MFCC (mel-frequency cepstral coefficients)
A dozen or so numbers per audio frame describing the rough shape of a sound — the workhorse of speech recognition for thirty years.
The recipe: slice the audio into frames of about 25 ms, run an FFT on each, sum the energy in triangular filters spaced along the mel scale, take the log, then apply a discrete cosine transform. Keep the first 12 or 13 coefficients. They capture the outline of the spectrum — the formants that make a vowel a vowel — while discarding pitch, which is exactly what you want when "ah" should be recognised whether a bass or a soprano says it. The cepstrum, a deliberate anagram of spectrum, dates from Bogert, Healy and Tukey in 1963; Davis and Mermelstein showed in 1980 that the mel variant worked best for speech.
Until the 2010s nearly every speech recogniser fed MFCCs into hidden Markov models. Deep networks prefer the log-mel spectrogram and skip the DCT, but MFCCs linger in speaker identification, music classification and every audio-analysis textbook.