Joint development of models for understanding, translating, and restoring Togolese national languages.
Fifty languages. Fifty hours of validated speech for each. Six thousand translated sentence pairs. The figures behind the partnership announced in Geneva on 12 July 2026 say how much is being undertaken — and, by implication, how much was missing.
Two ecosystems, one ambition
Togo AI Lab, the pan-African platform Zindi and Umbaji have joined forces to build models that can understand, translate and render Togo's national languages. The project was presented at the AI for Good global summit in Geneva, and sits inside Togo's national artificial intelligence strategy.
The goal has two halves that hold each other up. One is opening digital public services to people who do not use them in French. The other is correcting an absence: African languages remain close to invisible in AI models, for want of the data that would let them in.
Too many African languages remain under-represented.Celina Lee, CEO of Zindi
Ten languages named, fifty in scope
The announced scope covers fifty national languages in the end. Ten are already named, and they trace a linguistic map of the country rather than a convenient selection:
- Ewe and Mina (Gen), the most widely spoken in the south.
- Kabiyè and Tem (Kotokoli), in the centre and the north.
- Moba, Gourmantchéma (Gourma) and Bassar (Ntcham), along the northern edges.
- Losso (Naoudem), Akposso and Ana-Ifè, long absent from any digital corpus.
What the project targets, in numbers
An open-source collection platform is to gather speech and text directly from the communities who speak these languages. The target is public, and so checkable:
- 50 hours of validated speech per language.
- 6,000 translated sentence pairs per language.
- Four international competitions on Zindi, with a combined $40,000 prize pool.
- Three families of models in scope: speech recognition, speech synthesis and translation.
It is precisely the chain Umbaji has worked on from the start, except that the scale changes by an order of magnitude. Our earlier campaigns covered eleven African languages across six countries; this one targets fifty within a single country.
Who does what
Zindi, founded in 2018, claims more than 100,000 AI specialists across more than 180 countries: that is where the competitions run, and so where outside teams put the models to the test. It is not a first between Zindi and Togo — the two had already worked together in 2024 on infrastructure prediction.
Togo AI Lab carries the public side, backed among others by Berkeley's Center for Effective Global Action, the Patrick J. McGovern Foundation, the Development Innovation Fund and ProDigiT. Umbaji brings collection and validation: the least spectacular part, and the one everything else depends on.
Infrastructure, not a demo
Language models are essential public infrastructure.Cina Lawson, Minister for Public Service Efficiency and Digital Transformation
That sentence is worth pausing on. A demo is judged by the effect it produces on stage; infrastructure is judged by what it carries for years. Treating a language model like a road or a power grid means accepting that it is funded over time, measured in public, and built with the people who will use it.
It is the conviction Umbaji is built on: African language technology is built with local ecosystems, not on their behalf. Do you speak one of these languages, or work on the data that documents them? Get in touch.
