How a computer learns Hebrew vowels

October 18, 2026

Three letters, three different words. What the small dots and lines tell us, and what we found when we counted every vowel mark in the Tanakh: 304,094 words. An article for beginners.

Here are three Hebrew letters: ספר. How do you read them? It could be סֵפֶר (sefer, a book), סַפָּר (sapar, a barber) or סִפֵּר (siper, he told a story). Same letters, three words. The whole difference is a few small dots and lines: the vowels. In this article we look at what they are, which patterns hide inside them, and how a computer can learn to add them by itself.

What is nikud?

Hebrew letters are mostly consonants. The vowels are small marks added around the letters, called nikud, which means "dots". Here they are on the letter bet:

Notice two things. Some sounds have two marks, like qamats and patah for a: they once sounded different, and today they mostly sound the same. And sometimes a letter helps, like the vav in שָׁלוֹם. There is one more dot, the dagesh, a dot inside the letter: it turns v into b (ב → בּ).

We counted every vowel in the Tanakh

To learn how vowels behave, we did something simple but big: we took the whole vocalized Tanakh, 304,094 words, and counted every single mark.

The winner is the sheva, the two tiny dots that are often silent. Right behind it is the qamats, an a sound. Add the patah, the other a, and the a sounds alone are about a third of all vowels: Hebrew loves a. The rarest is the hataf qamats, fewer than 1 in 1,000 vowels.

Seven patterns we found

1. Neighbours give vowels away

Guessing a vowel with no clues at all means choosing from about nine options. But if you know the letter, the vowel just before it and the letter right after it, only about two or three are left. Vowels are not random: their neighbours give them away. This is the secret behind a computer nakdan.

2. Vowels like to rhyme

Many words repeat the same sound: melekh, eretz, both e–e. And a short vowel before the throat letters א ה ח ע tends to match the short "hataf" vowel after it: a segol before a hataf-segol is 13 times more common than chance, as in נֶאֱמָן (ne'eman, faithful). Linguists call this vowel harmony.

3. Spelling changed over the centuries

The o sound can be written with a dot alone (קֹל) or with a vav and a dot (קוֹל). In the five books of the Torah the vav is used only about a third of the time. In books usually dated later, like Esther and Daniel, it is used more than half the time.

Without anyone planning it, the spelling tells us something about the history of the books.

4. Words change at the end of a verse

At the end of a verse, where the reader pauses, about 1 word in 4 (25%) changes its vowels. The most common change: a patah becomes a qamats. מַיִם (mayim, water) inside a verse becomes מָיִם, with a longer, fuller a, at its end.

5. Little rules you can count

The letters ב ג ד כ פ ת get a dagesh 84% of the time at the start of a word, but only 15% of the time right after a vowel. And about 2,599 words end with a "hidden" patah, like רוּחַ (ruach), where you say the a before the last letter.

6. The next letter changes “and”

Hebrew glues “and” to the front of a word as a single letter, ו, normally pointed with a sheva: וְ. Two things turn it into a shuruk, וּ: the next letter being ב, מ or פ, the three you make with your lips (וּבֵין uvein, and between; וּמִן umin, and from), or the next letter already carrying a sheva of its own (וּלְיָמִים ulyamim, and for days). Between them they explain every shuruk vav in the Tanakh: 5,535 before ב, מ or פ, plus 2,966 before a sheva, make 8,501 — and we found not one shuruk vav outside these two cases.

The rule is less tidy in the other direction, which is just the kind of trap a computer has to learn. Before ב, מ or פ the shuruk wins 5,535 times out of 5,685, about 97.4%, once we set aside the name וׇפְסִי (Vophsi, Numbers 13:14), whose vav is not “and” at all; the other 150 take a qamats instead, where the stress falls right after the vav, as in תֹהוּ וָבֹהוּ (tohu vavohu, Genesis 1:2) — and never once, in 488 cases, when the vav’s word is tied to the next by a maqaf. A following sheva is not enough by itself either, because the storytelling “and then” vav stands before a sheva too, and takes a patah: וַיְהִי (vayehi, and it came to pass). And Vophsi is not alone: וָוֵי (vavei, the hooks of the pillars) opens with a vav that belongs to the word itself. Telling those apart from a prefix is part of the job.

