTER:


Use of Linear Phase Filters:
For practical design of upsampler filters, we use delayed symmetric
filters.
Zero phase filters are symmetric about zero. Delayed symmetric filters
have symmetry along a shifted point. They are also called as Linear phase
filters. The advantage of this approach is the filtering can happen at real
time. But disadvantage is that the output is not sharp.
Use of ARMA filters:
Using ARMA filters, we can get sharper output. But we cannot get
real time processing using ARMA filters.
For offline processing, we can use two ARMA filters, one Causal and
second Anti-causal filter. Then we just add the outputs and well have
desired result. The filter kernel of Causal and Anti-causal filters will be just
the flips of one another.

DOWNSAMPLING:
As said, Downsampling is decreasing the sampling rate of a signal.
Lets consider a simple case of downsampling a signal to half of its original
sampling rate.
Simplest way to do this is to forget every other sample and well have
the desired sampling rate.

But if we reduce the sampling rate just by selecting every other sample of
x(n), the resulting with folding frequency Fs/2 . The frequency spectrum
tries to spread up but it cannot do so. Hence it winds up on itself.

To avoid aliasing, we must first reduce the bandwidth of x(n) to max =


/2. So we should cut out high frequency contents to avoid aliasing.

You might also like