Speech recognition: how a machine writes “pare a pair of pears” correctly

How does a machine turn a voice into text? Here's a hint: it doesn't listen. It looks. In 1952, at Bell Labs, a machine called Audrey recognized the digits zero to nine, but it was tuned to a single voice. In 1997, Dragon NaturallySpeaking let people dictate to a computer without pausing between words. In 2022, Whisper, released openly, learned from six hundred and eighty thousand hours of speech: today it transcribes dozens of languages. Three steps. Let's go. Step one: sound becomes an image. Speech makes the air vibrate, and a microphone measures that sound wave thousands of times a second. The machine sorts it by frequency, low to high, and colours each moment by its energy: that's the spectrogram, a picture of your voice. Step two: the image is cut into thin slices, twenty-five milliseconds each. Slice by slice, a neural network estimates the likeliest sounds: the probabilities rise, and a string of sounds appears. Step three: context decides. One sound, several spellings. Listen: pare a pair of pears. Same sound, three meanings! A language model reads the whole sentence and picks the right spelling: to peel, a set of two… and the fruit. That's exactly what happens in the KOAEE assistant. In every sector, everyone taps the mic and speaks their own language: Japanese, Korean, English. It understands speech in sixteen languages, and replies in the same one. In French and English, dictation runs on Voxtral, by Mistral, and stays in Europe. To understand AI in 150 seconds, subscribe to the channel. The AI that stays between you. Try it at koaee.ai.