Languages¶
Status: initial survey, October 2026. Resource-availability findings come from the sources linked below; (verify) marks items to confirm.
Scope¶
South Africa recognises 12 official languages — English, Afrikaans, isiZulu, isiXhosa, isiNdebele, Sesotho, Sesotho sa Leboa (Sepedi), Setswana, siSwati, Tshivenda, Xitsonga, and South African Sign Language (official since 2023) — plus the Khoi, Nama and San languages recognised in the Constitution (e.g. Khoekhoegowab, N|uu, ǃXun). Reservoir aims to cover all, in phases.
Principle: languages are not translations¶
Plant names, kinship terms, ritual vocabulary and proverbs often have no equivalent in English. The data model stores names per language as primary entries (with orthography, dialect, audio and contributor), not as translations of an English field. See data model.
Technology availability (what we found)¶
| Language | Notes from initial survey |
|---|---|
| isiZulu, isiXhosa | Supported by Lelapa AI's InkubaLM; Nguni morphological analysers/taggers from NWU/SADiLaR; covered by Sunbird's 51-language Whisper adaptation |
| Afrikaans | Best resourced of the group; Common Voice Scripted Speech 25.0 includes an Afrikaans dataset; Whisper trained on ~4 hours |
| Setswana | Common Voice 25.0 has a small Setswana dataset (~5 h, 17 speakers); covered by Sunbird Whisper adaptation |
| Sesotho | SADiLaR morphological annotations exist; NLP/ASR coverage appears thinner |
| isiNdebele, siSwati | Some NWU/SADiLaR annotation work for the Nguni languages; otherwise sparse (verify) |
| Sepedi, Tshivenda, Xitsonga | Appear least covered in the sources surveyed — priority for data collection and for avoiding over-reliance on generic models (verify) |
| SASL | Separate modality; requires video/sign-recognition research and Deaf community leadership |
| Khoi & San languages | Very small speaker populations, some critically endangered; require community-led, non-extractive approaches |
Key external resources:
- SADiLaR — national digital language resource centre; catalogue freely available for research
- Masakhane — grassroots African NLP research community
- Lelapa AI — InkubaLM (0.4B, isiZulu/isiXhosa/Swahili/Yoruba/Hausa) and Vulavula speech/text services
- MzansiText and MzansiLM — open corpus and decoder-only model for South African languages (2026)
- Mozilla Common Voice — open speech datasets (coverage varies by language and release)
- Sunbird AI Whisper adaptation — ASR for 51 African languages incl. Afrikaans, Tswana, Xhosa, Zulu
Implications¶
- We cannot assume a single multilingual LLM handles all languages well. The AI layer is provider-agnostic with per-language routing and evaluation (ADR-0006).
- Reservoir can be a data source as well as a consumer. With consent, recorded names, words and phrases can become an open, ethically-collected dataset for language technology — but only for records marked
training_allowed. - Every language needs a human custodian. No language ships as "supported" without a native-speaker reviewer (language custodians).
- UI localisation is a design constraint from day one — string externalisation, text expansion, right orthography, fonts that render clicks (ǃ ǀ ǁ ǂ) and tone/diacritic marks correctly.
- Audio matters. For many users and for oral cultures, speaking is primary and typing is a barrier.
Evaluation¶
Before declaring a language "supported" we require: a UI translation reviewed by a custodian, a minimum glossary of reviewed plant names, and an AI-quality evaluation set (see AI agents).