EE-421

Digital Signal Processing

Complex Engineering Problem
Submitted To:
Ms. Nazo Haroon
Submitted By:
Attia Ali
Reg. No:
UW-21M-EE-BSC-006
Date of Submission:
29th January, 2025

Department of Electrical Engineering

 Wah Engineering College. (Wah Cantt).

Part-(1)

The sampling frequency (also known as the sampling rate), denoted as FsF_s, is a critical
parameter in digital signal processing, representing the number of samples taken per second
from a continuous signal to convert it into a discrete signal. For the noisy audio signal in this
case, the sampling frequency is obtained directly when reading the audio file using the audioread
function in MATLAB.

Here’s how the sampling frequency is determined:

1. Loading the Audio Signal: The audioread function is used to read the audio file (in this
case, AUDIO.mp3). The function returns two outputs:
o audioData: The actual sampled audio signal.
o Fs: The sampling frequency (in Hz), which is a scalar value representing how
many times per second the audio was sampled during recording or encoding.

The line of code used to load the audio file and retrieve the sampling frequency is:

[audioData, Fs] = audioread('AUDIO.mp3');

2. Understanding the Sampling Frequency: The sampling frequency, Fs, defines the
temporal resolution of the audio signal. It indicates how often the continuous waveform
of the audio is sampled to generate a discrete signal. A typical value for audio signals is
44,100 Hz (standard CD quality), but it can vary depending on the file format and how
the audio was recorded or encoded.
3. Output and Display: After loading the audio, the sampling frequency is displayed using
the disp function to confirm the sampling rate used for the audio file:
4. disp(['Sampling Frequency (Fs): ', num2str(Fs), ' Hz']);

In summary, the sampling frequency FsF_s of the noisy audio signal is found directly from the
metadata of the audio file when it is read using MATLAB’s audioread function. This is an
essential step, as it provides the necessary information for performing frequency-domain
analysis and designing appropriate digital filters to remove noise.

Part-(2)

The selection of the sampling frequency (FsF_s) plays a crucial role in ensuring the accurate
representation of an analog signal in its discrete form. In the context of audio signal processing,
the goal is to sample the continuous audio signal at a sufficiently high rate so that it can be
reconstructed accurately without losing critical information or introducing distortions, such as
aliasing.
Intuition Behind the Selection of FsF_s

The sampling frequency should be chosen based on the Nyquist-Shannon Sampling Theorem,
which provides the theoretical foundation for determining an appropriate sampling rate.
According to this theorem:

 The sampling frequency (FsF_s) must be at least twice the highest frequency present in
the continuous signal, known as the Nyquist rate, in order to avoid aliasing.

Where fmaxf_{\text{max}} is the maximum frequency present in the analog signal.

This requirement ensures that the discrete signal retains all the relevant frequency components of
the original signal.

Avoiding Aliasing in Audio Signals

In the case of audio signals, the maximum frequency of interest is typically the human audible
range, which spans from approximately 20 Hz to 20,000 Hz (20 kHz). However, practical
considerations suggest that audio signals are often low-pass filtered before sampling to remove
frequencies higher than the audible range, further reducing the risk of aliasing.

1. Human Hearing Range: The upper limit of human hearing is around 20 kHz. If the
sampling frequency were chosen to be just slightly higher than 20 kHz (say 22 kHz), it
would technically be sufficient to capture all audible frequencies, but it may still lead to
practical limitations in representing the finer details of the audio signal.
2. Industry Standard Sampling Rates: The most commonly used sampling frequencies
for audio signals are:
o 44,100 Hz (CD quality): This is a standard used for audio CDs, providing a
sampling rate that is more than twice the maximum frequency of human hearing
(i.e., it satisfies the Nyquist criterion). The extra margin above 40 kHz allows for
some filtering of the higher frequencies before sampling.
o 48,000 Hz: Commonly used in professional audio and video equipment.
o 96,000 Hz and 192,000 Hz: These higher sampling rates are used in high-
resolution audio formats for capturing ultra-high frequency details.

