What we’ve read
Two public studies we found interesting while building this keyboard. Here they are with their numbers — including the one that proves us wrong.
37,000 people, 160 countries
Anna Feit (ETH Zürich), Sunjun Kim and Antti Oulasvirta (Aalto University), with the University of Cambridge. Presented at MobileHCI 2019, Taipei.
The largest measurement ever made of typing on phones. What it establishes:
- 38 words per minute on average with two thumbs — and 74% of people type that way.
- The fastest person in the sample: 85 words per minute.
- People aged 10–19 type about 10 words per minute faster than people in their forties.
- Word prediction does not save time. According to the researchers, the time spent considering a suggestion often exceeds the time typing the letters would have taken.
Aalto University’s announcement →
The dataset,
made public →
170 people, and the price of a bad correction
Ohoud Alharbi, Wolfgang Stuerzlinger and Felix Putze, 2020. Proceedings of the ACM on Human-Computer Interaction (ISS ’20).
An experiment simulating autocorrection and word prediction, to observe what people do when those mechanisms get it wrong.
- Repairing a wrong autocorrection takes 5.5 seconds on average.
- Word prediction saves 3.43 characters per phrase, but adds about two seconds.
- Mistakes at the start of a sentence are the hardest to repair: the cursor has to travel back there.
What proves us wrong
Both studies agree on one point that does not favour us: autocorrection does help typing speed. In the Aalto measurement, using autocorrection goes together with faster entry. In the 2020 study, it has no significant effect on overall speed, even counting the repairs.
So why does this keyboard not autocorrect? Because average speed is not the only thing that counts. An unwanted correction costs 5.5 seconds and the feeling that the device wrote in your place. We would rather leave you the last word, and make accepting a suggestion fast enough that the gain comes back — but chosen.
And this only counts against us once: when the rule is written. Writing “dont → don’t” takes ten seconds, once in your life. After that, every measurement swings our way: the rule applies without you reading a suggestion (the two seconds of deliberation are gone), and it cannot be wrong, because it guesses nothing — you wrote it. So there are never 5.5 seconds to pay to repair it.
A statistical autocorrection places its bet again on every word, and can lose every time. A rule is won once and for all.
Publishing what contradicts us is part of the method. A page that quoted only the convenient findings would not be worth reading.
Where our own measurements live
The figures we established ourselves — what language the world writes in, country by country, and which correction dictionaries iOS ships — live on the map of languages, with their sources.