Chapter 14 · Images, sound and video
Speech recognition
Speech models turn sound waves into spectrograms and then into text tokens.
Chapter 14 · Images, sound and video
Speech models turn sound waves into spectrograms and then into text tokens.
A microphone records air pressure wobbling: just one number per moment.
Nothing here understands language yet. It only matches patterns of frequency over time, then tidies the result.
A microphone measures air pressure thousands of times a second. On its own that wiggly line is hard to read.
Pick a made-up word, play it, then tap a column to see what that slice looks like and which sound it matches.
34 slices become 2 sounds: “see”. Green marks the slices that survive the squash.
Want to see your own voice? Your microphone stays off until you press this, and the sound is drawn in your browser only. Nothing is recorded, saved or sent.
A teaching toy: five made-up sounds and a nearest-match test per slice. A real recogniser uses a large network that reads many slices at once and knows which words exist.
What does a spectrogram show?
Why do labels repeat across slices?
What does the collapse rule do?