Short answer. Africa holds over 2,000 languages, close to a third of the world's total, yet 88% of them are severely underrepresented or ignored in computational linguistics under the Joshi et al. classification. A 2025 PRISMA review found just 74 published ASR datasets covering 111 African languages — roughly 11,206 hours across five years, fewer than 15% of them reproducible. The gap is not about speaker numbers (Hausa ~80M, Amharic ~60M, Swahili well over 100M): the top 100 NLP languages cover about 96% of world GDP but under 60% of world population, and commercial attention followed the GDP. What changes it is capacity on the continent — 60% of the population is under 25, with about 12 million entering the workforce each year.
Key takeaways
- Africa is home to over 2,000 languages, close to a third of the world's total, yet 88% are severely underrepresented or completely ignored in computational linguistics per the Joshi et al. classification.
- A 2025 PRISMA systematic review found 74 published ASR datasets covering 111 African languages totalling roughly 11,206 hours of speech across five years of research, with fewer than 15% providing reproducible materials.
- Around 814 African languages are classified as endangered, with Nigeria facing threats to 171, Cameroon 75 and Ivory Coast 65.
- The top 100 languages in NLP cover about 96% of world GDP but fewer than 60% of the world's population, which explains why commercial attention has not followed speaker counts rather than a lack of speakers.
- Africa has 60% of its population under 25 and roughly 12 million young people entering the workforce annually, producing a large, often multilingual workforce.
- Kenya's 2026 draft AI policy proposes fair-pay benchmarking against international rates, addressing reported earnings of $1.46 to $3.74 an hour against $21 to $27 for equivalent US roles.
How large is the gap between Africa's languages and its AI training data?
The gap is severe by every published measure. A systematic literature review published in 2025, following PRISMA methodology and covering research from January 2020 to July 2025, catalogued the entire published landscape of automatic speech recognition for African languages: 74 datasets, 111 languages, roughly 11,206 hours of speech in total. Automatic speech recognition (ASR) is the technology that converts spoken audio into text, and it depends entirely on transcribed audio in the target language to train on.
Eleven thousand hours for 111 languages, across five years of published research, is a small figure next to a single well-resourced commercial ASR programme for English, which might use tens of thousands of hours for one language alone. The same review found that fewer than 15% of the studies provided reproducible materials, and that dataset licensing was frequently unclear, so even the small amount of data that exists is often not practically usable by the next team that needs it.
Meanwhile, roughly 814 African languages are classified as endangered. Nigeria alone has 171 languages facing the most severe threat levels, Cameroon has 75, Ivory Coast has 65. These are languages where the window for documentation is closing while the AI systems that could support them are being built without them.
There is also a broader diversity problem behind the numbers: Africa is home to over 2,000 languages, close to a third of all languages spoken on Earth, and under the Joshi et al. classification — a widely used scheme for ranking how well a language is served by digital and computational resources — 88% of African languages are either severely underrepresented or completely ignored.
Why does this gap persist if African languages have so many speakers?
It persists because commercial interest has followed GDP rather than population, not because these languages are small. Hausa has roughly 80 million speakers, Yoruba and Igbo each have tens of millions, Amharic has around 60 million, and Swahili is a lingua franca across East Africa with well over 100 million speakers including second-language users. These are the primary languages of major economies, not obscure tongues spoken by isolated communities.
The AI4D African Language Program made the underlying mechanism explicit: the top 100 languages in NLP cover about 96% of world GDP but fewer than 60% of the world's population. As the programme's authors put it, the low economic interest African languages represent for the companies driving NLP means the work has to be taken up by African researchers themselves.
That explains the shape of what has been built so far. The most significant African language resources have come from grassroots and academic initiatives rather than commercial ones: Masakhane, a grassroots organisation of African technologists creating datasets and models since 2019; the AI4D African Language Program's crowd-sourcing challenges and research fellowships; the African Languages Lab; and benchmarks like IrokoBench and AfroLID built by African researchers. What has been missing is production capacity — the ability to take a client requirement for 500 hours of verified Wolof speech, or a Yoruba instruction-tuning dataset, and deliver it at commercial scale, on a timeline, with documented quality.
Does the continent have the workforce to do this at scale?
Yes: Africa has 60% of its population under the age of 25, the youngest continent by a wide margin, with roughly 12 million young people entering the workforce every year. Mobile penetration sits around 495 million subscribers and rising, with smartphone use climbing steadily. Kenya's technology ecosystem has earned the "Silicon Savannah" label; Nigeria's population of over 220 million supports a substantial technology sector; South Africa has sophisticated financial and professional services markets.
