Signal Processing

Signal processing is the branch of engineering and mathematics devoted to analyzing, modifying, and synthesizing signals: quantities that vary over time or space and carry information, such as sound, images, video, biomedical signals (the electrocardiogram, or ECG, and the electroencephalogram, or EEG), and sensor readings. As the IEEE Signal Processing Society frames it, the word «signal» spans audio and speech, radar and sonar, and medical and musical signals alike.

It is not an «artificial intelligence» technique but a classical field —older than modern AI— that AI draws on constantly. Its aim is to pull useful information out of a signal, clean it of noise, or transform it so it can be transmitted, stored, or interpreted.

Time domain, frequency domain, and the Fourier transform

The same signal can be described in two ways. In the time domain, you watch how its amplitude changes moment to moment —the way a microphone captures a sound wave. In the frequency domain, you see which frequencies (tones) the signal is made of and how strong each one is. The Fourier transform, named after the mathematician Jean-Baptiste Joseph Fourier, is the bridge between the two: it decomposes a signal into the sum of the sine waves that compose it, letting you move from time to frequency and back. Many problems that are hard to solve by staring at a waveform become obvious in the frequency domain.

Sampling and filtering

Filtering is the other basic operation. A low-pass filter lets low frequencies through and attenuates high ones (smoothing a signal and removing high-pitched noise); a high-pass filter does the opposite (stripping out, say, the slow baseline drift of an ECG); a band-pass filter keeps only a slice of frequencies. Filtering, at heart, is deciding which part of the spectrum to keep and which to discard.

Signal processing and artificial intelligence

For decades, feeding a raw signal to a model gave poor results, so engineers hand-crafted «features» that summarized what mattered. In speech recognition, the canonical example is the MFCCs (mel-frequency cepstral coefficients), which compress the speech spectrum by mimicking how the human ear perceives sound. Today, deep learning often learns directly from the raw signal —the waveform or the spectrogram— and discovers useful features on its own. But signal processing has not vanished: it remains essential in acquisition (sensors, analog-to-digital conversion), in preprocessing (sampling, filtering, normalization), and in countless embedded systems where a large model simply does not fit. AI uses signal processing; it does not replace it.