Letter Frequency Counter

What is letter frequency?

Letter frequency is how often each letter turns up in a body of writing. In English the answer is famously uneven: E is worth about one letter in eight, while Z is worth about one in sixteen hundred. That lopsidedness is what makes Scrabble tile values work, what decides which keys wear smooth first, and what breaks every substitution cipher ever invented.

Almost every page that publishes a letter frequency table prints one table, unlabelled, as though it were a physical constant. It is not. A frequency table is a measurement of a particular pile of text, and different piles give different answers. This tool ships two tables, names the corpus behind each, and lets you switch between them so you can watch the expected column move.

Paste anything into the box and it is counted here, in your own browser. The counts themselves are exact. The interesting part is the column next to them, which tells you how far each letter sits from the reference — and whether a gap that size is remarkable or completely ordinary.

How to use the letter frequency counter

1. Paste or type your text into the box. Counting starts immediately, in your browser. Accented letters count as their base letter, so café adds an e; spaces, digits, punctuation and non-Latin script are tallied separately as ignored characters.
2. Choose which table to compare against. Twelve public-domain novels, or Google Books. They disagree, which is the point — flipping between them changes the expected share of every letter and the sign of some of the differences.
3. Read the verdict before the table. It tells you whether your sample is long enough to interpret at all, and whether its distance from the reference is unusual or is simply what text that length looks like. Sort A to Z or commonest first, and copy the whole thing as CSV.

Real text does not converge on English — it converges on its author

Here is the result that shaped this page. Take the twelve-novel corpus and draw letters from it independently at random. As the sample grows, its distribution closes on the corpus table exactly as sampling theory says it should: by 200,000 letters the distance is 0.0040, essentially a perfect match. Now take real continuous passages of the same length from the same books. They sit at 0.0301 — seven and a half times further away.

And the gap widens as the sample grows. At 30 letters a real passage is no worse than a random draw (0.3018 against 0.3237). At 1,000 letters it is 1.3 times worse, at 10,000 letters 2.4 times, at 200,000 letters seven and a half times. Past a few thousand letters you are not measuring English more precisely any more. You are measuring one author more precisely, and the author is not the language.

That has a consequence for the reference tables themselves. The twelve-novel table here and Norvig's table built from 3,563,505,777,820 letters of Google Books sit 0.0305 apart. The median distance between two individual novels in this corpus is 0.0449, and the closest two are 0.0236 apart — so seven of the sixty-six pairs of novels are closer to each other than the two reference tables are to each other. A corpus 668,000 times larger did not settle the question. It answered a different one.

You can see the disagreement in the order. Both tables open E T A O and then split: the novels give ETAONISHRDLU, Google Books gives ETAOINSRHLDC, and only fifteen of the twenty-six letters hold the same rank in both. The biggest single gap is H, at 6.18 per cent in the novels against 5.05 per cent in Google Books. Fiction is full of he, she, that and the. Google Books is full of everything else.

How much text you need, and what this tool cannot tell you

The amount of text required turns out to be a property of the question, not of the text. Thirty letters is enough to break a Caesar cipher 98.0 per cent of the time — and the two tables that disagree with each other above pick the same shift 99.3 per cent of the time at that length. Yet in those same thirty letters, the commonest letter is actually E only 41.3 per cent of the time. Breaking a cipher needs the ranking roughly right in aggregate; reading one frequency off the table needs that single letter individually settled.

Scale it up and the pattern holds. At 1,000 letters the commonest letter is right 96.7 per cent of the time. But the exact top six letters come out in the right order only 22.5 per cent of the time even at 100,000 letters, and only nine of the twelve complete novels reproduce the corpus's own top three. If you have seen a page confidently list the top ten English letters in order, that ordering is more fragile than it looks.

This is also why the tool marks a letter wide rather than significant. Textbook statistical significance treats your text as independent random draws, which prose is not, so on any long sample it flags nearly every letter as meaningfully off. Instead a letter is marked when it sits further from the table than the twelve individual books sit from each other. H alone runs from 4.90 to 8.10 per cent between two ordinary novels — a spread of 3.20 percentage points on a letter worth 6.18. Judging against sampling noise would call ordinary books extraordinary.

The honest limits: this counts a to z only, folding accents onto their base letters, so it measures the Latin skeleton of a language rather than its real alphabet — Spanish ñ and German ß are not counted as themselves. Both reference tables are English, so text in another language will read as unusual against them, correctly. Below about 200 letters the tool declines to interpret, because at that length the answer is noise. And it can tell you that a text is unusual, never why.

Why is it free?

Because counting letters costs us nothing. The whole analysis — the counts, the distances, the reference tables, the convergence curve behind the verdict — runs as a few kilobytes of JavaScript inside your own browser. Your text is never uploaded, never stored and never seen by us, because there is no server involved to see it.

So there is no account, no sign-up, no upload limit and no watermark on the CSV. Paste a private manuscript if you like; it does not leave the machine you are reading this on. The tool is free for the same reason a pocket calculator is free to use twice.