MSP Homework 1 – 2025

March 2025

Please include your ID number in your submission.

Question 1
A real-time continuous signal xc (t) is band-limited to frequencies below 5 kHz, meaning

Xc (jΩ) = 0 for |Ω| ≥ 2π · 5000.

This signal is sampled at a rate of 10,000 samples per second to produce a discrete-time
sequence:
x[n] = xc (nT ), with T = 10−4 .

Let X[k] be the 1000-point DFT of x[n].

1. Which continuous-time frequency does the index k = 150 in X[k] correspond to?

2. Which continuous-time frequency does the index k = 800 in X[k] correspond to?

1
Question 2
Show that the Short-Time Fourier Transform (STFT), defined as
∞
X
X[n, λ) = x[n + m]w[m]e−jλm
m=−∞

has the following properties:

1) Linearity: If x[n] = ax1 [n] + bx2 [n], then

X[n, λ) = aX1 [n, λ) + bX2 [n, λ).

2) Shifting: If y[n] = x[n − n0 ], then

Y [n, λ) = X[n − n0 , λ).

3) Modulation: If y[n] = ejω0 n x[n], then

Y [n, λ) = ejω0 n X[n, λ − ω0 ).

4) Conjugate Symmetry: If x[n] is real, then

X[n, λ) = X[n, −λ).

2
Question 3
Given the signal: h  πn i2
x[n] = sin u[n]
2
where the STFT is computed using the window:

1, 0 ≤ n ≤ 13
w[n] =
0, otherwise

Define:  
2πk
X[n, k] = X n, , 0 ≤ k ≤ 6.
7
1. Determine X[0, k] for 0 ≤ k ≤ 6.

2. Evaluate 6k=0 X[n, k] for 0 ≤ n < ∞.

Question 4
Previously, we defined the time-dependent Fourier transform of the signal x[m] such that,
for fixed n, it is the DTFT of the sequence x[n + m]w[m], where w[m] is a window function.
Now, define the time-dependent autocorrelation function as:
Z π
1
c[n, m] = |X[n, λ)|2 ejλm dλ
2π −π

This function satisfies the following:

a) Show that if x[n] is real, then

i.e., for fixed n, c[n, m] is the aperiodic autocorrelation of the sequence x[n + r]w[r].

b) Show that c[n, m] is an even function of m for fixed n, and derive the equivalent
expression:
X∞
c[n, m] = x[r]x[r − m]hm [n − r]
r=−∞

where
hm [r] = w[−r]w[−(m + r)].

3
Question 5 (From past exam)
We sample the note A2 (a pitch frequency of 110 Hz with 3 overtones) at a rate of 550
samples per second, and then perform a length-N DFT (with a square window).
(A.) If N = 5, which frequency bins have values that are non-zero? (comment: bin #1
is second after bin #0.)
(B.) If N = 15, which frequency bins have values that are non-zero?
(C.) If N = 16, which frequency bins have values that are non-zero?
(D.) If N = 32 = 16 signal samples + 16 zero-padding samples, is the result identical to
the N=16 case?
(E.) If N = 30 = 15 signal samples + 15 zero-padding samples, is the result identical to
the N=30 case?

4
Pure Data Assignment: Melodic Sequencer
Two synthesis methods were discussed in the recitation: additive and subtractive. The
subtractive synthesizer was shown in use with a sequencer. In this assignment, you’ll design
a melodic sequencer using additive synthesis.

Goal: Use a sequencer to trigger additive synthesis to create a melody.

• The sequencer could send bangs to multiple synths, each tuned to a specific pitch.

• The sequencer could control the pitch of a single synth dynamically.

Instructions:

• Create your patch in Pure Data.

• Submit your .pd files along with your homework.

• The choice of melody is entirely up to you.

If you’d like a suggested sequence to start with, try the following A-minor arpeggio
(frequencies in Hz):

A (440), C (523), E (659), C (523), A (440), C (523), E (659), C (523)

Experimenting with pitch and the pattern is highly encouraged.

5
The First Python Assignment: STFT Trade-Off Explo-
ration
In this assignment, we will explore the frequency-time resolution trade-off inherent in
the Short-Time Fourier Transform (STFT).
Each student will receive a unique signal provided as a CSV file. This signal is composed
of five elementary parts:

• Two impulses

• Two sinusoidal components

• One “interference” sinusoid

Objective: Identify the frequencies of the sinusoidal components and the time-locations
of the impulses using the STFT.
The signals are sampled at 1000 Hz.

Instructions:

• Implement the STFT using the SciPy Signal Processing Library.

• The function scipy.signal.stft returns:

– An array of frequencies
– An array of time values
– A matrix representing the spectrogram

• You may also use scipy.signal.find peaks to help identify frequency or time fea-
tures. Be mindful of how you choose threshold values when filtering peaks.

Discussion:

• What trade-offs arise when selecting different window sizes?

• How does window size affect your ability to resolve:

– Time-localized phenomena (e.g., impulses)

• Which window type and size produced the most useful results for your signal?

6
Deliverables:

• A brief explanation of the window choices you experimented with

• Your best estimates for:

– The two sinusoidal frequencies

• Your Python code (included as an attachment or appendix)