A grassroots research collective building NLP for African languages — by Africans, for Africa. Over 2,000 languages on the continent; when Masakhane started, fewer than ten had meaningful AI coverage.
What It Is
Masakhane (from Nguni languages, meaning "we build together") is a distributed research community focused on natural language processing for African languages. Started in 2019, the collective coordinates researchers across the continent and diaspora to build datasets, models, and benchmarks for languages that mainstream AI ignores.
Africa has over 2,000 languages. When Masakhane started, fewer than ten had meaningful NLP coverage. The collective addresses this through participatory research — native speakers contribute data, validate translations, and guide model development.
Why It Matters for Redistribution
Masakhane demonstrates what locally-led AI development looks like.
Who benefits? Speakers of African languages gain access to translation, text analysis, and other NLP capabilities. Researchers in Africa build expertise rather than importing solutions. The research community gets high-quality datasets that didn't exist before.
Who pays? The work is largely volunteer and grant-funded. Researchers contribute time; funders like Mozilla and the Lacuna Fund provide resources. This model has limitations (burnout, sustainability questions) but avoids extraction by commercial entities.
What changes? The typical pattern — Silicon Valley builds for English, translations follow maybe — gets inverted. Languages without commercial value still get served because the research agenda isn't set by advertising revenue.
The participatory model also matters. Masakhane papers frequently have 50+ authors because the work depends on distributed contributors. This is a different relationship between communities and AI development than "we built this for you."
As Kate Crawford argues in Atlas of AI, the material conditions of AI production shape what gets built (Crawford, 2021). Masakhane shows that changing those conditions — who participates, where, with what incentives — changes outcomes.
Masakhane is also a direct response to the dynamics described in Who Builds AI for the Rest of the World. The standard model — Silicon Valley builds, the Global South adopts — produces AI that reflects English-speaking priorities. Masakhane inverts this by building from the languages up. Their participatory approach to data collection addresses Who Owns the Training Data not through licensing debates but through community ownership from the start. And the volunteer model, while fragile, avoids the extractive dynamics documented in Data Labeling in the Global South — contributors are researchers and community members, not outsourced labor.
Current Status
Active and growing. Key outputs include AfroMT (machine translation), AfriSenti (sentiment analysis), and AfriSpeech (speech recognition). Multiple datasets now included in standard benchmarks.
The collective faces familiar challenges for volunteer research: funding instability, contributor capacity limits, and the gap between prototype and production deployment. But the model has proven replicable — similar efforts exist for other underserved language communities.
How to Get Involved
- Contribute data: Native speakers can help with translation and validation
- Join the community: Masakhane website — open to researchers at all levels
- Use the outputs: Models and datasets available through Hugging Face
- Cite the work: Academic visibility helps sustain the project
Related
- Lelapa AI
- Mozilla Common Voice
- Hugging Face
- The Digital Divide in AI
- Who Builds AI for the Rest of the World
- Who Owns the Training Data
- Data Labeling in the Global South
- Atlas of AI