Open-source · BSD-2-Clause

Every authority.
Every script.
One map.

Interscript encodes the world's romanization systems — BGN/PCGN, ISO, UN, ALA-LC, ODNI, ICAO, DIN — as machine-readable maps that produce byte-identical output across Ruby, TypeScript, and the browser. 289 systems, 35 authorities, 31 script pairs, all checked into git and tested against the published reference documents.

Try live transliteration → Browse 289 systems

Systems: 289
Authorities: 35
Scripts: 31
Runtimes: 3

Warming up the engine…

Cyrillic1/3

Антон

BGN/PCGNUkrainian · 2019

Антон

Arabic1/3

عَبد الله

BGN/PCGNArabic · 1956

عَبد الله

Devanagari1/3

महात्मा

UNHindi · 2016

महात्मा

Han1/3

台北

Academia SinicaTongyong · 2002

台北

Ethiopic1/2

ኢትዮጵያ

ALA-LCAmharic · 2011

ኢትዮጵያ

Greek1/3

Αθήνα

ISOGreek · 843/1997

Αθήνα

Authority-backed. Real documents. Real test vectors.

BGN/PCGN
ISO
UNGEGN
ALA-LC
ODNI
ICAO
DIN
MOFA
Academia Sinica

What it is

Transliteration as data, not code.

Most romanization libraries hand-roll one system per language. Interscript encodes the entire corpus of authority-backed systems using a single declarative map format. Maps are checked into git, versioned, independently testable, and compiled to a typed JSON IR that any runtime can execute.

Author the map

A declarative Ruby DSL captures the rules of an authority document verbatim, including all special cases and notes.

stage :main do
  parallel do
    sub "щ" => "shch"
    sub "ъ" => "ʺ"
    sub "Антон" => "Anton"
  end
end

Compile to JSON IR

The Ruby Compiler::JsonIR emits a typed, versioned intermediate representation any runtime can load.

{
  "schemaVersion": 1,
  "systemCode": "bgnpcgn-ukr-Cyrl-Latn-2019",
  "stages": [
    { "kind": "stage", "name": "main",
      "rules": [ /* parallel sub-rules */ ] }
  ]
}

Run anywhere

Ruby gem, TypeScript package, and in-browser widget — all consume the same IR and produce the same bytes for the same input.

Interscript.transliterate(
  "bgnpcgn-ukr-Cyrl-Latn-2019",
  "Антон"
)
# => "Anton"

A handful of the catalogue

From 289 systems. Five shown.

Each system has a published authority document, machine-readable map, exhaustive test vectors, and live in-browser preview. The five below are running right now on this page.

BGN/PCGN Ukrainian · 2019 Антон Anton bgnpcgn-ukr-Cyrl-Latn-2019
BGN/PCGN German · 2000 Tschüß! Tschueß! bgnpcgn-deu-Latn-Latn-2000
ODNI Russian · 2015 привет мир privet mir odni-rus-Cyrl-Latn-2015
ALA-LC Amharic · 2011 ኢትዮጵያ ʼiteyop̣eyā alalc-amh-Ethi-Latn-2011
UN Tamil · 1972 தமிழ் tamil̮ un-tam-Taml-Latn-1972

Browse all 289 systems →

The neural layer

Where maps end, models begin.

Some conversions have no authority document: restoring the diacritics a scribe left out, reading Thai or Urdu as phonemes. Those ship as checksummed neural artifacts under the same discipline — measured metrics, teacher next to student, identical bytes from every runtime.

diacritization heb-diac-1.0 DER 17.46 beam-4
g2p tha-g2p-base-1.0 PER 9.19
diacritization urd-diac-1.0 CER 3.74
g2p · client tha-g2p-small-1.0 int8/int4 · 246/193 MiB · PER 2.85

The model catalogue →

Get started

Install in one line.

Open source under BSD-2-Clause. Maintained by Ribose Inc.

# Ruby
gem install interscript

# TypeScript / JavaScript
npm install interscript-ts

# Or load a single map's IR directly
curl interscript.org/maps/<system-code>.json

Every authority. Every script. One map.

Install in one line.

Every authority.
Every script.
One map.