How the Syllable Counter Works (and Where It Is Wrong)

Syllables do most of the work in both readability scores on this site. Reading Ease and Flesch–Kincaid Grade combine average sentence length with average syllables per word, and the coefficients weight the second far more heavily: a rise of 0.01 in syllables-per-word costs about 0.85 Reading Ease points, while one extra word of sentence length costs about 1. Yet that total is the one estimated number in the pipeline. The readability checker does not look words up; it applies four spelling rules to each one and adds up the results. This guide states the rules, runs a dozen words through them, shows where they fail, and explains why a text-level score survives errors that a single word does not.

The four rules, in order

The counter handles one word at a time and never looks at a word's neighbours, its position in the sentence, or its meaning.

1. Reduce the word to English letters

The word is lowercased and every character that is not one of the letters a to z is removed: apostrophes, hyphens, digits, accented letters, and letters from any other script. "Don't" is counted as "dont", "e-mail" as "email", and "café" as "caf". This is the step that makes the heuristic English-only by construction. If nothing survives it, as with "500", the word contributes 0 syllables.

2. Count the vowel groups

The letters a, e, i, o, u and y are vowels — y always, whatever its position. A run of consecutive vowel letters is one group, however long. "Beautiful" has three groups, eau, i and u, and comes back as 3; "yellow" begins with ye as a single group and comes back as 2.

3. Subtract one for a silent final e

If the word ends in e, does not end in le, and has more than one vowel group, one group is removed. All three conditions must hold. "Cake" has two groups and ends in e, so it is counted as 1. "Table" ends in le, so it keeps 2. "The" has only one group, so the third condition leaves it at 1.

4. Never return less than one

Any word with at least one letter left is worth at least one syllable. "Rhythm" contains no vowel letter other than y and is counted as 1; "strengths" has a single group and is counted as 1.

That is the whole algorithm: no exceptions list, no lookup table, no model of stress — which is why it runs on the longest text the tool accepts without a network request, and why it is wrong in the ways described next.

Twelve words through the counter

Every count in the table comes from the same countSyllables function the checker runs on your text. The last column gives the dictionary syllabification for the five words the methodology page spot-checked; the other seven are shown for the mechanism alone and make no claim about how they are spoken.

Word Vowel groups seen Final-e rule Engine result Dictionary (spot-check)
beautiful eau · i · u — 3 3 — correct
readability ea · a · i · i · y — 5 5 — correct
strengths e — 1 1 — correct
created ea · e — 2 3 — under-counted
business u · i · e — 3 2 — over-counted
table a · e exempt (ends in -le) 2 —
whole o · e exempt (ends in -le) 2 —
cake a · e subtracted 1 —
the e kept (only one group) 1 —
yellow ye · o — 2 —
rhythm y — 1 —
don't o (apostrophe removed first) — 1 —

The two ways it is wrong

Under-count: vowels that touch but are spoken apart. In "created" the letters e and a sit side by side, so rule 2 sees one group where speech has two — the engine returns 2 against a dictionary count of 3. The rule cannot tell one vowel sound from a boundary between two, so any word with neighbouring vowel letters spoken apart loses a syllable the same way.

Over-count: vowels that are written but not spoken. In "business" the i is silent, yet it is a vowel letter and forms a group of its own; the engine returns 3 against 2. Rule 3 handles a silent e only when it is the last letter of the word, so a silent vowel anywhere else is always counted.

The final-e rule is spelled, not spoken. It fires on every final e outside an le ending, pronounced or not, and exempts every le ending: "whole" is exempt on the same terms as "table" and comes back as 2. It is right exactly as often as English spelling is regular.

Two properties follow: the errors are deterministic — "created" is counted as 2 every time it appears, in every draft — and they run in both directions.

Why a text-level score survives per-word errors

Reading Ease never sees an individual word's count, only the total divided by the number of words, so when some words are over-counted and others under-counted the total moves less than the sum of the errors.

Take "The business created a plan." The engine counts 8 syllables in 5 words. Correct the two documented miscounts by hand — "business" to 2, "created" to 3, everything else left as the engine has it — and the total is 8. The two errors cancel exactly: the Reading Ease the checker reports, 66.4, is the figure the corrected count produces, 66.4.

Now take "She created what he created." Both miscounts point the same way. The engine counts 7 syllables in 5 words; the corrected total is 9. That is a gap of 2 syllables across only 5 words — 0.40 syllables per word — and at about 0.85 points per hundredth it moves Reading Ease from the reported 83.3 to 49.5. Nothing cancelled, because nothing was there to cancel against.

The lesson is not that the heuristic is unreliable but that its reliability depends on length. Over a passage of ordinary prose the two kinds of miscount are mixed together, and the methodology page's finding — totals typically within a few points of hand computation — describes that mixture. Over a 5-word sentence, one repeated word decides the score. That is one more reason, alongside those in the Reading Ease guide, to distrust any score on a very short text; the Flesch–Kincaid grade is built from the same syllables-per-word figure and inherits the same error.

