Low-resource language machine translation
Accurate translation for the ~7,000 human languages, most of which have little written text, no standard orthography or are primarily spoken.
State of the art
Meta NLLB-200 translates 200 languages including ~150 low-resource ones (2022, Nature 2024).
- Only ~1–2k parallel sentences exist for many languages
- Most languages are unwritten, so speech-only data must be collected
- No reliable automatic quality metrics for these languages
Who’s working on it
- Meta AI ↗Company
Released NLLB-200 (Nature 2024) and then Omnilingual ASR (November 2025), covering 1,600+ languages including 500 never before transcribed, with open weights and corpus
- Vukosi MarivateResearcherUniversity of Pretoria
Co-founder of Masakhane and a lead of African Next Voices (2025), which built large culturally grounded datasets for African languages
- Masakhane ↗Program
Grassroots African NLP community of 1,000+ participants across 30 countries that produced the benchmark datasets most low-resource African MT is now evaluated on
- Lelapa AI ↗Startup
South African lab whose InkubaLM (2024) is a 0.4B-parameter model matching far larger systems on African-language translation and sentiment tasks
- GoogleCompany
Added 110 languages to Translate in 2024 using zero-shot LLM transfer, and ships the MADLAD-400 corpus and models for long-tail languages
- SIL International ↗Program
Holds the field linguistics corpora and orthography work for thousands of unwritten languages, which is the data bottleneck no amount of compute substitutes for
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| infrastructure | Community-led speech and text data collection for 1,000+ underserved languages | $$ $10M – $100M |
Depends on
Required by
No other capability builds on this directly.