A word counter has to decide where one word ends, and most decide by looking for spaces. We tested this on the Universal Declaration of Human Rights, taking each translation as the unicode-org UDHR project publishes it, prefatory note included. Split at spaces, the Simplified Chinese file holds 97 words. The dictionary segmenter that browsers expose to JavaScript, run in Node.js, finds 1,606. Microsoft Word for Mac 16.114.3 reports 2,879, because it treats every Han character and every full-width punctuation mark as a word. Thai and Lao go the other way: Word agrees with the space count, at 341 and 422, while the segmenter finds 2,331 and 2,411.
A space in Thai, Lao or Khmer is a caesura between phrases, and Chinese writes none
The Unicode standard UAX #29 (revision 49, 1 September 2026) says "reliable detection of word boundaries in languages such as Thai, Lao, Chinese, or Japanese requires the use of dictionary lookup or other mechanisms". Microsoft's page on text boundaries groups Thai with Japanese and Chinese as scripts without word spaces.
Yet the Thai, Lao and Khmer files are full of spaces, falling at the end of a phrase or clause like a pause in verse. In Khmer, Article 1 has 7 pieces between spaces and 30 segmenter words; in Lao, 5 and 42.
Microsoft's word processor gives each Han character and full-width comma a beat of its own
The help page "Show word count" states no rule for East Asian text, so we ran Word for Mac 16.114.3 on each file.
Word keeps a separate statistic, listed in its VBA reference as "Count of characters for Asian languages". It read 2,873 for Simplified Chinese, 2,691 for Traditional and 4,039 for Japanese, and our own count of Han characters, kana and full-width forms gives the same three. Word's word count is that figure plus everything else split at spaces: 2,873 plus 6 makes 2,879. The file has 2,675 Han characters, so 198 of its "words" are punctuation: commas ,, enumeration commas 、, full stops 。, semicolons, a colon and book-title marks 《》.
For Thai, Khmer, Lao and Burmese, Word matched the space split exactly on every UDHR file, and came within 1.6% of it on Wikipedia articles we used as a cross-check (CC BY-SA 4.0, revisions recorded). Setting the document language changed nothing. Word for Windows, Word on the web and Google Docs were not tested.
The sentence usually quoted as Word's rule, "In East Asian languages, each character counts as a word", is from the 2003 reference for the NumChars property of Microsoft Office Document Imaging, an OCR component. It describes characters recognised in a scan; Word is not its subject.
The browser's dictionary segmenter divides the line by its own lexicon
The browser segmenter, Intl.Segmenter, runs on ICU, whose user guide says ICU "provides dictionary support for word boundaries in Chinese, Japanese, Thai, Lao, Khmer and Burmese", applied automatically when such text appears. In our runs a language tag, or none, gave identical counts.
The count is reproducible, but nothing makes it the true one. In the Chinese Wikipedia article 汉语 (revision 95316942, CC BY-SA 4.0) it cut 五分之一 (wǔ fēn zhī yī, one fifth) into 五分 (wǔ fēn) and 之一 (zhī yī, one of). The Simplified Chinese UDHR file works out at 1.67 Han characters per segmenter word, on our arithmetic. Our reading-time guide found 1.26 to 1.48 across whole books, and Brysbaert's 2019 meta-analysis of reading rate converts Chinese at 1.5. No constant turns characters into words.
The Thai file carries four bracketed editorial corrections beside the misspellings they fix, and the segmenter counts both. The 2,331 is the file as published; resolved, it is 2,325.
Safari's engine reads a numeral as a rest where Node's reads a word
We ran the same call in Node.js 26.13 (V8, ICU 79.1) and in JavaScriptCore, Safari's engine, on macOS 26.7.2. JavaScriptCore marks a number written in digits, Arabic, Khmer or Burmese, as not word-like; Node counts it. In the other eight files that removes 30 to 34 words, mostly article numbers. The Chinese files number their articles in Han characters, so the gap is 0 for Traditional and 3 for Simplified. No other browser was run.
Vietnamese scans by the syllable under every method we tried
The Vietnamese file gives 2,502 by spaces, by Node's segmenter and in Word; Safari's engine gives 2,468, the numeral gap again. Microsoft's text-boundaries page explains why they agree: "spaces separate syllables rather than words."
Indonesian, a spaced control, moves too: 1,640 by spaces and in Word, 1,726 by segmenter, all 86 extra from splitting hyphenated tokens. The Badan Bahasa spelling guide EYD V, "Tanda Hubung", joins a reduplicated form such as anak-anak with a hyphen, without saying whether that makes one word. Of the 86, 65 are exact repetitions such as hak-hak, 80 if affixed forms like bersama-sama are included. Malay was not tested.
This counter was rebuilt to print its method beside the count
Until 9 November 2026 our counter split at spaces only, giving the Simplified Chinese file 97 words and a reading time under a minute. Its keyword check found 自由 (zìyóu, freedom) zero times in a text that contains it 31 times, because JavaScript's \b boundary recognises only ASCII letters, digits and the underscore, and Thai, Khmer and Lao keywords failed alike. It counted characters in UTF-16 units, so a skin-toned thumbs-up and a family emoji, with a space between, read 13.
The revision of 9 November 2026 counts at spaces, and with the browser's segmenter inside Chinese, Japanese, Thai, Lao, Khmer and Burmese. For unspaced text the word counter prints three rows: the space split, its own count and the Word for Mac rule. Pasting the Simplified Chinese UDHR file, prefatory note included, gives 97, 1,606 and 2,879, the same in Node and in Safari's engine. An unspaced keyword is counted wherever it occurs, and the result says so: 自由 gives 31. Text and keyword are normalised to NFC, so quyền, missed entirely in the partly decomposed Vietnamese file before, is found 67 times. Mostly Thai, Lao, Khmer or Burmese text shows "Not estimated" for reading time, and Chinese and Japanese are timed at reading-aloud rates, which run slower than silent reading.
Each script has its own metre, and a word count reports only one of them
For Chinese and Japanese, the character count is the most stable figure, depending on neither a dictionary nor a word processor; any word count needs its method stated. For Thai, Lao, Khmer, and Burmese, a space-based count is really a phrase count, while a segmenter count is only valid if you name the engine and version. For Vietnamese, all the tools we tested count syllables.
Characters have their own trap, measured in what a character limit does to Khmer and why your name does not fit in the box. Sorting is a third question, with no Thai collation in MySQL, as one collation for Southeast Asia sets out. A reading time built on 2,879 Chinese words differs from one built on 1,606, though both describe one page.