Words that carry no syllables

Rule 1 has a consequence unrelated to pronunciation. The tokenizer counts a run of digits as a word — "12" is a word, and "2,500" is two — but the syllable counter strips digits before it counts, so every number written in figures is a word with 0 syllables. Numbers therefore lower syllables-per-word and raise Reading Ease, not because they are easy to read but because they are invisible to the counter.

Compare "The invoice lists 12 chairs and 4 tables." with "The invoice lists twelve chairs and four tables." Both are 8 words in 1 sentence. With figures, the engine counts 8 syllables and reports a Reading Ease of 114.1 and a grade of −0.7; spelled out, it counts 10 syllables, a Reading Ease of 93.0 and a grade of 2.3. The sentence did not become harder to read; two words became countable. Hyphens and apostrophes are stripped for syllable counting but keep the word whole for word counting, so "don't" is one word of 1 syllable — see what counts as a word.

Why text in other languages cannot be scored

Rule 1 keeps only the letters a to z. An accented letter is not among them, so "café" is reduced to "caf" and counted as 1 — a count of what was left, not of the word. A word in a script with no a-to-z letters at all is reduced to nothing and counted as 0, as "книгу" is here.

The readability function still runs on such text, because the Unicode-aware tokenizer counts the words correctly; what it cannot do is count their syllables. Score "Мы читаем книгу дома." and the engine finds 4 words, 1 sentence and 0 syllables. With syllables-per-word at zero the 84.6 term vanishes, and the formula returns a Reading Ease of 202.8, a grade of −14.0 and the band "very easy — about 5th-grade level". Those figures describe a formula with one input missing, not the sentence; the checker reports them because the scale is unbounded and it does not clamp, but they are a diagnostic, not a score. For a Latin-script language the failure is quieter but no smaller: accented letters are discarded and the rest grouped by English spelling rules, so the number that appears is not a measurement of that language.

How to read the output

  • Treat the syllable total, and the scores built on it, as a strong signal with a known error rather than a laboratory measurement.
  • Compare drafts of the same text rather than absolute scores. Every miscount is deterministic, so a revision that changes a handful of words changes the score only through those words; the rest carries the same errors in both versions and drops out of the comparison.
  • Prefer passages to fragments. Cancellation needs enough words for both error directions to appear.
  • Do not score figures, other scripts or accented text and expect a readability verdict. Spell numbers out if you want them counted as spoken, and keep the checker for English prose.

For what the scores mean once the syllable count is in hand, see the Reading Ease and Flesch–Kincaid grade guides, or the full list of guides.

Frequently asked questions

Why does the checker count "created" as 2 syllables?

The counter works on vowel groups, and the letters "ea" in "created" form one group even though they are spoken as two syllables. The engine returns 2; dictionary syllabification gives 3. This is one of the two error directions documented on the methodology page. The other is "business", where a vowel letter that is not pronounced is still counted: 3 from the engine against 2 in a dictionary.

Does the syllable counter use a dictionary?

No. It applies four spelling rules to each word — keep only the letters a to z, count runs of vowels, subtract one for a silent final e outside -le endings, never return less than one — and there is no lookup table or exceptions list. That is why it runs instantly on any length of text with no network request, and also why it is wrong on words whose pronunciation departs from their spelling.

Do numbers written as digits count as syllables?

No. A digit is not one of the letters a to z, so it is removed before counting and the word contributes 0 syllables while still counting as a word. In the worked example on this page, writing two quantities as figures instead of words moved the Reading Ease from 93.0 to 114.1 with no other change to the sentence.

Can I score text written in another language?

Not meaningfully. The counter strips every character outside a to z before it counts, so a word in a non-Latin script has nothing left and returns 0 syllables, and an accented word such as "café" loses its accented letter and is counted as 1. The readability formula still produces a number for such text, but the number describes the formula running with one input missing, not the text.

If individual words are miscounted, can the readability score be trusted?

As a signal, yes; as a measurement, no. Miscounts in opposite directions partially cancel across a text: in "The business created a plan." the engine total of 8 matches the hand-corrected total of 8 exactly, despite two wrongly counted words. Miscounts that all point the same way do not cancel, and in a short text they dominate: in "She created what he created." the engine total is 7 against a corrected 9, which alone moves Reading Ease from 83.3 to 49.5. Longer passages give both directions room to appear.

Why do other tools report a different syllable total for the same text?

Because each uses its own heuristic, and none of them is a dictionary. Small differences between tools are expected and are not evidence that either is broken. The useful question is which rules a tool applies, since that tells you which of your words it will get wrong — which is why this guide states ours in full.

Every number in this guide was computed at build time by the same engine that runs the readability checker. When you use the tool itself, scoring happens locally in your browser — nothing you paste is transmitted or stored. Syllable counting is heuristic, as documented on the methodology page.