Method

Searching Syriac, a script that leaves out its vowels

Ephrem the Syrian is catalogued as ʾAprém, Afrem, Ephraem and ܐܦܪܝܡ, and the script itself writes no vowels at all. We have added Syriac to tetrak-translit with a fold that keys on consonants alone, and this is what that buys, what it costs, and where we would welcome the help of people who read Syriac

28 October 2026 15 min read Stephen Masters

Ephrem the Syrian (c. 306–373 AD) was born in Nisibis, Mesopotamia, and later moved to Edessa. He was known for his vivid poetic sermons and liturgical hymns (madrāšê) and is venerated as a saint across Eastern and Western Christianity. As a result, he is among the most copied authors of the Syriac churches. He is also among the most translated: his works reached Armenian early, and the commentary on the Pauline epistles attributed to him is known only through its Armenian translation. The Mekhitarist fathers printed that translation at their press on San Lazzaro in Venice in 1836, as the third volume of his works in Armenian, Srboyn Epʻremi matenagrutʻiwnkʻ. The Armenian Institute’s library in London holds a copy, and the Austrian National Library has digitised one. So his name turns up in catalogues of Syriac manuscripts and of Armenian ones, and in a good many forms:

ʾAprém, as the Library of Congress romanises the Syriac;
Afrem, as the West Syriac tradition writes it (Mor Afrem), with the soft p pronounced as f;
Ephraem, the Latin form, in the Library of Congress heading Ephraem, Syrus, Saint, 303-373 (LC authority records disagree slightly with historians on birth date);
Ephrem, as most English writing has it;
ܐܦܪܝܡ, as the Syriac script writes it;
and Եփրեմ, in Armenian.

Set the Armenian aside for a moment and look at what the others share. Not the vowels: e, a, é and ae disagree about every one of them. The Syriac spelling has five letters, ܐ ܦ ܪ ܝ ܡ, and only three of them are consonants in any sense a reader of Latin would recognise: p, r and m. The first, alaph, is a glottal stop that most romanisations drop. The fourth, yod, is a consonant that is also doing the job of the vowel é. Everything else a reader sees in ʾAprém has been supplied by someone who knew how to pronounce the word, because the script itself is not particularly phonetic.

That is the problem Syriac poses to search, and this article is about how we have handled it in tetrak-translit, our transliteration library.

Why Syriac, and why now

We have been thinking about Syriac OCR for a while. Syriac manuscripts survive in large numbers, many have been photographed, and very few have been transcribed. In late October we attended a Zoom presentation by Dr Ani Shahinian, cataloguer of Armenian manuscripts at the Hill Museum & Manuscript Library (HMML). As she showed HMML’s online collections, we noticed what looked like a large Syriac collection beside the Armenian. The catalogue is published online, so we had a look at what it holds, and we saw the same thing we had seen with Armenian: titles and names recorded twice, once in the script and once in a romanisation, and a search that only finds a record if you type the form its cataloguer chose.

A recogniser is a long piece of work. A transliteration library that understands the script is a short one, and it is the part that makes the recogniser’s output, and every catalogue record, findable. So we started there.

What makes Syriac different

Syriac is a literary form of Aramaic, first written at Edessa, and the language of a large body of Christian literature from late antiquity onwards. Several things about the script matter for search.

It is written right to left

Like Arabic and Hebrew, Syriac runs right to left, and its letters join to their neighbours. Unicode stores the letters in the order they are read, so in the data ܐܦܪܝܡ begins with alaph, and the right-to-left display is the browser’s job, governed by the Unicode Bidirectional Algorithm. Mostly the browser gets it right. Where it goes wrong is a line that mixes directions: a Syriac title followed by an English one, joined with a separator, can come out with the pieces in the wrong order. The fix is to isolate each catalogue value, with HTML’s <bdi> element or dir="auto", so that the browser lays out each one by itself. We found that the hard way, putting Syriac titles on a page that had only ever shown Armenian and English.

It has no capital letters

Each Syriac letter changes shape as it joins its neighbours, but there is no upper and lower case. That is one less thing for a fold to do, but a library that reads romanisations back into the script has to know not to look for capitals that do not exist.

It is normally written without vowels

This is the big one. Syriac is an abjad: its twenty-two letters are consonants, and three of them, alaph, waw and yod, double as long vowels (the matres lectionis). The short vowels, and the quality of the long ones, are marked, when they are marked at all, by small signs above or below the letters, and there are two systems for doing so:

the West Syriac system, attributed to Jacob of Edessa, uses small forms of Greek vowel letters;
the East Syriac system, the older, uses dots in different positions.