Thus, selecting a sampling frequency of 44,100 Hz or higher ensures that the signal can be
accurately digitized without aliasing, as this rate comfortably exceeds twice the maximum
audible frequency (i.e., 40 kHz).
Practical Considerations to Avoid Aliasing

1. Anti-Aliasing Filter: In real-world audio applications, an anti-aliasing filter is often

applied before the analog-to-digital conversion (ADC) process to attenuate frequencies
above the Nyquist frequency (Fs/2F_s / 2) that could cause aliasing. The presence of this
filter ensures that only the desired frequency components below the Nyquist frequency
are sampled.
2. Margin for Imperfections: While a sampling rate of 44.1 kHz is sufficient in theory, a
small margin (e.g., 1.1 times the maximum frequency) is often used in practice to
account for imperfections in both the anti-aliasing filter and the ADC process itself. This
margin helps to ensure that even with practical limitations, the audio signal is sampled
correctly and without aliasing.

Part-(3)

In the context of frequency domain analysis of the noisy audio signal, the Discrete Fourier
Transform (DFT) is used to convert the time-domain signal into the frequency domain,
allowing for the analysis of the frequency components of the signal. The DFT computation
requires specifying a value for NN, which refers to the number of points in the DFT
calculation.

Why and What Value of N Was Chosen?

In the provided code, the DFT is computed using a value of N=1024N = 1024, which is a power
of 2. The choice of NN is influenced by several factors:

1. Computational Efficiency: When NN is a power of 2 (e.g., 1024), the Fast Fourier

Transform (FFT) algorithm can be used, which is a more efficient implementation of
the DFT. FFT algorithms operate at O(Nlog⁡N)O(N \log N) time complexity, making the
computation faster compared to a direct DFT implementation, which operates at
O(N2)O(N^2).
2. Resolution of the Frequency Spectrum: The value of NN determines the frequency
resolution of the FFT. The higher the value of NN, the more frequency bins will be
available for resolution. With N=1024N = 1024, the frequency resolution is relatively
high, allowing for finer details in the frequency spectrum. This is particularly useful for
noise detection and filtering tasks, where identifying distinct frequency components is
important.
3. Compatibility with Signal Length: N=1024N = 1024 is chosen to balance
computational efficiency and the resolution of the frequency spectrum. It’s a standard
value used in many signal processing applications for audio signals. The choice of
N=1024N = 1024 implies that the signal will be padded with zeros (if the signal length is
smaller than 1024) or truncated (if the signal length exceeds 1024).
Implications of Making NN Smaller or Larger than the Input Signal Data Length

1. Smaller N: If N is smaller than the input signal length, the following implications arise:
o Frequency Resolution Decreases: A smaller NN leads to fewer frequency bins
in the resulting frequency spectrum. This reduces the ability to distinguish
between closely spaced frequencies, making the analysis less precise. In audio
signals, this could lead to missed identification of certain frequency components
or the inability to resolve finer details.
o Faster Computation: The computation will be faster because fewer points need
to be processed. However, this comes at the cost of reduced frequency resolution.

2. Larger N: If N is larger than the input signal length, the following implications arise:
o Zero Padding: When NN exceeds the signal length, the signal is zero-padded
(i.e., additional zeros are added to the end of the signal). This increases the length
of the signal without adding any new information but provides a finer frequency
resolution in the frequency domain. While this can improve the accuracy of
frequency representation, it does not provide more detail about the actual signal
content.
o Higher Computational Load: A larger NN requires more computational
resources and time for the FFT calculation. This could lead to inefficiency,
especially when working with long audio signals or when real-time processing is
required.

Part-(4)

The Discrete Fourier Transform (DFT) is a key mathematical tool in signal processing that
converts a time-domain signal into the frequency domain. However, directly computing the DFT
can be computationally expensive, especially for large signals, as its complexity is
O(N2)O(N^2), where NN is the number of data points. In practical applications, this high
computational cost can significantly slow down processing, particularly when working with
large datasets like audio signals.

To mitigate this issue, the Fast Fourier Transform (FFT) algorithm is employed. The FFT is
an optimized version of the DFT that reduces the computational complexity from O(N2)O(N^2)
to O(Nlog⁡N)O(N \log N), making it much more efficient for large datasets. MATLAB provides
an efficient implementation of the FFT algorithm through the fft function.

Justification for Using FFT Algorithm in MATLAB

1. Computational Efficiency:
o The primary reason for using the FFT algorithm is its faster computation. For
large datasets, the standard DFT would take considerably longer to compute due
to its quadratic time complexity (O(N2)O(N^2)). In contrast, the FFT reduces this
to O(Nlog⁡N)O(N \log N), which allows for much faster computation, particularly
when processing long audio signals.
 o In practical applications, such as real-time audio processing, the efficiency of the
FFT is critical for maintaining low-latency and real-time performance.

