What is speech recognition? AI that listens to spoken words and turns them into written text

AI that listens to spoken words and turns them into written text.

Speech recognition turns the sound coming into a microphone into text. The AI chops the audio into tiny slices, reads the pattern of pitch and loudness in each one, and looks at the slices around it to decide which word it's hearing. That's why the same sound can be written differently depending on the rest of the sentence.

Accuracy jumped in the early 2010s when deep learning arrived. In 2022 a model called Whisper, trained on about 680,000 hours of audio, was released for anyone to download. It still slips up with background noise, strong accents and rare words.

Think of it like this

Think of "their", "there" and "they're". They sound the same, so the word on its own won't tell you which to write, but the whole sentence will. Speech recognition does the same: it listens to what comes before and after, then picks the most likely words.

For example

When you talk to your phone and the words appear in your message, that's speech recognition.

Common mix-ups

Mix-upSpeech recognition just turns each sound into a letter.
ActuallyIt even fills in things that aren't in the sound. You can't hear full stops or spaces in speech, yet the transcript has them, and if you say "three thousand won" it may write "3,000 won" in figures.
Mix-upToday's speech recognition never makes mistakes.
ActuallyThere are also reports of it inventing words nobody said during silent stretches. For important transcripts, like a recorded lesson or an interview, listen back to the recording and check names and numbers first.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: Which words does speech recognition get wrong most often in your class?
Activity: Each group reads the same five sentences into a phone's dictation. Change the conditions — a quiet room, music playing, sentences with regional words or subject jargon — and record how many words come out wrong in a table. Compare the conditions, then deliberately write sentences with sound-alike words (there/their, to/two) and check those too.