I built the first Tetun translator in 2014. Here's what I built next.
From a weekend language-learning hack to 57,000 users. Twelve years later, here's the rebuild — with AI, Timorese voice audio, dual orthography, and a 90,000+ sentence parallel corpus.
2014 — the weekend project
I moved to Timor-Leste in the early 2010s and started trying to learn Tetun (pronounced teh-TOON). The resources available online were fragmentary — a handful of dictionaries, a few grammar chapters, no way to get a fast gloss for a phrase I'd just heard in the taxi. So in 2014 I built what was, as far as I know, the first public English ↔ Tetun web translator and hosted it at translate.tetumdili.com.
It was nothing fancy: a Python back-end, a curated English–Tetun glossary of a few thousand headwords, phrase-match lookup for longer strings. I launched publicly in 2015. Within a year it was in active daily use by expats, Timorese diaspora, development workers, and — to my surprise — by a lot of Timorese students.
2017 — 57,000 users, top TL website
By early 2017 the translator had served more than 57,000 users globally, around 700 daily users, and was ranked among the most-visited websites in Timor-Leste. It wasn't monetised; it wasn't advertised; there was no app. People found it because Google did. I kept adding vocabulary and fixing phrase issues as people wrote in.
In March 2019 the companion iOS app shipped. It used the same backing corpus and the same phrase-match logic. Same URL, same data, now in your pocket.
Then machine translation happened
Google Translate added Tetun in 2024. Generic large language models started producing plausible Tetun output. On paper, a handcrafted translator looked less necessary. In practice, the opposite — generic MT systems still treat Tetun as a low-resource language. They hallucinate loanwords, miss aspect markers, confuse INL and DIT orthography, and produce output that reads like Portuguese scaffolding with Tetun words bolted on.
The right answer isn't pure MT. The right answer is grounded translation: draft with a capable model, then check every sentence against curated linguistic data, and lint the output against the rules that actually matter in Tetun. That's the rebuild.
2026 — the rebuild
Twelve years in, here's what's under the hood:
- 28,000+ active English → Tetun glossary rows, sourced from the tetun.org dictionary (Catharina Williams-van Klinken, DIT), DIT's justice, health, tourism, and IT glossaries, and the INL Matadalan Ortográfiku (2003).
- 90,000+ parallel English–Tetun sentences from MADLAD-400-Tetun, the timor-leste.gov.tl corpus, and DIT publications. BM25 retrieval at query time so the translator has real examples to ground on.
- 79 orthography and grammar rules, coded as a lint catalog and applied to every output. See the full list on the grammar page.
- Dual orthography support — both INL (official, Decree 1/2004) and DIT (academic), with a converter between them. Explained in full on the orthography page.
- Native Timorese voice audio. Six Timorese voices — a narrator, a young woman, a young man, and more — via ElevenLabs. Tap the speaker icon on any Tetun output or dictionary entry.
- Domain modes — general, justice, health, tourism, government. The translator weights its corpus retrieval and glossary toward the domain you select, so legal vocabulary doesn't bleed into tourism text and vice versa.
- Review mode and alternatives. The translator surfaces two or three phrasings with rationales, and a review pass that audits its own output against the glossary, corpus, and linter — then flags mismatches.
- Programmatic dictionary. Every English headword has a dedicated page with audio, INL and DIT forms, and example sentences from the corpus. 14,000+ dictionary pages live.
What's different from other Tetun tools
Most Tetun translators are one of two things: a dictionary with a search box, or a pass-through to a generic MT model. Neither is enough. Good Tetun output needs both the lexical grounding of a curated dictionary and the sentence-level flexibility of a capable model — plus a linter that catches what neither of them will.
That's what this rebuild does. It's the same project I started in 2014 — just with twelve years of linguistic data under it and the model pool that finally makes it viable at sentence scale.
Free, no ads, no signups
The translator, dictionary, grammar catalog, orthography converter, and audio are free to use. No account required. No advertising. Source data citations are on every page.
Tony Franklin is a software engineer based in Timor-Leste. He built the first English ↔ Tetun online translator in 2014 and rebuilt it in 2026.