Digital Signal Processing Tutoring

Digital signal processing is signals and systems with numbers instead of curves. Every idea from the earlier course returns in discrete form, and the new part is learning what a computer can and cannot do with a finite block of samples. Students are often comfortable with the z-transform on paper and uneasy with the DFT, because leakage, bin spacing, and zero padding behave in ways the continuous math never showed. Your tutor starts with small numeric examples and then connects them back to the formulas.

Course content usually includes sampling and aliasing, the DTFT, the DFT and FFT, the z-transform and system functions, FIR and IIR filter structures, filter design by windowing and by the bilinear transform, and sometimes multirate processing. Labs use MATLAB or Python with NumPy and SciPy. Sessions follow Oppenheim and Schafer, Proakis and Manolakis, Lyons, or your professor's notes, and we read your code and plots alongside the math.

A typical DFT problem, worked

A signal is sampled at $f_s = 8000\ \text{Hz}$ and we take an $N = 1024$ point FFT. Where do tones at 1000 Hz, 1010 Hz, and 5000 Hz show up?

Bin spacing is $f_s / N = 8000 / 1024 = 7.8125\ \text{Hz}$, and bin $k$ sits at $k \times 7.8125\ \text{Hz}$. The usable range is 0 to $f_s / 2 = 4000\ \text{Hz}$, which is bins 0 to 512.

The 1000 Hz tone lands at $k = 1000 / 7.8125 = 128$ exactly, so its energy stays in one bin (plus a mirror image at bin $N - k = 896$). The 1010 Hz tone lands at $k = 129.28$, between bins 129 and 130. Its energy spreads into neighboring bins, which is leakage. A Hann window cuts the far-off leakage, at the cost of a wider main lobe.

The 5000 Hz tone is above $f_s / 2$, so it aliases to $|5000 - 8000| = 3000\ \text{Hz}$, which is bin 384. In the FFT output it looks exactly like a real 3000 Hz tone. Only an anti-aliasing filter ahead of the sampler can prevent that.

Where students usually get stuck

• Reading bins above $N/2$ as high frequencies, when they represent negative frequencies
• Expecting zero padding to improve resolution. It interpolates the spectrum, while resolution depends on how long you record.
• Mixing normalized frequency in radians per sample with hertz. A value of $\pi$ corresponds to half the sampling rate.
• Forgetting to prewarp the cutoff frequencies before the bilinear transform
• Picking filter order by guesswork instead of working from the transition width and stopband attenuation