Word prediction is the assistive-technology name for software that offers the rest of what you are typing so that you can accept it instead of producing it keystroke by keystroke. It has been used for decades by people with motor impairments, repetitive strain injuries, tremor, and dyslexia, and the way it is judged has never changed: how many keystrokes does it save, and how much attention does it cost to read the offer? What has changed recently is the answer to the first half.
What word prediction actually does
Classic word prediction watches the characters you have typed so far and offers candidate words from a lexicon, usually ranked by frequency and by what you have written before. You type rec and it offers receive, recent, recommend. Choosing one saves the remaining characters of that word.
That is a real saving, and for someone for whom every keystroke has a cost it adds up over a day. But it is bounded by the length of a word. The ceiling on classic word prediction is a handful of characters per acceptance, and to get them you have to read a list, find the right entry, and select it — which is itself effort.
What a language model changed
A model that predicts from the meaning of the sentence so far, rather than from a lexicon, can offer the rest of the phrase instead of the rest of the word. That moves the ceiling: one keypress can replace forty or fifty characters rather than five.
| Classic word prediction | Phrase-level prediction | |
|---|---|---|
| Offers | The rest of the current word | The rest of the phrase or sentence |
| Saved per acceptance | A few characters | Tens of characters |
| How you choose | Pick from a ranked list | One key accepts, ignoring it costs nothing |
| Source | Word frequency and a personal lexicon | Sentence context |
| Fails by | Offering the wrong word | Offering a plausible phrase you did not mean |
The second failure mode is the one to take seriously. A wrong word in a list is easy to skip. A fluent, wrong sentence is easy to accept by accident, and for someone using prediction precisely because reviewing and correcting text is expensive, an unhelpful suggestion costs more than it does for a fast typist. This is why a phrase-level tool has to be willing to say nothing.
What to look for
One low-effort action to accept
Whether the offer is three words or fifteen, accepting should be one key — and it should be a key you can choose. A tool that requires reaching for a number row or a modifier combination has given back some of what it saved.
System-wide, not one application
Prediction built into a single word processor helps only in that word processor. Most typing during a working day happens in a browser, a mail client and a chat window. Ask what happens in those.
Local, if your text cannot leave the machine
Prediction that runs on a server sees everything you type. In clinical, legal and public-sector settings that is often not permissible regardless of how good the tool is. A model that runs on your own machine sidesteps the question entirely — and keeps working with no connection.
Willing to stay quiet
A tool that offers something at every caret position trains you to ignore it, and an ignored suggestion still cost you the glance. Fewer, better offers beat constant ones — particularly when reading is the expensive part.
Coexists with the rest of your setup
It should not demand exclusive control of the keyboard, and it should not conflict with a screen reader, a dictation tool, or an on-screen keyboard. Word prediction is one layer among several, not a replacement for any of them.
Where it does not help
It is worth being direct about the boundaries, because the category is often oversold.
- It is not dictation. If typing is not viable at all, speech input is the tool, and prediction has nothing to add to it.
- It is not AAC. Augmentative and alternative communication systems are communication devices in their own right, with their own vocabularies and access methods. Word prediction inside one is a feature of that system, not something you bolt on from outside.
- It is not a spell checker. A word you accept from a suggestion arrives spelled correctly, which is genuinely useful if spelling is effortful — but nothing here corrects a word you have already typed.
- It cannot work where an application tells Windows nothing. Some programs draw their own text and expose no accessible field. No system-wide tool can read those, and any that claims otherwise is guessing.
Where Typeahead fits
Typeahead is phrase-level word prediction that runs on your own Windows machine. It shows the continuation as grey text at the caret; one key — Tab by default, or the right arrow, or a key you choose — accepts it, and ignoring it and carrying on typing costs nothing. The model is downloaded once and runs locally, so nothing you type is sent anywhere and it works with the network off.
It withholds a suggestion when the model is not confident, which makes it quieter than it could be on purpose: roughly one caret position in three that would once have produced grey text now produces none, and the proportion of shown suggestions that turn out to be useful rose accordingly. For anyone paying attention costs to read each offer, that trade is the right way round.
If you are evaluating it for reduced typing effort specifically, the accessibility page covers the day-to-day side; if the constraint is that text may not leave the machine, local versus cloud prediction covers what each actually does with your keystrokes.
Frequently Asked Questions
What is word prediction software?
How is this different from the word prediction on my phone?
Does word prediction software work in every application?
Does word prediction need an internet connection?
Is word prediction a replacement for dictation or AAC?
Does it help with spelling?
Get Typeahead
One-time purchase, no subscription. 7-day free trial with every feature unlocked.
Get it on Microsoft Store