Decoder. plain-English AI glossary

Mel Spectrogram

● Core

A spectrogram where frequencies are warped to the mel scale - which better matches human hearing than linear frequency.

Think of it like

A piano keyboard: low keys are spaced widely (low frequencies), high keys are spaced closely (we hear pitch logarithmically).

Example

Raw spectrogram: 0-22050 Hz linearly. Mel spectrogram: frequencies warped so differences we actually perceive are evenly spaced.

How it actually works

Humans perceive pitch logarithmically, not linearly. Mel scale (0-2595 mels) mimics this. Transform: take STFT magnitude, apply triangular filterbank on mel scale, take log (perceptual weighting). Smaller, more interpretable than raw spectrograms. Used as input to speech recognition, music tagging, audio classification. Variants: MFCC (adds DCT, compresses further). Provides good frequency resolution for speech and music.

For product teams

Standard input representation for audio models - improves performance and reduces data size.

For engineers

Convert waveform -> STFT -> power spectrum -> mel-scale filterbank -> log compression. Libraries: librosa, scipy.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open DecoderAdd to Chrome