waveform → samples → spectrum
continuous pressure
microphone output, smooth in time
sampled at sr Hz
float array, indexed by n / sr
DFT magnitude
0
sr/2
energy per frequency bin
sample
DFT
aliasing — energy above Nyquist folds
sr/2 (Nyquist)
f_true = 7 kHz
alias = 3 kHz
fix: low-pass before downsample
2026 sample rates
8 kHztelephony — avoid for ASR
16 kHzASR standard (Whisper, Parakeet)
24 kHzmodern TTS (Kokoro, F5, xTTS)
44.1 kHzCD audio, music
48 kHzfilm, pro audio, hi-fi TTS
first question before any ML pipeline