Advanced Text Statistics Tool

• Word counter
• Text analyzer
• Readability checker
• Word frequency
• Keyword extractor
• Text splitter

An average tells you almost nothing on its own. Two documents can share an average sentence length of twenty words while one reads evenly and the other alternates five-word fragments with thirty-five-word paragraphs-in-a-sentence. The second is the one a reader struggles with, and only a distribution shows it. This tool measures the shape of your writing — how lengths are spread, how uneven they are, how varied the vocabulary is — so you can decide what to edit before you send anything to an editor or run heavier language tooling over it.

Benefits

• Distributions, not just averages: word length and sentence length are each shown as a full histogram, with mean, median, standard deviation, minimum and maximum beside it.
• A diversity score that survives length: MTLD alongside the raw type-token ratio, so two documents of different sizes can actually be compared.
• Sentence detection that is right: decimals, abbreviations, initialisms, domains, filenames and ellipses stay inside their sentence instead of splitting it.
• Correct counts in every script: Chinese, Japanese, Thai, Khmer, Lao and Burmese are segmented properly rather than counted as one word per run.
• A punctuation profile: every mark counted, with density per sentence and per hundred words.
• Runs on your device, not our servers: no upload, and the page keeps working offline once it has loaded.

A 25 MB document is profiled in about a second on a desktop machine, with the work kept off the page so typing never stutters.

The text is walked three times: once over characters, once for sentence boundaries, once for words. Each pass writes into fixed-size counters rather than building lists, which is why a 25 MB document finishes in about a second and why the work can sit off the page entirely — typing never stutters, and there is no button asking you to opt in to the analysis you already asked for.

What each number means

• Words: runs of letters and, optionally, digits. Apostrophes and hyphens join a word when the options say so, so don't and state-of-the-art can each be one word or several.
• Sentences: boundaries found by the smart detector described below, or by a plain split on . ! ? if you choose simple detection.
• Paragraphs and lines: a line is a line with something on it; a paragraph is a block of lines separated from the next block by a blank line. Both are reported because neither is what everybody means.
• Mean, median and standard deviation: the average, the middle value, and how far from the average a typical value sits. A large standard deviation on sentence length is the signature of uneven pacing.
• Type-token ratio: unique word forms divided by all words, as a percentage. Simple, familiar, and length-dependent — see the diversity section.
• MTLD: how many words pass before the running type-token ratio decays through 0.72, averaged forwards and backwards. Higher means more varied.
• Hapax legomena: words used exactly once.

A full stop is not a sentence boundary. It is a character that is usually a sentence boundary, and the exceptions are common enough that a naive split gets ordinary business prose wrong. Smart detection is the default here and it holds a sentence together across all of the following.

How sentence detection works

• Decimals and version numbers — $3.50 and 1.2.3 are inside a sentence, not two or three of them.
• Dotted titles — Dr., Mrs., Prof., Lt. and the rest never end a sentence.
• Dotted abbreviations that sometimes do — etc., et al., e.g., i.e., Vol., Fig., Jan. Each of these breaks only when the next thing is a capital letter or an opening quote, which is the same evidence a reader uses.
• Initialisms — U.S.A. followed by a capital is a boundary; followed by a lower-case word it is not.
• Domains and filenames — example.com and report.txt are not two sentences.
• Ellipses — Wait… nothing happened is one sentence; Wait… Nothing happened is two.
• Closing quotation marks and brackets belong to the sentence they finish, so “yes.” and (see the note.) are not left starting the next one.
• A blank line ends a sentence, so a heading or a list item is a sentence of its own rather than being glued to the paragraph below it.
• Full-width stops, the Devanagari danda, and the Arabic question mark and full stop all end sentences, so a Japanese or Hindi document is not reported as containing one.

Simple detection is kept alongside it, and it does what almost every other online counter does: break at every . ! ? and nothing else. It is there so you can reproduce another tool's number when you need to explain a difference, not because it is more correct.

The familiar measure is the type-token ratio — unique word forms divided by total words. It is easy to explain and it has one flaw that matters more than it sounds: it falls as a text gets longer, whatever the writer does, because common words repeat and the pool of new ones is finite. A 500-word essay and a 90,000-word book by the same author will score very differently, and the difference is length, not vocabulary.

Vocabulary diversity

• Type-token ratio is shown because it is the number people know, and because on texts of similar length it is perfectly usable.
• MTLD is shown beside it because it is not length-dependent. It counts how many words pass before the running ratio falls through a fixed threshold, then reports words per fall — a rate, so document length cancels out.
• MTLD is computed forwards and backwards and averaged, because a text that front-loads its unusual words scores differently in each direction.
• Below about fifty words there is nothing to measure and no MTLD is shown, rather than a confident number invented from too little text.
• Hapax legomena — words used exactly once — is the third view. Together the three separate a wide vocabulary from a repetitive one and from a short one.

Paste text or open a plain-text file, then read the length distributions, diversity scores and punctuation profile.

