Speech systems note

Audio Is Computer Vision

A spectrogram is a picture of sound: time on one axis, frequency on the other, color as energy.

July 2026 · spectrograms · speech recognition · video analysis

Spectrograms Convolutions Speech Recognition Video Analysis

Once audio is drawn as a spectrogram, the analogy to computer vision becomes very direct.

  • Image classification: pixels -> convolution or encoder -> label.
  • Audio classification: spectrogram -> convolution or encoder -> label.
  • Speech recognition: spectrogram over time -> encoder/decoder -> word sequence.
  • Video analysis: frames over time -> image encoder plus temporal model -> action or event.

The common move is learning local patterns on a grid. In images those patterns are edges, textures, and parts. In spectrograms they are bursts, harmonics, formants, silence, and other time-frequency shapes.

Speech recognition and video analysis then add the same extra problem: order over time. You are not just classifying one grid; you are reading a sequence.

Plainly: the analogy is not perfect, but it is the first mental model I use when explaining speech recognition to someone who already understands vision.