Typing accuracy: what it really measures and what a good score is

How is typing accuracy calculated?

The number most sites show is keypress accuracy. The typing accuracy formula is:

accuracy = correct character keypresses ÷ all character keypresses × 100

If you pressed 160 keys and 6 were wrong, that's 154 ÷ 160 = 96.25%.

Two details decide whether this number is honest. The first is whether backspace counts as a keypress. We leave it out. If you included it, every correction would add a "correct" keypress and pad the figure. The second is whether a corrected error still counts as an error. It does here. You pressed the wrong key; deleting it later doesn't change that. A test that quietly forgives fixed errors will report higher accuracy for exactly the same typing, so check this before comparing.

This is the measure on our headline row, and it's the one our 90% leaderboard floor uses. It describes your fingers. It does not describe your text.

Uncorrected errors: what's left in the text

Take the text you were shown and the text you ended up with. The Levenshtein distance between them (the fewest single-character edits that turn one into the other) divided by the longer of the two lengths is the uncorrected error rate. In the example above, two errors were left in a 156-character final text: 2 ÷ 156 = 1.28%.

This is the measure text-entry researchers use for the finished product, and it's the one Dhakal and colleagues reported in their 2018 study of 136 million keystrokes. Across 168,960 volunteers the mean uncorrected error rate was 1.17%, and nine in ten people left under 2.66%. Trained typists left 1.02% uncorrected; untrained, 1.23%. The gap is smaller than most people expect, because almost everyone fixes almost everything. What differs is how much fixing costs.

A low uncorrected rate next to a middling keypress accuracy means a lot of backspacing.

Mistake percentage, the number exams want

"Percentage of mistakes" is the phrase that shows up in exam notices and in searches from candidates. On our results screen it's 100 minus keypress accuracy: wrong character keypresses over all character keypresses. For the example, 6 ÷ 160 = 3.75%.

Researchers call this the total error rate: corrected and uncorrected errors added together. Arif and Stuerzlinger's 2009 analysis of text-entry metrics called it "the most powerful error rate metric at the present time", because it charges you for the effort of correcting. They also found something that should worry anyone comparing accuracy across sites. Changing how errors were handled shifted the error metrics by roughly 18% to 66%, depending on the metric, while WPM showed no significant change for expert keyboard typists. Speed holds up across definitions. Accuracy doesn't.

Exam boards often count words, not keys. India's CBSE defines accuracy as net words per minute (correct words only) divided by gross (all words typed). Its own illustration has a candidate typing 390 words in 10 minutes with 2 wrong: 38.8 ÷ 39 = 99.49%. That number only sees the finished text, so fixed errors cost time but never accuracy, and one wrong letter costs a whole word. At 97% keypress accuracy with errors left in, about one word in six would be wrong (the arithmetic is below), so word accuracy can read far lower or far higher than keypress accuracy depending on how much you fix.

US federal hiring sets its bar as an error count instead: 40 WPM on a five-minute sample with three or fewer errors. That's about 1,000 characters, so three errors means roughly 99.7% of the text left clean.

Indian exam boards often weight "full" and "half" mistakes differently and count against the passage. Our mistake percentage is stricter than most of those schemes, so passing comfortably here is a reasonable sign. More in the government exams article; the exam practice pages run the common formats.

KSPC, the number that explains the gap

Keystrokes per character: every character keypress plus every backspace, divided by the length of the text you finished with. In the example, (160 + 4) ÷ 156 = 1.05. A KSPC of 1.00 means no wasted motion. The mean in Dhakal's data was 1.17, so the average online typist spends 17% more keystrokes than the text needs.

We almost left this odd unit out. On our own runs it tracked the gap between raw and net more closely than accuracy did. Dhakal found the same thing at scale: speed correlates with KSPC at minus 0.40, with error corrections at minus 0.36, and with uncorrected errors at only minus 0.21. What slows you down isn't the errors you leave. It's fixing the ones you make.

