Digitizing Indonesia's indigenous languages, and where translators come in
Indonesia is the standard example in any discussion of linguistic diversity, and for once the standard example earns its place. The national language agency, Badan Pengembangan dan Pembinaan Bahasa, has mapped and verified more than 700 living languages across the archipelago. Only Papua New Guinea has more. The figure is provisional. Surveys are still running in the east, and the line between a language and a dialect is drawn there by fieldworkers, not by nature.

What makes Indonesia a useful case study is not the number. It is the range of conditions contained within a single national policy framework. Javanese has something like 80 million speakers. Several languages in Maluku have fewer than fifty. Both sit under the same ministry, the same curriculum rules, and the same Unicode Consortium.

Vitality is not the same as size
Badan Bahasa grades vitality on a scale running from aman, rentan, mengalami kemunduran, terancam punah, kritis, hingga punah. A dozen or so languages already carry the last label, most of them in Maluku and Papua: Tandia, Mawes, Kajeli, Piru, Moksela, and others recorded in the final years of a single elderly speaker.

The uncomfortable finding is at the other end of the scale. Household surveys and school data have shown Javanese and Sundanese losing ground in urban homes, where parents who speak the language fluently raise children in Indonesian. Eighty million speakers today do not guarantee eight million in three generations. Vitality is a transmission rate, not a headcount.

The Ministry's Revitalisasi Bahasa Daerah program, launched as one of the Merdeka Belajar episodes in early 2022, responded to this by integrating regional languages into schools through teacher training and student competitions in storytelling, poetry, and scriptwriting. It began with a handful of languages and has expanded each year. Whether it changes what happens at the dinner table is a separate question, and the honest answer is that nobody knows yet.

The script problem
Indonesia is unusual in how many of its writing systems survived into the digital era, and how badly most of them were served by it.

Unicode has been kind, eventually. Lontara came in around 2005, Balinese in 2006, Sundanese in 2008, Javanese in 2009, Batak in 2010, Makassar as recently as 2018. Encoding is the precondition for everything else. A script without codepoints cannot be searched, indexed, spellchecked, or entered on a government form.

Encoding alone does not finish the job. A large share of Javanese-script material typed over the past twenty years used legacy fonts that map aksara onto Latin codepoints. Open one of those files without that exact font installed, and you get nonsense. A search engine sees the same nonsense. The text exists only as an image of itself, and converting it back is manual work that nobody funds.

Then there is Pegon, the Arabic-derived script used for Javanese, Sundanese, and Madurese, which carries an enormous pesantren manuscript tradition. It has no Unicode block of its own. It is written with Arabic codepoints plus a scatter of additions, which leaves encoding practice inconsistent from project to project. For anyone building a searchable corpus, that inconsistency is the whole obstacle.

The Kongres Aksara Jawa held in Yogyakarta in 2021 addressed exactly this class of problem: standardizing keyboard layouts, fonts, and transliteration rules so that the script could function as a working technology rather than as heritage decoration.

Manuscripts
Bali's Gedong Kirtya in Singaraja houses several thousand lontar (palm-leaf) manuscripts covering law, medicine, ritual, and literature. Batak pustaha on tree bark, Bugis chronicles, Acehnese religious texts, Minangkabau tambo: the archipelago's written record is spread across village collections, family houses, and provincial museums, in materials that insects and humidity continually work on.

Two programs have done most of the heavy lifting. The British Library's Endangered Archives Programme has funded digitization across Aceh, Riau, Sumatra Barat, and elsewhere. DREAMSEA, the Digital Repository of Endangered and Affected Manuscripts in Southeast Asia, is based at UIN Syarif Hidayatullah Jakarta in partnership with Universität Hamburg and has photographed manuscripts in situ across the region, leaving the originals with their owners.

Both produce images. Images are not text. A digitized lontar is preserved against fire and rot, and it remains unreadable to a search query until somebody transliterates it. Handwriting recognition for Balinese, Javanese, and Pegon scripts is an active area of research, with no production-grade tool yet. This is a bottleneck where trained philologists and translators are the constraint, not the technology.

Speech
Recording practice for spoken varieties is well established, so there is little excuse for doing it badly. Uncompressed audio, video where feasible, time-aligned transcription in ELAN, and metadata recorded at the moment of capture rather than reconstructed later. Who spoke, where, in which variety, on what topic, and with what permission.

Indonesian collections tend to fall down on the last two. A recording labeled "bahasa Sasak, Lombok, 2011" is nearly useless if Sasak has several mutually distinct varieties and the file does not specify which one. Archives that migrate formats as technology shifts also matter: PARADISEC in Australia holds substantial Indonesian and Papuan material, and it exists precisely because DAT tapes and MiniDiscs are already hard to read.

The speech-technology picture is thin. Mozilla Common Voice has Indonesian, but the contributions are overwhelmingly standard Indonesian rather than regional languages, and the speaker base skews young, male, and Javanese-adjacent. Anyone building an ASR model for a regional language in Indonesia today starts nearly from zero.

Corpora, machine translation, and why this concerns translators
Here is the part with direct professional consequences.

The NusaCrowd initiative and the NusaX dataset, assembled by Indonesian NLP researchers, brought together parallel and annotated data for roughly 10 regional languages, including Acehnese, Balinese, Banjarese, Buginese, Madurese, Minangkabau, Ngaju, and Toba Batak. Genuinely valuable work. Ten out of more than seven hundred.

Meta's NLLB-200 covers a similar handful: Acehnese in both Latin and Jawi orthographies, Balinese, Banjar, Buginese, Javanese, Minangkabau, Sundanese. Output quality varies wildly across directions and domains.

The register problem is the one to watch. For most low-resource languages worldwide, the largest available parallel corpus is scripture, because Bible translation has been running for two centuries and produces aligned text. Train on that, and the model learns a solemn, archaic, homiletic register. Ask it to render a tenancy agreement or a vaccination leaflet and it will produce something that sounds like a sermon. A reviewer who does not speak the target language cannot detect this at all.

So the professional stake is fairly clear:

Orthography and standardization. Communities are still deciding how to write their languages in Latin script for social media, signage, and school materials. Linguistically trained translators are useful in that argument, and often absent from it.

Terminology. Revitalization needs modern vocabulary, not only folklore. Somebody has to build padanan istilah for administration, health, law, and technology, and defend those choices publicly.

Register coverage. The gap in every Indonesian regional-language corpus is legal, medical, and administrative text. That is the material translators produce for a living.

Evaluation. There is no established quality framework for these pairs and, in most cases, no certification path. Who is competent to sign off on a Buginese medical consent form? The question is not rhetorical, and it carries liability.

Ethics. Recordings of elders are the pooled voice of a few families. Once a corpus enters somebody else's model, it cannot be withdrawn. Te Hiku Media in New Zealand built its own Māori speech recognition and licensed the data under kaitiakitanga rather than hand it to outside firms; that model is directly applicable here, and largely unadopted.

What the digital layer can and cannot do
Digitization does two things well. It buys time against decay, and it lowers the cost of access for the next generation of learners and researchers. Neither is small.

It does not produce speakers. A perfectly encoded, fully searchable, well-archived corpus of a language that nobody speaks at home is a very good tombstone. The Indonesian programs that show real movement are the ones where the digital work sits underneath something human: a pesantren teaching in Pegon, a village school running a Sasak curriculum, a grandmother in Wamena who has been given a reason to keep talking in her own language.

Build the infrastructure. Then go find the grandmother.


Figures for language counts and vitality categories move between survey rounds. Check the current Badan Bahasa release before citing them in published work.