Most text has neither. In the catalogue we looked at, almost none of the titles written in Syriac script carried a single vowel point.

And it has other dots too

Not every dot is a vowel. Two dots over a word, seyāmē, mark it as a plural. A dot above or below one of six letters (b g d k p t) says whether it is pronounced hard or soft, which is why in the West Syriac pronunciation of Fenqitho (ܦܢܩܝܬܐ) the ܦ (p) is f and the ܬ (t) is th. There are feminine dots and dots that tell homographs apart. None of these is a letter, and a fold has to ignore them all.

One encoding, three styles

Syriac is written in three main styles: Estrangela, the oldest, used in ancient manuscripts and now for headings; Serto, the cursive West Syriac style; and the East Syriac style. They look quite different on the page, but they are one script and one encoding. Nothing in the text says which style a manuscript used, so nothing in our library tries to tell them apart.

Not all Syriac is in Syriac

Syriac letters have also been used to write other languages, most of all Arabic: Garshuni. A catalogue of Syriac manuscripts is full of it, along with Malayalam, Turkish and other languages in Syriac script. Our library detects all of it as Syriac script, which it is. Whether it is Syriac language is a different question, and another question the script alone cannot answer.

What the Library of Congress table asks for

The romanisation used by English-language libraries is the ALA-LC Syriac table, in its 2012 version, the only one there has been. It is a careful table. It maps each consonant to one Latin letter (with ʾ and ʿ for alaph and ʿē), each vowel point to a vowel, and handles the places where yod and waw carry a vowel, so ܝ with the right point becomes í, and a final alaph after ā or e is dropped (zadíqā, not zadíqāʾ).

And then, for unvocalised text, it says this:

If the text to be transliterated has no vowels, supply them, starting from the Syriac vowels prescribed in J. Payne Smith’s Syriac English Dictionary.

That is an instruction to a cataloguer who reads Syriac. No table can follow it. To supply the vowels of a word you have to know which word it is, in which form, and in which tradition, since the table goes on to list where East Syriac differs from the dictionary.

So our library renders exactly what is written. A pointed text is romanised as the table says, vowel for vowel. An unpointed one is romanised as its consonants: ܡܠܟܐ is mlkʾ, where a cataloguer would write malkā. That is less pleasant to read, but it is true to the text, and it never invents a vowel that the manuscript did not write. Our library has the same rule for every script: a transliteration is a display of the text, and it must not guess.

We did look at doing better. Payne Smith’s dictionary is out of copyright and has been scanned, but the available text of it is not clean enough to build on, and a dictionary’s headwords do not cover the inflected forms a title is written in. Supplying vowels is a real project, one we would only take on alongside a Syriac recogniser, and one we would want to do with Syriac scholars rather than for them.

Folding on consonants

Display is only half of what the library does. The other half is the fold: a function that maps every spelling of a word to one key, so that one search finds them all. We have written about folding for Armenian, where the key is roughly the word as a Library of Congress cataloguer would romanise it, with the marks taken off.

That approach cannot work for Syriac, because the romanisations disagree most about exactly the thing the script leaves out. ALA-LC writes Penqítā for the book of hymns for the year’s feasts; the West Syriac tradition writes Fenqīthō, with o for ā and the soft consonants spelled out; a popular spelling drops the marks, Fenqitho; and the script writes ܦܢܩܝܬܐ. The vowels differ in every one. What they share is the consonants, so the Syriac key is the consonants:

Spelling	Key
ܦܢܩܝܬܐ	`pnkt`
Penqítā	`pnkt`
Fenqīthō	`pnkt`
Fenqitho	`pnkt`
ܐܦܪܝܡ, ʾAprém, Afrem, Ephraem, Ephrem	`prm`

The rules, in order:

Syriac script is rendered through the ALA-LC consonants, with any vowel points it carries, and Garshuni’s extra letters on the consonants they stand for.
Latin spellings have their two-letter consonants reduced: sh is šin, and the West Syriac soft consonants th, ph, kh, gh, dh and bh are their letters; f is p and c is k. Only Latin gets this step. The script is rendered one letter at a time, so ܕܗ is a d followed by an h, not a soft d.
q becomes k, and every vowel is dropped, with w, y, j and v, which are waw and yod in one spelling or another (Jacob, David).
Diacritics go, so ṭ meets t, ṣ meets s and ḥ meets h. Alaph and ʿē, ʾ and ʿ, go with them, as most romanisations drop them.

