11.2 DSP features for machine learning
You can turn a raw waveform into features a model can learn from.
Before:06. Deep Learning
FFTs, spectrograms, mel features and MFCCs turn raw waveforms into representations models learn from — the front end of every audio and vibration system, with librosa and scipy as the working tools. It sits early in the bridge module because features precede models. The deep-learning-learns-features slogan gets its correction here: on small datasets, engineered spectral features routinely beat end-to-end learning, and knowing both approaches is the professional position.
Work through these
FFT, STFT and spectrograms
Moving a signal into the frequency domain, doing it over short windows so the result varies with time, and displaying that as an image. This transformation is the foundation of nearly all audio machine learning.
Mel spectrograms and MFCCs
Two representations built to match how human hearing works, one a warped frequency scale and the other a compact summary of it. These remain the standard features for speech.
Filtering, denoising, windowing
Removing noise, restricting to a band of interest, and shaping the window so the transform behaves. These preparation steps affect results more than the model choice does.
Feature extraction with librosa and scipy.signal
The two libraries that implement all of the above, one specialised for audio and one general. Knowing which to reach for saves reimplementing standard transforms.
Sign in to keep your progress.
Free resources
We haven't checked most of these for screen reader use yet.
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.