Notes · Speech from first principles
Mel filterbank
also: MFCC front end · log-mel
A set of overlapping triangular weighting filters spaced on the mel scale — roughly linear below 1 kHz, logarithmic above — that sum spectral energy into a few dozen bands. It is a cheaper, blurrier route to the same formant envelope the cepstrum recovers: averaging within each band smooths away the glottal source’s fine harmonic comb, and the perceptual spacing puts resolution where vowels actually differ.
Take the log of the band energies and a discrete cosine transform — a cepstrum on the warped spectrum — and you have MFCCs, the front end that defined speech recognition for decades. Modern encoders often drop the DCT and feed log-mel energies straight into a convolutional stack, but the filterbank itself has proved stubbornly hard to beat.