What N-ATLaS is
N-ATLaS stands for Nigerian Atlas for Languages & AI at Scale. It is Nigeria's first open-source multilingual AI project, launched on 20 September 2025 by the Minister of Communications, Innovation and Digital Economy, Dr Bosun Tijani, at the 80th UN General Assembly in New York.
It has two parts:
- A text model that reads and writes Háusá, Ìgbò, Yorùbá and English. It is a fine-tuned version of Meta's Llama 8B model.
- Four speech recognition models that turn speech into text, one each for Háusá, Ìgbò, Yorùbá and Nigerian-accented English. They are fine-tuned from OpenAI's Whisper Small model.
Who built it
N-ATLaS is an initiative of the Federal Ministry of Communications, Innovation and Digital Economy, built with the National Centre for Artificial Intelligence and Robotics (NCAIR, part of NITDA) and the Office for Nigerian Digital Innovation (ONDI), and powered by Awarri Technologies, a Lagos AI company.
How Naija Transcribe uses it
- We use the four speech models. We converted them to smaller formats so they run inside your web browser: Full (8-bit, about 287 MB) and Lite (4-bit, about 203 MB, almost as accurate). Downloads resume where they stopped if your connection drops. That is why your audio never has to leave your device.
- We do not use the N-ATLaS text model for translation. In our tests a hosted AI model translated more accurately, so Translate sends your transcript text to one through OpenRouter, only when you press it.
How accurate it is
We measured the speech models on Google's FLEURS test recordings: 100 clips per language of people reading sentences aloud. Word error rate is the share of words the transcript gets wrong, so lower is better.
| Language | Words wrong | Ignoring tone marks | Stock Whisper Small |
|---|---|---|---|
| Háusá | 38.5% | 38.5% | 92.1% |
| Ìgbò | 45.4% | 39.4% | No Ìgbò support |
| Yorùbá | 64.7% | 33.5% | 113.8% |
So N-ATLaS is far better than the general model it was built from, but about one word in three still needs fixing. Names and places are the most common mistakes, and the Yorùbá model sometimes leaves out tone marks. The smaller in-browser version is about 3 points less accurate than the full model. Real voice notes with background noise may score differently. Full method and results are in Naija Eval.
The licence
N-ATLaS is free to use under its own licence, with three conditions that matter here: services may have at most 1,000 active users in any 30 days unless they get a separate licence from Awarri; anything built from the models must credit Awarri Technologies and the Federal Government of Nigeria and keep the same licence; and uses such as surveillance, discriminatory profiling, disinformation and military applications are not allowed. See Legal and licences.
Is N-ATLaS up right now?
A live check of the N-ATLaS models on Hugging Face by Down in Nigeria. Naija Transcribe keeps its own copies of the speech models, so it keeps working even when Hugging Face is down.
Learn more
- The models: huggingface.co/NCAIR1
- National AI Innovation Challenge, built on N-ATLaS: ncair.nitda.gov.ng/naic
