94 Million Hausa Speakers, and AI Still Barely Understands Them. What Three Years of Grassroots Work Taught Me.

A developer spent three years building AI tools for Hausa, a language spoken by 94 million people but understood only 10-20% by ChatGPT and LLMs. The effort, rooted in the community AI Bauchi founded in Bauchi State, Nigeria, tackled challenges including orthographic inconsistencies, script duality, and code-switching. The project produced a Hausa text-to-speech model, a Hausa-to-Sayawa translator, and the HausaMediaLab library.

Hausa is spoken by close to 94 million people in Nigeria alone, with tens of millions more across Niger and the wider Sahel. Nature has reported that ChatGPT and LLMs correctly understands only 10 to 20 percent of sentences written in Hausa. I've spent the last three years trying to close that gap. First alone, then with a community I founded in Bauchi State that's since grown past 500 people. This isn't really a project retrospective. It's an honest account of what it actually takes to build language technology for a language the industry has quietly decided isn't worth the training data, and what leading the people doing that work with me has taught me. The easy explanation for Hausa's AI gap is "not enough data." That's true, but it undersells the problem. Hausa has technical properties that make naive data collection actively misleading, not just insufficient. Start with orthography. Standard written Hausa uses three hooked consonants, ɓ, ɗ, and ƙ, that don't exist on a standard keyboard. In formal, edited text they're there. In the informal Hausa that actually dominates the web social media, SMS, WhatsApp, the actual keyboards people type on they're routinely flattened to plain b, d, and k. Scrape the open web for Hausa and you get a corpus quietly missing phonemic information in a large share of its examples, and that gap propagates into anything trained on it, from tokenizers to text to speech models. Then there's script. Hausa has been written in two systems for over a century: Boko, the Latin based script that's now standard, and Ajami, an Arabic based script still used in religious and literary contexts. A dataset built only from Boko sources is already excluding a real slice of the written language. Then code-switching. Everyday Hausa, especially online, mixes in English freely and carries a long history of Arabic loanwords from Islamic scholarly tradition. A model that treats Hausa as a clean, isolated language misses how the language is actually used in practice. Put those together and "low-resource language" starts to look less like "not enough Hausa exists" and more like "the Hausa that's easy to scrape is a degraded sample of the Hausa people actually speak." Standard multilingual tokenizers, trained on that same degraded sample, fragment Hausa words into far more subword pieces than an equivalent English sentence needs. That makes inference more expensive and generation noticeably worse, since the model has fewer clean examples to learn from per token. That's the actual problem. Not "AI doesn't know Hausa." AI has been fed a thin, flattened version of Hausa, and it shows. I founded AI Bauchi in October 2022 out of Abubakar Tafawa Balewa University, with a mission that hasn't changed since day one: make AI indigenous and accessible in Bauchi State, not just consumed from outside it. The language work grew out of that mission, not the other way around. The tempting move, and the one I've watched well-funded initiatives make, is to announce a "Hausa LLM" on day one. I didn't do that, and I don't think a volunteer community should. A full generative language model needs enormous clean data and compute to be worth the name. What a community of volunteers with day jobs and a group chat actually has is time, native speaker knowledge, and the ability to ship something narrow and useful while the harder foundation gets built underneath it. So the sequence was deliberate. A Hausa text to speech model first, because it forces you to solve the orthography and pronunciation problem directly. A Hausa to Sayawa translator alongside it, because it forces the team to think about Hausa in relation to another indigenous language instead of only English. A media processing library, HausaMediaLab, that turns both into something a developer outside the project can actually pull in and use. Each of those was scoped to ship, not to impress. Each also produces cleaner labeled data as a byproduct, which is exactly the fuel a more ambitious language model eventually needs. Here's the decision I'd defend hardest to another technical lead: the TTS project's contributors aren't only machine learning engineers. They're ML engineers, data scientists, linguists, and software developers, working together on purpose. That's not a diversity nicety, it's an engineering requirement. An ML engineer can build a text to speech pipeline. An ML engineer