Spectrograms, Mel Scale & Audio Features
Neural nets do not consume raw waveforms well. They consume spectrograms. They consume mel spectrograms even better. Every ASR, TTS, and audio classifier in 2026 lives or dies by this single preprocessing choice. Take a 10-second 16 kHz clip. That is 160,000 floats, all in [-1, 1], almost perfectly uncorrelated with the label "dog barking" or "the word cat". The raw waveform has the information but in a form the model cannot easily extract. Two identical phonemes spoken 100 ms apart have completely different raw samples. A spectrogram fixes this. It collapses the temporal detail where human perception ignores it (microsecond jitter) and preserves the structure where perception attends (which frequencies are energetic, over time windows of 10–25 ms). Mel spectrograms push further. Humans perceive pitch logarithmically: 100 Hz vs 200 Hz sounds "the same distance apart" as 1000 Hz vs 2000 Hz. The mel scale warps the frequency axis to match. A mel-scaled spectrogram is the single most important feature in speech ML from 2010 through 2026. Waveform to STFT to mel spectrogram to MFCC ladder STFT (Short-Time Fourier Transform). Slice the waveform into overlapping frames (typical: 25 ms window, 10 ms hop = 400 samples / 160 samples at 16 kHz). Multiply each frame by a window function (Hann is the default; Hamming slightly different tradeoff). FFT each…
Spectrograms, Mel Scale & Audio Features: Neural nets do not consume raw waveforms well. They consume spectrograms. They consume mel spectrograms even better.…
This free lesson is part of the AI Engineering from Scratch curriculum. Read the full explanation, run the lesson code, and verify the result in the interactive reader or from the repository source.
Browse the complete course catalog or open this lesson on GitHub.