Mel spectrogram

A mel spectrogram is the time-frequency picture of audio that speech models actually consume — a spectrogram warped to the mel scale, which spaces frequencies the way human hearing does.

Nearly every modern speech model — Whisper included — reads mel spectrograms rather than raw waveforms. The representation compresses what matters perceptually into an image-like input neural networks handle well.

Related terms

MFCC (Mel-frequency cepstral coefficients)MFCCs are a compact numeric summary of a sound's spectral shape on the mel scale — the dominant audio features… WaveformA waveform is the visual plot of an audio signal's amplitude over time — the shape you scrub through in every … FrequencyFrequency is how many times per second a sound wave cycles, measured in hertz (Hz) — perceived as pitch, with …