2014 exam worked solutions

Monday, 16 November 2015 11:23 AM

ELEC4620 2014 Exam

Page 2
Monday, 16 November 2015 11:23 AM

A square wave us made up of infinitely many sinusoids, so to

Theoretically sample it, you need an infinite sampling frequency

Reason 'c' is a better answer, is that while 16kHz will theoretically give
a higher resolution, the extra bandwidth and computational power
is wasted as 12kHz will easily Reconstruct the signal.

ELEC4620 2014 Exam

Page 3
Monday, 16 November 2015 11:23 AM

What is this even supposed to mean….

More often than not, this is false
Smoother -> wider mainlobe
This would be 'most' correct

ELEC4620 2014 Exam

Page 4
Monday, 16 November 2015 11:23 AM

ELEC4620 2014 Exam

Page 5
Monday, 16 November 2015 11:23 AM

ELEC4620 2014 Exam

Page 6
Monday, 16 November 2015 11:23 AM

The number of poles should equal

Number of zeros. 13 is not an option so (E)

Order = number of zeros = 13

When there is a zero located on the unit circle, you know the
response will be driven to zero. Each set of conjugate pairs only
count for 1 time the response is driven to zero, so total number
of times is:
5*[conjugate pairs] + 1*[the zero located at (-1,0)] = 6
The zeros that don’t fall on the unit circle are not counted. While they
drive the magnitude response down, they do not force it to be zero.

ELEC4620 2014 Exam

Page 7
Monday, 16 November 2015 11:23 AM

Quote from lecture slides referring to FIR

P-McC design method:
"This method designs an optimal linear
phase filter directly from the design
specifications."

Points on the unit circle are reciprocals of themselves and the wording can be
open to interpretation (ie: they are always called conjugate pairs), so I think it's
best to go with the 100% true answer of (c)

ELEC4620 2014 Exam

Page 8
Monday, 16 November 2015 11:23 AM

The reason for Blackman being the worst choice is that it has the widest
mainlobe. Because the signals are equal magnitudes and close together, we
would need a sharp, small mainlobe width (this is the key factor due to close spacing).
Rectangular window would be best choice in this case.

ELEC4620 2014 Exam

Page 9
Monday, 16 November 2015 11:23 AM

Do to the low attenuation of the rectangular window, if the

Signal is not incredibly dominant, it will be over powered by
The noise floor due to the flatness of the sidelobe rolloff. The
Best choice would be blackman due to the large rolloff. This
Would make a low amplitude signal apparent when spaced
Far away.

I think I'd say rectangular is the worst. Due to the fact

that even if you do pin point the low amplitude
frequency, due to the higher noise floor, the signal will
be choppy as hell. Blackman would be better in that you
could increase the filter order to create a narrower
mainlobe, but you can't do that for a rectangular as the
high noise will always be there.

All but rectangular have the possibility to go beyond 40dB in

the stop band sue to sideband roll off.

Both have the potential to

have stopband past 60dB.
Just need to design the filter
properly.

ELEC4620 2014 Exam

Page 10
Monday, 16 November 2015 11:23 AM

Table from lecture notes

Rectangular has the 'sharpest' main lobe, meaning

the steepest or narrowest transition band. And also
according to this textbook extract Pg361 table 9-2
http://www.utdallas.edu/~raja1/EE4361%20Spring%
2014/Lecture%20Notes/FIR%20Windows.pdf

On slide 27 in the DFT slides it says that constant

geometry is easier to implement in hardware. And when
listening to the lecture he mentions it as good for
hardware, so I would probably bet on B. But then again,
it depends what is meant by "most suitable".

Good-Thomas is least suitable. Because N

cannot be factored into two relatively prime
factors nm such that N = nm. This is because N
factors to 2^8, so any factor of N will be
factorable by 2 and cannot be relatively prime
with any other factor of N.

ELEC4620 2014 Exam

Page 11
Monday, 16 November 2015 11:23 AM

Constant geometry is not an in place algorithm, as it uses two

arrays and 'ping pongs' data between the two. Where as the
other algorithms uses the single arrays and updates data as it
goes.

In lecture 10 - slide 29 it says that the radix-4 uses 50%

fewer memory calls, so A is false.

Purely real DFTs exhibit the property of DFT symmetry.

Even or odd length does not matter and neither does
the first term. As long as the circular symmetry is met.
Suppose the DFT vector was length 7 [4 5 -6 7 7 -6 5]:

Purely imaginary DFTs must have the first DC

term = 0, and be asymmetric

ELEC4620 2014 Exam

Page 12
Monday, 16 November 2015 11:23 AM

The function form used in b, c and d can be correctly used,

but each fft will require padding of length 7 which is the
total N-1 in this case. The conv function only requires the
two vectors as input and automatically zero pads. Two
correct answers would be
-> conv([1 2 3 4], [5 6 7 8])
-> ifft(fft[1 2 3 4], 7).*fft([5 6 7 8],7))

Correct term would have been:

ifft(fft([1 2 3 4 ]).*fft([5 6 7 8]))

ELEC4620 2014 Exam

Page 13
Monday, 16 November 2015 11:23 AM

(I) If impulse is real, than the impulse response is symmetric

(II) Since impulse is real, there are no sin terms since they hold
the complex term, so a sum of cosines is possible.
(III) same deal as linear phase filters.

ELEC4620 2014 Exam

Page 14
Monday, 16 November 2015 11:23 AM

http://www.ti.com/lit/an/spra291/spra291.pdf
(a) see section 3.1
(b) see section 3.2

There are two approximation methods for n bit length, signal to

noise ratio:
- Fixed point arithmetic: 5dB/bit
- Quantization: 6dB/bit
So in this case we use 5dB*16bits = 80dB out of band.

ELEC4620 2014 Exam

Page 15
Monday, 16 November 2015 11:23 AM

If sampling frequency is 24kHz, then every second

there are 24k data points. If a normal FIR filter, then it
is a convolution operation of sample data and filter, an
O(NM) operation. Hence, multiplications = 872*24k =
20.9 Million.

ELEC4620 2014 Exam

Page 16
Monday, 16 November 2015 11:23 AM

- Sampling can be lower than twice the highest audio signal due to Nyquist. Hence 20kHz is un usable.
- 100kHz is unreasonable as the signal will be oversampled, wasting computational power, create large digital files and taking longer to perform
operations on the signals.
- 44kHz would be chosen such that the Nyquist criteria is met and the signal is not oversampled so much that computations are wasted.

The data has to be compressed and digitally stored on the physical media, making production cheaper and much more efficient.
So when the audio signal needs to be played, up sampling is used to generate the original signal. Processes such as SINC
interpolation can be used when the signal has been up sampled to restore the original 44.1kHz sample frequency.

From Lecture slides: The dither signal improves the SNR by breaking up the spurious line spectra that may occur. The dither has
the greatest effect on small signals near the quantization noise floor. It can add about 12 dB to the dynamic range and signals
below the noise floor can be represented. Although a quantizer is non-linear, dither makes it behave linearly.

Oversampling, to spread the images in the frequency domain and make for less reliance on expensive
analogue filtering as it relies on digital implementation. Due to the image spreading

ELEC4620 2014 Exam