One letter needed care. Šin, ܫ, is š in ALA-LC and sh nearly everywhere else. If the key wrote it sh, then ܫ would meet ܣܚ, s followed by ḥ, a different pair of letters. So šin is written c in the key, a letter that step 2 guarantees nothing else will leave behind. Šḥímā, Shhimo and ܫܚܝܡܐ all fold to chm.

What it costs

A consonantal key is coarse, and that has costs we would rather state up front.

It over-merges. Words that differ only in their vowels, or only in alaph, ʿē, waw and yod, share a key. ܥܠ (on) and ܠܐ (not) both fold to l. The biblical Ephraim folds to prm, the same key as Ephrem. In a search over titles and names, most such collisions are harmless, because a search for a two-word title needs both keys to match, but a one-word search with a short key finds more than it should.

It under-matches more than the Armenian fold does. We measured the fold against that catalogue, which records most Syriac titles both in the script and in a romanisation or translation, by checking how often the two forms of a title reach the same key. For Armenian, the same measurement finds nearly every pair. For Syriac, it finds most of them, but noticeably fewer. The pairs it misses are mostly not failures of the fold but things a fold cannot do:

Names in another language. Edessa is the Greek name of the city the Syriac calls ܐܘܪܗܝ, Urhay; Nisibis is ܢܨܝܒܝܢ. Bar Hebraeus is the Latin form of Bar ʿEbrāyā, with an h the Syriac does not have and a Latin ending. No spelling rule turns one into the other. That is a job for a name authority, which knows that two names belong to the same place or person, not for a fold. Our test index already has one for people: when it is built, each creator is looked up in a snapshot of Wikidata’s names for the region, and the person’s names in every language are indexed beside the catalogued one. A search for Abul-faraj, or for the Armenian Գրիգոր Բար-Հեբրեոս, finds Bar Hebraeus’s works, though neither form appears in the catalogue. The same snapshot already holds the places, with Edessa, Urhay and ܐܘܪܗܝ under one entry and Nisibis beside ܢܨܝܒܝܢ, but the index does not use them yet: a place in a record is not a creator, and joining it needs its own rule for telling one Edessa from another.
Translated titles. A cataloguer’s Book of hymns is not a romanisation of anything.
Garshuni romanised as Arabic. Arabic romanisation writes j where Syriac has ܓ, g, so Jirjis and ܓܪܓܝܣ part company.

And it can intrude on other scripts. When a search index holds Armenian and Syriac records together, every query has to be folded both ways, because the index cannot know which script the reader had in mind. An Armenian query folded under the Syriac rules loses its vowels too, and can find Syriac records that happen to share its consonants. In our test index, a search for Avetaran, the Armenian for the Gospel, reaches Syriac records whose key is also trn. We keep the Syriac keys only on records that are Syriac, and rank a match on them below a match on the Armenian keys, so every Armenian Gospel book comes first and the Syriac strays fall to the end of the list. Ranking, not matching, keeps them in their place.

Two scripts, one search

The reward for all of this is the search the opening promised. In our test index over that catalogue, one search for Ephrem finds the Syriac manuscripts through ܐܦܪܝܡ and the Armenian ones through Եփրեմ, together, under the one name a reader is likely to type. The Armenian fold and the Syriac fold never meet. Each takes its own script’s spellings to its own key, and the query is folded both ways. Where a name authority links the two, an Armenian-script search for Եփրեմ can reach the Syriac records too.

What we would like help with

We are not Syriac scholars. We built the tables from the standard and the fold from the catalogue, and a fold built that way is bound to have mistakes that a reader of Syriac would see at once. If you read it, we would be grateful for:

Romanisations we have not met. Our test cases cover the Library of Congress table, the West Syriac spellings, East Syriac ones and popular ones, for liturgical books, authors, places and common words. There are certainly traditions we have missed: French and German scholarly spellings, the romanisations of particular churches or catalogues.
Places where the fold is wrong. Two spellings of one word that do not reach one key, or two different words that do and should not.
Better ways to supply vowels. If you know of a lexical resource that could, with the right permissions, support vowel suggestions for unpointed text, we would like to hear about it.

The test cases are published, in tests/data/syr_fold_cases.tsv, and the rules are documented in the library’s fold reference. The easiest way to tell us is an issue on the library, but a message to either of us on LinkedIn works too.