Introducing Namu AI Lab

Namu AI Lab is the research unit of Namu. We study speech technology for Hausa and Zarma in five research programmes, and we publish how we measure the work.

Namu AI Lab is the research unit of Namu. We study speech technology for Hausa and Zarma, two languages of the Sahel, and we publish that work here with the way we measured it.

What the Lab is

Namu builds speech-native models, datasets and products for African languages. The products live at namu-ai.org. The Lab publishes the research underneath them.

We start with Hausa, and Zarma comes next. Every programme page ends with what we do not claim yet.

Why the Sahel’s languages

Hausa has about 94 million first- and second-language speakers (Ethnologue, 2026). Zarma has about six million (Ethnologue, 26th edition, 2023) and is one of the main languages of Niamey.

Speech technology has not kept up with either language. Most public Hausa speech data was read aloud from a script, in Nigeria. In the Hausa test set of FLEURS, a benchmark many papers report, 620 of 621 utterances are labelled female. As far as we can find, Zarma has no Common Voice collection, no FLEURS test set and no adapter in Meta’s MMS recogniser.

Namu holds that access to technology should not depend on speaking a global language or being able to read and write. So the Lab works on speech, and it starts with Niger, which public Hausa datasets mostly leave out.

Five programmes

The work is organised in five programmes. Rumbu and Tankaɗe bring data in and make it usable. Hira and the Zarma programme build on that data. Maʼauni measures all of it.

Hira (conversation) is active. It studies full-duplex speech in Hausa, with models that listen while they speak. Our working system, Aïr, has an ear, a brain and a mouth. The brain is AfriqueQwen3.5-9B from McGill NLP (CC BY 4.0), and we do not retrain its weights. The mouth is built on Kyutai’s Moshi (weights under CC BY 4.0). Aïr’s live hearing still goes through a written transcript, so it is not yet end-to-end speech-to-speech.

Rumbu (granary) is opening. It will build a consented Hausa speech commons, Niger first, including two-person conversations on separate channels, ordinary phone calls and speakers who do not read. Nothing has been collected yet. The consent design and the pay rate will be published before recording starts.

Zarma from the ground up is starting. Namu has no Zarma data, model or measurement yet. The programme begins with a written convention vetted by Zarma writers and a small consented test set, and only then turns to models.

Tankaɗe (sifting) is active. It studies how radio archives, field recordings and crowd audio become data a model can learn from and a reader can trust. Our pipelines run internally. Their yields and thresholds will appear in a write-up before we quote them.

Maʼauni (measure) is active. It builds the test sets, text normalisation and listening protocols that make a Hausa or Zarma result mean something. Every Hausa error rate we report will come twice, strict and folded.

How we work

Consent comes before collection. Speakers will be told in their own language what a recording is for, and they will choose the uses they allow. They can withdraw. Everyone who records, transcribes or rates speech for us will be paid, and the rate will be public.

We publish methods and failures along with results. Hira exists because our first design, one model learning everything at once, picked up the rhythm of conversation and missed its sense. We will write that up as a finding.

Every number we publish will state its conditions and what it cannot show. Namu has not yet published a measured benchmark for any of its models. The one public result from our team comes from a competition. Abdourahamane Ide Salifou’s entry won the AI for Good Swahili ASR Challenge on Zindi in 2025, with a 17.81% word error rate on the private leaderboard (read the write-up). We keep competition results apart from model results.

We release tools, test sets and measurements where consent and licences allow. When something stays closed, we say what it is and why.

What comes next

Next, Maʼauni will publish its Hausa normaliser as v1, with test cases, then re-score open Hausa models on public test sets. Tankaɗe will write up its pipelines and their yields. Rumbu will publish its consent design and pay rate. Hira’s current step is an ear trained on real phone audio, so that Aïr’s hearing no longer goes through a transcript.

Two research notes appear alongside this one: Who is in the FLEURS Hausa test set and Why we report every Hausa score twice.

If you speak Hausa or Zarma, or you write, transcribe or study either language, each programme page says what help it needs. Write to us at contact@namuai.org.