Speech & audio

Voice typing: how dictation works

Voice typing listens to your speech, turns the sound into likely words with a trained model, and keeps revising its guess as you keep talking — and you can watch it happen in the demo.

See it for yourself

Press the button and talk. Interim guesses appear gray, then firm up — the revising is the interesting part.

How it works

Sound becomes clues

When you talk, your microphone picks up one long wiggly line of sound. The recognizer chops that line into tiny slices and turns each slice into a set of features — numbers describing what the sound is like at that instant, the way you might describe a voice as high or low, hissy or humming.

Clues become likely words

A trained model then asks: given these sounds, and given the words that came before, what is this person most likely saying? It is guessing — but guessing from enormous experience of how speech sounds and how sentences tend to go. "Recognize speech" and "wreck a nice beach" sound nearly identical; context is what tells them apart.

Why the gray words change

The recognizer never bets on one sentence at a time. It keeps several candidate sentences alive at once, and when a later word makes an earlier guess less likely, it goes back and swaps it. That is the flicker in the demo above: gray words are guesses still open to revision, and they firm up once the context around them settles.

Where it came from

Speech recognition is older than it feels. In 1952, Bell Labs built "Audrey," a machine that could recognize spoken digits — the numbers, and not much else. A decade later, in 1962, IBM's Shoebox understood sixteen words. That was the frontier: years of serious engineering to get a machine from ten words to sixteen.

The 1990s brought dictation home. Dragon shipped DragonDictate to consumers, then NaturallySpeaking in 1997, its first consumer product that could follow continuous speech, ordinary sentences at an ordinary pace. Under the hood these were statistical models: vast tallies of which sounds tend to mean which words, and which words tend to follow which. That approach ruled the field for years.

Then deep learning rewrote the accuracy numbers in the 2010s, and recognition became something a browser could hand to any web page through the Web Speech API — which is what the demo above uses. The machine that once managed sixteen words now sits behind a button on this page.

Who it helps — honestly

Voice typing earns its place wherever spelling or typing is the bottleneck between a person and their own ideas. That includes many people with dyslexia, who know exactly what they want to say while spelling it drains the effort; people with dysgraphia; and people with motor differences or RSI, for whom a keyboard is slow or painful. Speaking routes around the bottleneck: you think, you talk, the words arrive.

Now the honest limits. Recognition still makes mistakes, and it makes more of them with strong accents, background noise, and homophones — "their" and "there" sound identical, so the recognizer can only guess from context. It wants a quiet-ish space, which not everyone has. And dictating punctuation — saying "comma" and "new line" out loud — feels strange at first and takes practice before it stops interrupting your train of thought.

So treat dictation as a faster first draft, not a finished one — you will still proofread. Where Helperbird comes in: it brings voice typing to browsers and text fields that lack it on their own, notably Firefox. Try it on this page first; if watching your words land feels like a relief, that is your answer.

Try it in Helperbird

Helperbird brings these tools to any website, PDF, or Google Doc — no copying text anywhere, and nothing you read leaves your browser.

Voice typing Dictate into any text field on any website, including Google Docs.

Step-by-step: How to dictate an equation with your voice

Questions people ask

Is there speech to text in Firefox?

Firefox does not ship dictation for ordinary pages. Helperbird adds voice typing there, and in Chrome, Edge, and Safari — click into any text box and talk.

Is there a voice typing extension for any website?

Yes — that is Helperbird's speech-to-text: click into any text box, talk, and the words arrive, rather than dictating into one app and pasting.

Why does dictation change words after I said them?

Because it is guessing from context. The recognizer holds several candidate word sequences at once, and a later word can make an earlier guess less likely — so it swaps it. The revision you see is the statistics working, not a bug.

Sources

Share: X Facebook LinkedIn Email