What that means practically for multilingual data collection is a large, young, digitally fluent workforce that is often multilingual by default. A Nigerian annotator may speak Yoruba at home, English professionally, Nigerian Pidgin socially, and understand Hausa or Igbo from regional exposure — a combination that is genuinely difficult to source outside the continent and exactly what multilingual annotation programmes need.
It matters for quality, too, in a way that cost-focused discussions miss. Annotating Yoruba correctly requires knowing Yoruba, but annotating it well requires knowing how it is actually spoken in Lagos versus Ibadan, which tonal distinctions carry meaning in which contexts, and when a code-switch into English is natural speech rather than a specification violation. That knowledge does not transfer through a training document; it has to live with the people doing the work, which is one reason recruiting native contributors close to the language matters more than recruiting anywhere cheap.
What are the labour conditions like in African data annotation work?
They are a documented and serious concern, and any account of this industry that skips them is incomplete. Kenya's draft Artificial Intelligence and Other Emerging Technologies Policy, published for consultation in 2026, addresses data annotation, content moderation and AI quality evaluation roles specifically. It proposes a fair-pay reference framework benchmarked against international rates rather than domestic minimums, responding to a stark gap: reported earnings of roughly $1.46 to $3.74 an hour against $21 to $27 for equivalent United States roles.
The draft also proposes mandatory psychosocial support, written contracts, transparent pay reporting and grievance mechanisms. A 2026 CHI paper drawing on interviews with Kenyan data workers describes a "regime of entrapment" produced by precarious contracts and global labour arbitrage. This is the context any company operating in Africa is entering, and the terms on which the work is done are becoming a procurement question rather than only a values statement — buyers subject to EU documentation obligations will increasingly be asked how contributors were treated, not only what was delivered. There is an operational argument alongside the ethical one: in languages where the qualified annotator pool is genuinely small, fair terms and stable work are how a delivery operation retains the ability to deliver in Wolof or Tigrinya next quarter. Attrition in a rare-language programme is a capacity loss that can take months to rebuild, not an HR metric.
What difference does having delivery centres inside Africa actually make?
It changes recruitment reach, dialect coverage, field realism, data residency and turnaround — five distinct operational gains, not one. For a language like Wolof or Oromo, contributors cannot be sourced through a global job board; they are reached through local networks, community organisations, universities and word of mouth in the regions where the language is spoken, which requires people on the ground who know those networks.
Dialect coverage — recruiting representatively across the different regional varieties of a single language — also depends on physical presence. Swahili in Nairobi is not Swahili in Dar es Salaam; Hausa spoken in Kano differs from Hausa spoken in Niger. A programme that recruits from one location produces a dataset that teaches a model one variety and treats the others as errors; distributed recruitment across the actual geography of a language is the only way to get representative coverage.
Field conditions matter too: if a voice product will run on inexpensive phones in noisy environments with intermittent connectivity, the training data should be recorded that way, not through a clean, centrally-run studio session. Several African jurisdictions also have data protection frameworks with cross-border transfer restrictions, and voice data that identifies a speaker frequently qualifies as sensitive personal data, making in-region collection a compliance requirement for some client programmes rather than a convenience. And a regional hub with reliable power and connectivity supporting speech data collection in low-resource languages converts an infrastructure problem into a logistics one — recordings collected offline can be physically transported to a hub with bandwidth rather than pushed over a weak connection at the contributor's expense.
Lifewood's published work in this area includes a long-running supply relationship with a global voice AI company covering multilingual speech and LLM services, specifically expanding voice-AI coverage into low-resource Asian and African languages — sustained capacity to extend a client's language coverage into places where the data has to be created rather than sourced, an approach detailed further in how speech data is collected for low-resource languages and in the account of running one delivery playbook across a global operation.
Where is African language coverage heading over the next few years?
Coverage is improving as language work becomes a policy objective and not only a commercial calculation. African governments and research institutions are increasingly funding language technology work directly, creating a new class of buyer with different requirements: openness, documentation and explicit coverage mandates. The grassroots research community has also produced benchmarks — IrokoBench, AfroLID, AfriWOZ and the Masakhane corpus family — that make African language work measurable in ways it was not five years ago, letting a client evaluate whether a model actually works in Yoruba rather than assuming it does because the language appears on a support list. This sits alongside a broader shift toward treating a language as high- or low-resource based on available data rather than speaker count, which is the framing this piece has argued for throughout.
None of that closes the gap on its own. What closes it is the unglamorous work: finding speakers, obtaining consent, recording, transcribing, verifying, adjudicating and documenting, thousands of times over, in languages where none of this has been done before — work that has to happen where the languages are spoken.