How to get text statistics online

1. Add your text — Paste into the box, or open a .txt, .md, .log, .csv or .tsv file. Nothing is uploaded — the analysis runs on your device.
2. Choose how words and sentences are counted — Smart sentence detection keeps decimals, abbreviations and initialisms whole. Simple detection breaks at every full stop, which is what most other counters do. Numbers, apostrophes and hyphens can each be included or excluded from what counts as a word.
3. Read the profile — The distributions show the shape of your writing, not just its average. Standard deviation says how even your sentences are; MTLD says how varied your vocabulary is without penalising length.
4. Export or compare — Copy the report or download it as text. With Pro, export CSV or JSON, and profile up to 100 documents into one comparison table with all reports in a single archive.

Any time you are about to make an editing decision and want evidence rather than an impression, and any time you need to compare two versions of the same thing.

Ideal use cases

• Editing a draft: find the one paragraph-length sentence, and see whether the pacing is even or lurching.
• Comparing versions: profile the before and after and see which numbers actually moved.
• Comparing a set: with Pro, profile a folder of drafts into one table and find the outlier without opening any of them.
• Style guides: check a house limit on sentence length against the distribution rather than the average.
• Teaching and research: demonstrate lexical diversity measures on a real text, with the raw ratio and MTLD side by side.
• Survey and review text: open the CSV column of free-text answers and profile it as prose.
• Private drafts: unpublished manuscripts and confidential documents, with nothing uploaded.

What the numbers are, precisely, and where they stop. Everything below is a property of the measurement rather than a caveat about the tool being unfinished.

Key points

• Word and sentence length are measured in code points, so an accented letter or an emoji counts as one character rather than two.
• Word length counts the letters and digits of the token. Punctuation attached to a word — the comma after it, the quotation marks around it — is not part of the word.
• Sentence length is measured in words, using the same word definition your options select.
• The median comes from the exact distribution, not from a sample, so it is the true median at any document size.
• Standard deviation is the population value over all words or all sentences, not a sample estimate.
• The longest and shortest sentence are shown up to 4,000 characters; beyond that the preview is trimmed and says so, while the length reported is the real one.
• Ties go to the first occurrence, so the longest word shown is the first word of that length in the document.
• A distribution with a long tail is binned into 24 ranges rather than truncated — the handful of very long sentences is the finding, not noise to cut off.

This tool reads plain text. Anything else is refused at the file picker with a link to the tool here that extracts its text, because a Word document fed to a text scanner produces statistics about markup rather than about prose — a page full of confident, plausible, completely wrong numbers.

What it reads, and what it does not

• Reads: .txt, .text, .md, .markdown, .log, .csv and .tsv.
• Encodings: UTF-8, UTF-8 with a byte-order mark, UTF-16 in either byte order, and Windows-1252 as a fallback. A file saved from Windows Notepad as “Unicode” is UTF-16 and is read correctly rather than as a string of blanks.
• Word documents: extract the text first, then profile it.
• PDFs: extract the text first, then profile it.
• Slide decks: extract the text first, then profile it.
• Spreadsheets: export the column you want as CSV, which this tool reads directly.
• Images, audio and video: recognise or transcribe the text first.
• Binary files renamed to .txt: refused, rather than analysed into nonsense.

The numbers are only as meaningful as the comparison you put them in.

Quality considerations

• Compare like with like. Change one thing between two profiles — the draft, not the draft and the options at the same time.
• Use MTLD, not the type-token ratio, whenever the two documents differ in length by more than a little.
• Read the standard deviation next to the mean. A mean of 18 with a deviation of 4 and a mean of 18 with a deviation of 16 are different documents.
• Strip markup before profiling if you want statistics about the prose. Markdown syntax counts as punctuation, which is accurate but probably not what you are asking.
• For a legal or academic word count, check which definition of “word” is required and set the number, apostrophe and hyphen options to match it before quoting a figure.
• Do not read vocabulary diversity as a quality score. It measures variety, and variety is not always what a piece of writing needs.

When not to rely on it

• When you need a graded reading level — Flesch, Fog, SMOG or a grade equivalent. The readability checker on this site computes those.
• When you need which words appear and how often. The word frequency counter and the keyword extractor answer that.
• When sentence boundaries must match a specific publisher's house rules rather than ordinary written usage.
• When you need syllable, lemma, part-of-speech or n-gram statistics.
• As a plagiarism, grammar or SEO suite. It measures, it does not judge.

Free profiles one document up to 25 MB — around four million words — with every metric included: nothing on this page is behind the paywall. Pro removes the imposed size limit, raising the ceiling to what your device can handle, and adds the things scale needs: up to 100 documents in one pass, a comparison table across them, CSV and JSON export, and every report in a single archive. The tool needs JavaScript and a current browser; the analysis itself runs off the page, so a large document never freezes the tab.

Recommended editing workflow

• Profile the draft first, and read the sentence-length distribution before anything else — the tail is where the editing is.
• Open the longest sentence, split it, and profile again to confirm the deviation actually came down.
• Check vocabulary diversity with MTLD if the two drafts differ in length; with the type-token ratio if they do not.
• Use the trimmer, splitter or deduplicator for the line-level clean-up the profile pointed at.
• Finish with a readability check and a keyword pass before publishing.

For a full review, use this profile together with the readability checker for graded scores, the word frequency counter for which words you lean on, and the word counter when all you need is a number.