• A Fourier transform tells us which frequencies exist in a signal, but not when they occur. Speech changes constantly, so we apply it to short overlapping windows.
  • Stack those windows over time and you get a spectrogram: time on one axis, frequency on the other, and energy represented by intensity.
  • This is why convolutions work well on audio. A spectrogram is a grid with local patterns, much like an image.

Connections

  • related_to Spectrogram — STFT windows stacked over time