7. The tune decides some of the pointing

The te'amim — the small marks above and below the words that carry the chant and tell the reader where to pause — are usually thought of as music and punctuation. They also decide part of the pointing. The sharpest case is the dagesh — the dot inside a letter at the start of a word. A dagesh is not a vowel; it hardens a consonant, turning a v back into a b. But it is the clearest thing to count. After a word that ends in a consonant it is simply always there: we counted 29,109 of them, 100% with the dot. After a word that ends in a vowel it looks like a coin toss, 56.5% of 19,201. But it is not a coin toss at all. If the previous word carries a separating accent, the dot is there 98.8% of the time; if it carries a joining accent, only 9.1%; and when the two words are tied together with a maqaf, 7.8%. Knowing the tune raises the guess from 56.5% to 95.2%.

And it really is the vowels too, not only that dot. We took every word in four books the computer had never read whose letters allow more than one reading, and sorted them by what the readings actually disagree about. Where they differ only in their vowels, knowing the accents raises the answer from 72.95% to 76.42% — a bigger gain than for the dagesh, where it is 81.20% to 82.95%. The pausal change of pattern 4, patah becoming qamats, is exactly this: a vowel, decided by where the chant comes to rest.

This is the sharpest thing we found, and it changes what "a computer cannot know" means. Hand a program plain letters and some vowels become genuinely unguessable — not because Hebrew is undecidable, but because the answer was written in a channel we deleted. The Masoretes put the chant and the vowels in the same system, and the two hold each other up.

What Rashi already knew

We counted with a computer. Rashi counted in his head, nine hundred years ago, and in about 700 of his comments he talks about the vowels themselves — the names of the marks, the dagesh, and where the stress falls.

One of those remarks we were able to check exactly. On כִּי רָבָּה (Genesis 18:20) he writes that every רבה in Scripture is stressed on the ב — except this one, stressed on the ר, because here it does not mean “is great” but “has already grown great”. So we counted all 48 times the word appears in the Tanakh. Thirty-six are stressed on the ב. Eight are tied to the next word by a maqaf and carry no stress mark of their own. Three fall on the ה. And exactly one is stressed on the ר — Genesis 18:20. He made a claim about every verse in the Tanakh, without a concordance, and it holds.

His other remarks are the same kind of thing, and none of them is really about letters. קַשָּת, an archer, doubles its ש because it belongs to the craftsman pattern, like חַמָּר (donkey driver), גַּמָּל (camel driver) and צַיָּד (hunter). A word is pointed one way because it is a noun and another way because it is a verb. A dagesh can mark a letter that stands in for two.

That is the difference between him and our program. We can tell you how often a vowel follows the letter beside it. He could tell you why. When we tested our seven patterns on books the computer had never read, the ones that helped were about position and neighbours; the ones about spelling added nothing at all, because the letters already said it. What is missing is the thing Rashi had: knowing what the word is.

How a person does it

A trained reader vocalizes a new sentence far better than our program does. It is worth being precise about what they are doing, because it is not a faster version of counting letters. We sorted every spelling in the Tanakh that allows more than one reading, according to what its readings actually disagree about. The answer came out in three parts, and they are three different questions.

1. Which word is this? — 72.5% of the hard cases

Most ambiguous spellings are not one word with two pointings. They are different words. אדם is אָדָם (a man), אֱדֹם (Edom) or אֹדֶם (a ruby). קורה is a wooden beam or the verb “happens”. Nothing in the letters chooses between them. You choose by knowing what the sentence is about — and that is why this is the part a program built on spelling can never reach.

2. What is it built from? — the root and the pattern

Here is the surprise. A reader does not remember a similar-looking word and copy its vowels. They know the root, they recognise the pattern it has been poured into, and they build the vowels from the two. We measured how different those operations are: taking another form of the same root and copying its vowels gets about 1 word in 100 right. שָׁמַר, שֹׁמְרִים and יִשְׁמֹר share a root and share almost no vowels.

This is exactly what Rashi does. קַשָּת, an archer, doubles its ש because it belongs to the craftsman pattern — like חַמָּר, גַּמָּל, צַיָּד. He is not recalling a word. He is naming a pattern and building.

