05 / OBSERVATORY · LOCAL LANGUAGES Back to home
African and minority languages.
An AI that does not speak the language of its users is useless. In 2025–26 real open tools arrived for African languages, especially for voice: in many communities people speak more than they write. Indigenous languages of Latin America and Haitian Creole remain the most uncovered.
Updated September 2026
Updated September 2026
Voice and translation for local languages.
Model, who built it, languages, license, type.
| Model | Who | Languages | License | Type |
|---|---|---|---|---|
| Omnilingual ASR | Meta | 1,600+, hundreds never covered before | Apache 2.0 | speech → text |
| Sunflower 14B/32B | Sunbird AI (Uganda) | ~31 Ugandan languages, expanding | Apache 2.0 | language model and translation; 9 GB version |
| MADLAD-400 | 450+ | Apache 2.0 | translation | |
| N-ATLaS | Awarri and the Nigerian government | English, Hausa, Igbo, Yoruba | licenza commerciale oltre 1.000 utenti al mese | 8B language model |
| UlizaLlama3 | Jacaranda Health (Kenya) | Swahili, Hausa, Yoruba, Xhosa, Zulu | licenza Llama | maternal health |
| InkubaLM | Lelapa AI (South Africa) | Zulu, Xhosa, Swahili, Hausa, Yoruba | CC-BY-NC (non commerciale) | small 0.4B model |
| NLLB-200 | Meta | 200, including Quechua, Guaraní, Haitian Creole | CC-BY-NC (non commerciale) | translation |
| Latam-GPT | CENIA (Chile) and consortium | Spanish and Portuguese; indigenous languages as a goal | aperto (base Llama) | 70B language model |
The data behind the models.
Open speech and text corpora, and the benchmarks that measure quality.
- WAXAL (Google Research Africa, 2026): 27 Sub-Saharan languages, ~1,850 hours of transcribed speech, CC-BY-4.0 Google Research
- African Next Voices: 9,000+ hours in 18 languages on agriculture, health, education Polity
- AfroBench (64 languages, GitHub ) and IrokoBench (17 languages, NAACL 2025 ) to measure quality
- AfriMed-QA: pan-African medical questions, ACL 2025 award for social impact ACL 2025
- Mozilla Common Voice: strong on Swahili and Kinyarwanda, CC0
A community that organizes itself.
Networks, funds and continental strategy behind African-language AI.
Masakhane (masakhane.io ) and the LINGUA Africa programme with Microsoft, the Gates Foundation and Google.org; Lacuna Fund ; the African Union’s AI strategy (AU ); Smart Africa; Deep Learning Indaba (Deep Learning Indaba ).