How does AI understand speech? Speech-to-text explained simply

How does AI hear you?

AI doesn't just read text anymore. It can also hear you talk. You speak into your phone, and the words appear right away. But how does a machine listen?

SOUND WAVES

What is a voice to a machine?

To a machine, a voice is just a list of numbers: how the air vibrates, measured tens of thousands of times per second. Looking at those numbers, it's hard to tell what you said.

SPECTROGRAM

Turning sound into a picture

So the machine slices the sound very thin and splits each slice into low and high sounds. Line the slices up, and you get a picture called a spectrogram: bright spots are where there's sound.

READING THE PICTURE

Reading the picture into words

Each sound has its own shape: an "ah" is a few horizontal bars, a "sss" is a blur up high. A model that has heard hundreds of thousands of hours of recordings reads those shapes into words, using the meaning of the sentence too.

WHY IT GETS IT WRONG

Noise and people talking over each other

Anything that smudges the picture makes AI mishear. Coffee shop noise blurs the shapes of words. When two people talk at once, their shapes pile on top of each other.

NAMES

Names are easy to get wrong

AI rarely sees unusual names, so it often guesses a common word instead. A coworker named Leigh might turn into "lay". So always double-check names and numbers.

PART 2

It can listen. Can it talk?

AI can listen now. What about talking? How does a machine read text in a voice that sounds human? See you in part 2.