Syllable counter

What is a syllable counter?

A syllable counter takes text and reports how many syllables it contains — the beats you would clap out saying it aloud. People reach for one to fit a haiku to 5-7-5, to check a line of verse scans, to keep a lyric singable, or to see whether a piece of writing leans on long words. This one takes a single word or a whole page, counts it in English, Spanish, Portuguese or French, and shows the count for every word alongside the total.

The interesting part is not the counting. It is that for a surprising number of English words, there is no one number to report. The CMU Pronouncing Dictionary — the standard machine-readable record of how English words are said, built at Carnegie Mellon — lists more than one pronunciation for 1,468 ordinary words, and those pronunciations disagree about the syllable count. “Every” is two or three. “Our” is one or two. “Fire” is one or two. “Family” is two or three. Both readings are correct English; which one you use depends on how carefully you are speaking.

Weighted by how often words actually occur in prose, that is 1.39% of running words — roughly one word in seventy-two. So it is not a curiosity confined to rare vocabulary, and a counter that prints a single confident number for a paragraph containing “every” and “different” is not being more accurate than one that prints a range. It is being less honest.

How to use it

1. Pick the language first. The rules are genuinely different — English needs a dictionary because its spelling does not determine its pronunciation, while Spanish, Portuguese and French can be counted from their vowels. Switching the language re-counts the same text under the new rules.
2. Type or paste your text. A single word works, and so does a whole page. The totals update as you type, and the word-by-word breakdown underneath shows what each word contributed, so you can see immediately which one is throwing your line count off.
3. Read the total as a range when it shows one. If the count appears as two numbers with a dash, your text contains at least one word the dictionary gives two readings for, and those words are listed separately with both counts. Take the lower number for careful, compressed speech and the higher one for a fully articulated reading.

Why English needs a dictionary and Spanish does not

English spelling does not determine English pronunciation, so counting syllables from the letters alone is a heuristic and cannot be anything else. The cleanest demonstration is a pair of words with an identical ending: “alone” is two syllables and “abalone” is four. No rule written over the letters gets both, and this counter's letter rules get the second one wrong — they say three. That gap is exactly why a table of known answers has to exist underneath.

So the English counting works in two layers. Where the CMU dictionary has an answer for a common word, that answer is used, drawn from a table of 1,050 words the letter rules miss. Everything else falls through to the rules: silent final e, silent -ed, the vowel groups, and the hiatus patterns where a vowel run really is two beats. Counting the stressed vowels in a CMU pronunciation is an exact syllable count, which is what makes the dictionary a real check rather than a second opinion.

Spanish, Portuguese and French are a different situation, and it is worth saying why rather than treating all four the same. Their orthographies are close to phonemic: a vowel group with a diphthong rule genuinely tells you the syllable count. So “biblioteca” is four because “io” is a diphthong, “aeropuerto” is five because a and e are both strong vowels and open separate syllables, and “país” is two because the accent on the weak vowel breaks the diphthong. French needs its silent endings stripped first — “parlent” is one syllable, not the three its spelling suggests — and its digraphs collapsed, so “beau” is one and “oiseau” is two. None of the three needs an exception table.

Honest limits

How accurate is it? That question has no single answer either, and the reason is worth stating because it is where most published accuracy figures quietly mislead. Measured against the whole CMU dictionary, the letter rules plus the exception table get 88.79% of words exactly right and 99.56% within one syllable. Measured over only the distinct words that occur in real prose, 92.70%. Weighted by how often those words actually occur — which is what you experience typing ordinary text — 98.40%. The first figure is the one a paper would quote and the last is the one that describes using the tool, and the gap between them is entirely the dictionary being full of proper nouns and foreign names while prose is full of short common words.

Put the last figure beside the ambiguity figure and you get the honest summary of what this tool is: it is wrong about 1.6% of the time on ordinary text, and about 1.4% of the time the question itself has two right answers. Those are different words — where the dictionary gives two readings, the counter matches one of them 97.68% of the time — but they are the same size, which means the residual error is roughly as large as the genuine uncertainty in the language.

It counts syllables; it does not divide words into them. Where a syllable boundary falls is a separate question from how many there are, and the usual computational answer — the hyphenation patterns from TeX that word processors use — answers something else again: where a line may legally be broken, which deliberately omits perfectly good syllable breaks to avoid ugly typography. Printing a division built on those patterns and calling it syllabification would be a claim this tool cannot defend, so it does not make it.

The full pronunciation dictionary is not bundled. As a word-to-count table it is 353 KB compressed, which is more than twice the largest download this site is willing to make a visitor wait for, and trimming it to whatever happens to fit would be an arbitrary cut rather than a boundary anyone would recognise. What is bundled is the 1,050-word exception table for everyday vocabulary and the 1,468-word disagreement table, together under 20 KB.

Arabic is not offered, and the reason is the same one that stops this site scoring Arabic readability. Every rule here reads vowels off the spelling, and Arabic is normally written without the diacritics that mark short vowels — so where the syllables fall is not recoverable from the letters on the page. A number produced anyway would be a guess wearing the clothes of a measurement, and an option that silently does that is worse than an option that is missing.

One more thing the tables were cleaned of, because it would have been visible: the dictionary also stores the spelled-out reading of some abbreviations under an ordinary word's spelling. “Us” shares a spelling with U.S. and “am” with A.M., and taken at face value those two looked like the commonest disputed words in English. They are not disputed at all; they are two different words wearing one spelling. 124 entries of that kind were removed before anything shipped.

Why is it free?

Because it costs nothing to run. The counting happens in your browser, on your own machine — the dictionary tables are part of the page. Nothing you type is uploaded, no server sees your text, and there is no account, no limit on length, and no watermark on anything you copy out.

The script that reads the pronunciation dictionary and builds the tables is kept in the repository alongside the data it produces, and the test suite checks the counts against that dictionary rather than against expectations written by hand. That is the difference between a number this page can defend and a number it merely asserts.