the four-step ladder every audio model climbs
1. waveform
T samples in [-1, 1]
10 s at 16 kHz = 160,000
frame+FFT
2. STFT
|FFT| per 25 ms frame, 10 ms hop
filterbank
3. log-mel
80 mels, log scale, [T, 80]
DCT (opt)
4. model input
Whisper: log-mel
Parakeet: log-mel
ECAPA: log-mel
KWS tiny: MFCC-13
tensor: [batch, T, 80]
mel scale is log-perception of pitch
mel
Hz
0
1 kHz
4 kHz
8 kHz
m = 2595 · log10(1 + f / 700)
triangular mel filterbank
Hz (linear) — triangles widen with frequency
each mel bin = weighted sum of FFT bins