without a linguist on the team will get tone marking and hooked consonants subtly wrong in ways that are invisible to anyone who doesn't speak the language natively, and obvious to everyone who does. There's a second representation question underneath that one: which Hausa. The language spans real dialectal variation. The Kano variety often treated as a reference standard is not the only lived version of the language, and a single synthetic voice trained on one region's speech quietly tells everyone else their Hausa is the deviation. A linguist in the room is what catches that before it ships, not after. So the structure is: weekly progress meetings, native speakers reviewing linguistic decisions, engineers reviewing the model, and a public repo with issues explicitly tagged for newcomers so the contributor pipeline doesn't dry up when the founding few get busy. Alongside the language projects we run bootcamps, an eleven week AI bootcamp and a six week computer vision bootcamp, that exist specifically to grow that pipeline. Most contributors to the harder projects came up through one of those cohorts first. If I only talked about the model, I'd be underselling what actually moved. A few things I'd point to as the real impact: I spent six months in 2023 as an instructor for NITDA's AI Developer Training, which meant the ideas driving AI Bauchi's work weren't staying inside one community, they were reaching a national training cohort. At DevFest Bauchi 2023 I gave a talk called "Democratizing Indigenous AI," making the case to a room of general developers, most of whom had never thought about Hausa specifically, that inclusive AI development has to start by respecting the linguistic and cultural nuance of the communities it claims to serve. In 2025 we ran EmpowerHer AI Bauchi, built to get more young women into this pipeline early, because a community that only recruits people who already see themselves as engineers keeps producing the same blind spots. And we've taken this outside the university entirely, hosting secondary school students from I-Scholars International Academy for a STEM excursion at NCAIR, Nigeria's National Centre for Artificial Intelligence and Robotics, because the earliest point to fix "who gets to build this" is before someone's chosen a degree. None of that shows up in a commit history. All of it is why the commit history exists. In 2025, Nigeria's federal government unveiled its own multilingual model effort through NITDA and NCAIR, covering Hausa alongside Yoruba and Igbo. That's a good thing, and it's an enormous, well-funded validation of a problem AI Bauchi had already been working on, in public, since 2023. I don't think grassroots work like ours competes with that kind of national effort. I think it's what keeps it honest. A federally funded model can move fast and cover a lot of ground. A community-built one, with linguists in the room and native speakers reviewing every tone mark, is what makes sure the fast version doesn't flatten the same nuance the open web already flattened. Both are needed. I'd rather our work be the thing that keeps the bigger effort accountable than the thing racing it for a headline. Note to self before publishing:I want this section grounded in exactly what I used, not a general gesture at the program. Naming the actual services, credits, or mentorship conversations that fed into this work is also the part AWS reviewers will check most closely. The program's value to this work was never the swag. It was mentorship conversations with AWS specialists who'd already thought hard about deploying models into low connectivity, resource constrained environments, which is the exact condition Bauchi operates under, and credits that let a volunteer community run experiments it couldn't otherwise afford. S3, GPU training instances of Sagemaker Start narrower than feels ambitious. A translator or a TTS model that ships is worth more than a language model that stays a roadmap item. Recruit people who aren't engineers before you think you need them, because by the time you obviously need a linguist, you've already shipped something linguistically wrong. Treat your constraints, low connectivity, volunteer time, no enterprise compute budget, as a design spec, not an excuse. Some of the most careful data hygiene decisions in this project exist specifically because we couldn't afford to throw compute at a sloppy dataset the way a better funded team might. The text to speech model and the translator were never the destination. They were how a volunteer community earns the right to attempt something bigger, and that bigger thing is still the goal. If you're building language technology for a community that big tech has decided isn't worth the training data, or you're an AWS Community Builder working in a similarly underserved region or language, I'd genuinely like to compare notes. Drop a comment, or find AI Bauchi and say hello.