STTSpeech-to-Text

Sound has no spaces

Writing has blank space between words, but spoken sound runs together from start to finish. Stare at the waveform all you like — there's no mark that says "word ends here." The pauses a person actually takes while talking often don't line up with where meaning breaks anyway.

So before STT can turn sound into letters, it also has to decide where one chunk ends and the next begins. The exact same stretch of sound, cut in different places, turns into completely different sentences.

English piles on its own version of this trouble with run-together contractions and homophones, so the machine usually has to look at the surrounding words to cut things in the right place at all.

The sound gets sliced thin and spread out

Incoming sound is chopped into very short slices. One slice is far shorter than a blink of an eye — short enough that the sound barely changes within it.

For each slice, the system measures how much of each pitch is present and stacks that as a vertical strip. Line those strips up in time order, and the sound becomes a picture. That picture is a spectrogram. A human voice draws a fairly distinct pattern in this picture for every sound it makes, and that pattern is what lets a machine guess which sound it's looking at.

Transcription happens on top of this picture. Each slice gets scored for which letter or sound fragment it's likely to be, and those scores get stitched together in time order into words and sentences.

It picks whichever reading makes sense

A lot of speech can't be told apart by sound alone. Words like "flower" and "flour," or "there," "their," and "they're," sound nearly identical. In these cases the system doesn't just look at the sound score — it also weighs which reading makes sense given what comes before and after.

That's why the same sound can get written down as a different word depending on the sentence in front of it. On the flip side, a short phrase with no context around it gets misread often. So do unfamiliar names, shop names, and jargon that rarely shows up in training. Some services let you feed in a list of expected words ahead of time to raise accuracy on exactly this kind of case.

What comes out isn't a single answer but several candidates. The system picks the most plausible one to show you, and many services hand back how confident they are alongside it.

Conditions decide the score

Just as shade over a panel cuts its output, transcription is shaken hard by the conditions the sound arrives in. Café chatter, an air conditioner humming, and the echo of a room full of bare walls are the usual suspects. Move the mic farther away and the voice gets quieter while the noise stays the same — so things only get worse.

Several people talking at once is hard too. Overlapping sounds overlap in the picture as well, so pulling out just one person's share takes extra processing. That's why meeting-transcription services keep a separate feature just for splitting speakers apart.

Manner of speaking matters too. Speech styles that were common in training get transcribed well; regional accents, children's voices, and very fast talkers get misread relatively often.

Live and after-the-fact processing differ

Captions that pop up on screen right away process sound from the front as it arrives. The system has to judge without knowing what comes next, which is why text that just appeared sometimes quietly changes a moment later.

Processing a finished recording afterward lets the system look at the later context too and fix earlier guesses, so it comes out more accurate — at the cost of having to wait for the result. Processing on the device keeps sound from ever leaving it, which is reassuring; sending it to a server allows heavier processing. Which one you pick depends on how you weigh accuracy, speed, and privacy against each other.