2. Built-in MATLAB Support:

o MATLAB has a highly optimized fft function that is tailored for fast execution.
This built-in function is specifically designed to handle large datasets efficiently,
providing an out-of-the-box solution for frequency domain analysis without
needing the user to implement their own DFT algorithm.
o MATLAB's FFT implementation is optimized for both speed and accuracy,
leveraging hardware-specific optimizations for faster execution on modern
processors.

3. Handling Large Datasets:

o Audio signals, such as the one in this project, can have thousands or even
millions of samples. Computing the DFT on such large datasets would be
computationally impractical without the FFT. The FFT allows us to perform
frequency domain analysis even on long audio signals in a reasonable amount of
time.

4. Memory Efficiency:
o The FFT algorithm is not only faster but also more memory-efficient. The
standard DFT involves computing all pairwise multiplications of the signal,
which can be memory-intensive for large datasets. The FFT algorithm reduces
this by breaking the problem into smaller subproblems, using a divide-and-
conquer approach, which results in lower memory consumption.

5. Accuracy of Frequency Components:

o The FFT does not sacrifice accuracy in frequency component extraction. It
produces the same results as the DFT, but with much less computational
overhead. This allows for accurate frequency domain analysis, which is essential
for identifying and filtering noise in audio signals.

Part-(5)

The process to identify the noise frequencies involved the following steps:

1. FFT Computation: The FFT of the noisy audio signal was computed using the fft
function with N=1024N = 1024. This transformed the audio signal from the time domain
to the frequency domain, providing the frequency components and their corresponding
magnitudes.
2. Magnitude Calculation: The magnitude of the FFT was calculated using
abs(audioFFT), which gives the amplitude of each frequency component.
3. Thresholding for Noise Detection: A threshold was set to identify the frequencies
associated with noise. Specifically, the threshold was defined as 10% of the maximum
 magnitude in the frequency spectrum (threshold = max(magnitude) * 0.1). This was done
to isolate the significant peaks that were likely to correspond to noise.
4. Noise Frequency Identification: The frequencies where the magnitude exceeded the
threshold were considered noise. The find function was used to detect the indices where
the magnitude surpassed this threshold, and the corresponding frequencies were extracted
from the frequency axis f.
5. Displaying Noise Frequencies: The identified noise frequencies were then displayed in
the output using disp(noiseFreqs).

This approach effectively isolates the frequencies that contribute to noise in the signal, allowing
for targeted filtering in subsequent steps.

Part-(6)

In my code, a Low-pass filter was chosen to remove the noise from the noisy audio signal.
Based on the frequency analysis, the noise was found to be primarily in the higher frequency
range, while the desired signal occupies lower frequencies. A Low-pass filter allows the lower
frequency components (which represent the desired audio) to pass through, while attenuating the
higher frequency noise.

This approach is effective for removing the high-frequency noise and preserving the original
audio signal, which is typically in the lower frequency range. Therefore, a Low-pass filter was
selected as the most appropriate filtering technique for this purpos.

Part-(7)

In my filter design, I chose a Low-pass filter because the noise in the audio signal was
predominantly in the higher frequency range, while the desired audio signal resides in the
lower frequency range. The intuition behind selecting a Low-pass filter is straightforward: it
will allow the lower frequencies, which contain the important components of the audio signal, to
pass through while blocking the higher frequencies, where noise tends to occur.

For the design method, I used both FIR (Finite Impulse Response) and IIR (Infinite Impulse
Response) filters. The intuition behind this dual approach was to explore different filter
characteristics and compare their effectiveness in noise removal.

1. FIR Filter: The FIR filter was chosen because it provides a linear phase response,
ensuring no phase distortion of the audio signal. This is especially important when
dealing with signals where phase relationships need to be preserved, such as in audio
processing. The fir1 function was used to design the filter with a specified cutoff
frequency, which controls the transition between passing and attenuating frequencies.
2. IIR Filter: The IIR filter was selected because it is computationally efficient and
requires fewer coefficients to achieve the same level of filtering as the FIR filter. The
butter function was used to design a Butterworth IIR filter, which is known for its
 maximally flat response in the passband and smooth roll-off in the stopband, making it
a good choice for noise reduction.

By using both filter types, I was able to compare their performance in removing high-frequency
noise and decide on the most suitable approach for the given application.