Why 90% at 80 WPM is worse than it sounds

Assume errors land at random. A common word plus its space is about six keystrokes. At 90% accuracy, the chance all six are right is 0.9 to the sixth power: 0.531, about 53%. Nearly half of all words contain a mistake. At 97% the same sum gives 0.97 to the sixth, 0.833, so about one word in six.

If nothing were corrected, net WPM would be roughly raw WPM times that clean-word share. Raw 80 at 90% gives about 42 net. Raw 70 at 97% gives about 58. Slower fingers, faster result.

In practice you correct: a backspace or several, a retype and a pause to read what went wrong. Dhakal's volunteers averaged 239 ms between keystrokes, and a correction takes several of those. The net number climbs back toward raw, but raw itself falls, because the time went on fixing. Either way the 90% typist pays.

There's a second reason 90% is a cliff rather than a slope. Errors aren't random. They cluster on the same letter pairs, the same fingers, the same hand alternations. In Dhakal's bigram analysis, the timing of hand-alternating pairs predicted overall speed at r = minus 0.72, far more than repeated letters at minus 0.32. A 90% typist usually has a few specific weak spots dragging everything else down, and one aggregate percentage hides which ones. Our most-missed keys and per-word breakdown exist to show them, and the practice drill is built from them.

Is 90% typing accuracy good?

It's passable, not good. Our public leaderboards require 90% keypress accuracy.

We considered a much lower floor first. At 75%, one keypress in four is wrong. Nobody types prose at 75% by accident; it's what you get when someone sprays the keyboard and lets the net count keep whatever survives, which on a short test can be a high number. That's not the skill we want a leaderboard to rank.

At 95% we'd have shut out many honest typists on an ordinary day. In Dhakal's data, 6.3% of all keypresses were Backspace or Delete. Corrections and errors don't map one to one, but that points to typical keypress accuracy somewhere in the low-to-mid nineties. A 95% floor would cut a large share of ordinary typists; 90% keeps most of them and still removes the high-raw, low-net pattern we wanted off the boards.

The floor applies to the boards only. Your result is always shown to you, saved and counted toward your personal bests whatever the accuracy. The same server code that recomputes every result from its keystroke log applies it, so a browser that reports a flattering accuracy gets corrected before anything is ranked.

One admission: the floor is a blunt tool. 90% at 40 WPM and 90% at 140 WPM are different kinds of sloppy. We considered a sliding floor that rose with speed and decided it would be impossible to explain on a results screen. A single number people can remember beat a fairer one they couldn't.

Reading your own accuracy

Keypress accuracy under 95% with uncorrected errors near zero means you're fixing everything and paying for it. Your net WPM will rise fastest if you slow down until accuracy is 97% or better, then speed up again. This is the most common pattern we see.

Keypress accuracy near 98% with uncorrected errors at 2% or more means you type cleanly and don't look back. Fine for a timed test, bad for a transcription job, and exam modes will punish it.

KSPC above 1.2 means a sixth of your keystrokes are wasted, usually on the same few words. Open the per-word breakdown on the results screen and look at the words colored for low burst speed.

Consistency under 70% with decent accuracy usually means pauses, not errors; the per-second chart shows them as flat spots.

What accuracy doesn't measure

Accuracy says nothing about posture, finger assignment or whether you're touch typing. Dhakal's volunteers who reported using more fingers were faster (r = 0.38 between finger count and speed), but a percentage won't tell you how many you use. It also says nothing about the text: 90% on a quote full of capitals and semicolons is a different achievement from 90% on common lowercase words, which is why the punctuation test is worth taking separately.

The exact formulas, with the worked example these figures come from, are on the metrics page. How accuracy interacts with speed across sites is in why your WPM differs between sites.

Frequently asked questions

What is a good typing accuracy?

95% or higher on a normal test; 97% or more if you want your net WPM close to your raw WPM. Below 90%, about half your words contain a mistake.

Does fixing a typo raise my accuracy?

No. On this site a wrong key counts against you even after you delete it. Fixing it only lowers your uncorrected error rate.

What accuracy do typing jobs and exams expect?

Exams in India often allow a set share of mistakes against the passage; transcription work expects very few errors left in. Practicing at 97% keypress accuracy covers both.