The rare theory folder whose reps are genuinely fun and mostly free — NumPy on your laptop, then a ~$30 RTL-SDR dongle that turns the whole course into a working radio. Every project produces something you can hear, see, or decode.

Difficulty: 🟢 starter (hours–days) · 🟡 intermediate (a week or two) · 🔴 ambitious / portfolio-grade. ⭐ = highest signal.

The ladder

  • 🟢 ⭐ See a sound’s spectrum — record a few seconds of audio (a whistle, a chord, your voice), FFT it, and plot the spectrum and a spectrogram. Done when: you can point at the peaks and name the notes, and watch a rising whistle sweep up the spectrogram → spectral analysis. Do this first — an hour, and the frequency domain stops being abstract.

  • 🟢 Make aliasing happen — synthesise a tone above half your sample rate, sample it, and watch it appear as a different, lower tone. Done when: a 900 Hz tone sampled at 1 kHz plays back as 100 Hz, and you understand why → aliasing.

  • 🟢 Remove hum with a filter — take a recording with 50/60 Hz mains hum, design a notch filter (scipy.signal), and hear it vanish. Done when: the hum spike is gone from the spectrum and the audio is clean → filters.

  • 🟡 ⭐ Implement convolution from scratch, then prove the theorem — write the slide-multiply-sum by hand, verify it matches np.convolve, then compute the same result via FFT → multiply → inverse-FFT. Done when: both give the identical answer, and you’ve seen that convolution in time equals multiplication in frequency → convolution. The single most clarifying DSP exercise.

  • 🟡 Build a graphic equaliser — split audio into frequency bands, apply per-band gain, recombine. Done when: you can boost the bass and cut the treble on a real track and hear it → filters.

  • 🟡 FIR vs IIR, measured — design a low-pass both ways to the same spec, then compare tap count, phase response, and (for the IIR) plot the poles against the unit circle. Done when: you can state which you’d ship for audio (FIR, linear phase) vs tight embedded (IIR, efficient) and why → FIR vs IIR.

  • 🟡 A guitar tuner / pitch detector — capture audio, find the fundamental frequency, display the note. Done when: it correctly reads a plucked string, and you’ve handled the resolution/latency trade (you need a long enough window for low notes) → resolution.

  • 🔴 ⭐ Decode a real radio signal with an RTL-SDR — a ~$30 dongle, GNU Radio (or pyrtlsdr), and demodulate FM radio, then decode aircraft ADS-B transponders and plot the planes overhead. Done when: you’re seeing real aircraft on a map, decoded from raw I/Q you captured → SDR. The most satisfying payoff in this folder — and legal, because you’re only receiving.

  • 🔴 Build a digital modem — modulate a bitstream (BPSK or QPSK) onto a carrier, add noise, demodulate, and measure the bit error rate as noise rises. Done when: you can plot BER vs signal-to-noise and see it match theory → modulation.

  • 🔴 Connect it to ML — take an audio classification task, build the mel-spectrogram front-end yourself (don’t use a black-box loader), and feed it to a small CNN. Done when: you understand every step from waveform to the tensor the model sees → CNNs.

If you only do one

See a sound’s spectrum, then decode ADS-B with an RTL-SDR. The first is an hour and makes the frequency domain concrete; the second is a weekend and a dongle, and ends with real aircraft on a map decoded from radio you captured — which is about the most tangible thing any foundations folder can offer.

↑ Back to Digital Signal Processing