3. Where does it sit? — 8.7%

The rest is position: at the end of a verse the vowels lengthen, a maqaf ties two words into one, a prefix changes what follows it. This is the part the letters and the chant really do tell you, and it is the part our program handles best.

And no, this does not add up to perfection — not even for people. In one careful study, two native-speaker experts were asked to find known mistakes in text that had already been proofread by experts. One found 62% of them and the other 43%, and they often disagreed about which were mistakes. The proofread text itself still contained errors. Reading traditions differ too. The skill is not never being wrong; it is knowing which of the three questions you are facing, and noticing when you are unsure.

How a computer learns to add vowels

A menaked (nakdan) is a tool that adds vowels to Hebrew text. A computer doesn't understand Hebrew, but it can learn from patterns, the same ones we just found. Our experimental menaked tries four steps, in order:

1. Have I seen this word next to these words? The letters ספר can be book, barber or told, but after "to the house of" it is almost always לְבֵית הַסֵּפֶר, school.
2. Have I seen this word at all? Then use its most common reading.
3. Is it a word I know with a prefix? Hebrew glues little words to the front: and, the, in, to, from. Take them off, read the word, put them back.
4. Never seen it at all? Hebrew pours a root into a template, so a word we have never met is usually some familiar word’s shape with different root letters. We look for the known word of the same length that matches in the most places, copy its vowels, then correct the ending from the endings we know. That lifts whole words from 3.8% to 13.4%, and marks from about half to about three fifths — which still means it is wrong far more often than right, so a guessed word should never be presented as an answer.

We built one, as an experiment. Two numbers matter, and they are very different. On a whole biblical book hidden from it while it learned — genuinely new text — it gets 66% to 77% of words completely right, and 83% to 89% of the marks. That is the number that matters. On Tanakh and Rashi text this site already holds it reaches 99.97%, but that is memory rather than skill: there it is quoting a vocalized source it already knows. The best tools today, built on large AI models, reach about 19 words in 20.

The menaked is not published on the site at the moment. It is good on Tanakh and poor on everything else, and the next two paragraphs explain why — that seemed worth writing down even though the tool is not ready.

No menaked is perfect, not even the best ones — and the reference texts they are scored against contain mistakes of their own, so a claim of 100% could not even be checked. That is why a good menaked lets you click any word and pick another reading: the computer does the boring work, and you stay the teacher.

Two limits are worth stating plainly, because we measured both. First, the patterns above describe the vowels well but mostly do not predict them. We tested all seven on books the program had never read: the ones about position and neighbours helped a little, and the ones about spelling added nothing at all, because the letters already carried that information. Combining all seven was worse than using the single best one. Second, when the program meets a word it has never seen, the right pattern is usually already sitting in its dictionary under some other word — 73.5% of the time — and it simply picks the wrong one. That is the whole gap, and closing it needs what Rashi had: knowing the word's root and form, not more counting of letters.

And it only knows the Hebrew of the Tanakh. Give it a modern sentence and it answers in Biblical Hebrew, which is often the wrong answer: קורה comes out קוֹרָה, a wooden beam, where modern Hebrew means קוֹרֶה, “happens”; לנחש comes out לְנָחָשׁ, “to a snake”, instead of “to guess”. Those are not bad guesses — they are correct Biblical Hebrew in the wrong century. For modern Hebrew use Dicta's nakdan instead.

Recap and a short quiz

Vowels are the small marks that tell us how to read Hebrew letters. The most common are the sheva and the qamats. And vowels follow patterns, strong enough that a computer can guess them right about 7 times out of 10 in a book it has never seen. Try to answer before you open:

Want to look for patterns yourself? The Tanakh search finds gematria values, skip codes and first-letter acrostics in every book of Tanakh.

Sources

• The count: the vocalized Masoretic text (MAM) via Sefaria, 304,094 words, and the vocalized Rashi commentary (190,008 words).
• Our menaked's accuracy: scored on whole books held out of its dictionary (Job, Esther, Ruth, Hosea), which is a harder and more honest test than hiding scattered verses.
• The rule behind pattern 6, and the qamats before a stressed syllable: Gesenius’ Hebrew Grammar, §104.
• Accuracy of the advanced tools: Nakdimon (Gershuni & Pinter) and Dicta Nakdan (Shmidman et al.).