Part-(8)

For the given CEP involving noise removal from an audio signal, the FIR filter is more
suitable. The primary reasons for this choice are:

1. Linear Phase Response: One of the key advantages of FIR filters is their linear phase
response. This is important in audio processing, as it ensures that all frequency
components of the signal are delayed by the same amount, preventing any phase
distortion. In audio applications, phase distortion can lead to undesirable effects such as
smearing or warping of the sound, which is not ideal.
2. Stability: FIR filters are always stable, regardless of their order. In contrast, IIR filters
can potentially become unstable for higher orders, especially if the design parameters are
not carefully chosen. Stability is critical in real-time applications, where signal
processing needs to be consistent and reliable.
3. Precision: FIR filters are more suitable when precise frequency control is needed, as they
can be designed to have sharp cutoffs and more control over the frequency response.
This is particularly useful in the context of removing noise, where the goal is to isolate
and suppress unwanted frequencies effectively.

While IIR filters are more computationally efficient and require fewer coefficients to achieve
similar performance, the advantages of FIR filters in terms of stability, linear phase response,
and precision make them more suitable for this application. Therefore, FIR filters are preferred
for noise removal in audio signal processing.

Part-(9)
FIR Filters:

Advantages:

 Linear Phase Response: FIR filters provide a linear phase response, which ensures
that all frequency components of the signal are delayed equally, preserving the shape of
the waveform. This is important for audio applications where phase distortion is
undesirable.
  Stability: FIR filters are always stable regardless of their order, making them
predictable and reliable for real-time applications.
 Design Flexibility: FIR filters can be designed with a precise cutoff frequency, which
provides more control over the frequency response.

Disadvantages:

 Higher Computational Cost: FIR filters generally require more coefficients (taps) to
achieve a sharp cutoff, especially for filters with a steep roll-off. This can lead to a higher
computational burden.
 Higher Memory Usage: Due to the need for more coefficients, FIR filters tend to use
more memory than IIR filters.

IIR Filters:

Advantages:

 Computational Efficiency: IIR filters are computationally more efficient than FIR
filters because they require fewer coefficients to achieve similar performance. This leads
to a lower computational cost, making them suitable for systems with limited resources.
 Lower Memory Usage: IIR filters require fewer coefficients, thus consuming less
memory compared to FIR filters.
 Sharper Roll-Off: IIR filters can achieve a sharper roll-off in the frequency response
with fewer coefficients, making them more efficient in terms of the trade-off between
filter performance and resource usage.

Disadvantages:

 Nonlinear Phase Response: IIR filters generally have a nonlinear phase response,
which can introduce phase distortion in the signal. This can affect the quality of the audio
signal in applications where phase integrity is critical.
 Potential Instability: IIR filters are prone to instability, particularly as their order
increases. This requires careful design and validation to ensure that the filter remains
stable in all conditions.

Computational Efficiency:

IIR filters are generally more computationally efficient than FIR filters. This is because IIR
filters achieve similar performance with fewer coefficients, resulting in lower computational
complexity and memory usage. However, the choice between FIR and IIR depends on the
specific requirements of the application, such as the need for linear phase response or the
acceptable level of phase distortion.
 Part-(10)

In the given code, the pole-zero plots of both the FIR and IIR filters exhibit specific
symmetries, and here's a breakdown based on the filter design:

FIR Filter:

 Symmetry: The pole-zero plot of the FIR filter will show only zeros and no poles,
since FIR filters have no feedback (non-recursive). In the plot, the zeros are
symmetrically placed on the real axis or possibly along the unit circle in the complex
plane, depending on the filter’s order and design. This symmetry ensures that the FIR
filter maintains a linear phase response.
 Zero Placement: The zeros are typically distributed symmetrically, which is a result of
the filter's design, as FIR filters are often designed for a flat or linear phase response.

IIR Filter:

 Symmetry: The pole-zero plot of the IIR filter shows both poles and zeros. The poles
are typically placed inside the unit circle for stability, and the zeros may appear outside
or on the unit circle, depending on the filter's design. The plot exhibits bilateral
symmetry about the real axis, with the poles and zeros appearing in conjugate pairs.
 Pole and Zero Placement: The IIR filter's poles are crucial for determining the filter's
behavior and stability. If the poles are placed too close to the unit circle or outside it, the
filter may become unstable.

In summary, both FIR and IIR filters in the given code exhibit symmetry:

 FIR filter: Symmetry in zeros (placed along the unit circle or real axis, with no poles).
 IIR filter: Symmetry in both poles and zeros (poles and zeros appear in conjugate pairs,
with poles inside the unit circle for stability).

Code

Outputs