Introducing Namu AI Lab
Mouhamad Mamane
Why we build voice technology for the languages of the Sahel, what exists today, and how we want to work.
In Hausa, namu means “ours”. When we chose the name, we were thinking of the people we wanted to build for. They speak the languages of the Sahel all day long, at home, at the market, on the radio and on the phone, and they almost never hear a machine answer them in those languages.
At Namu AI Lab we build our voice models, publish our research, and open what we have built to anyone who wants to use it.
Why voice
My cofounder Abdourahamane and I grew up in Niger speaking Hausa. I often think of an evening in Niamey when I was about ten. The power went out while my family sat outside, and in the dark my uncle helped me find a question I could not have typed into any search bar. We simply talked until it took shape. Many of the people we know use their language most confidently when they speak it. It is the language they teach and trade in, the one they use to give directions, settle an argument or raise their children, and they rarely need to write any of it down.
Most of the technology that reaches us was built for a different kind of fluency. It expects people to read menus, type questions and spell every word correctly, usually in a language other than the one they speak at home. When it fails us, people often say that our languages are missing from the internet. I do not think that is right. Our languages are not missing. Machines were taught to look for them in text, and they live mostly in speech.
So we started with voice: technology that can hear what someone says, answer out loud, and carry the meaning from one language into another.
What exists today
Today, one model we trained speaks Hausa, Zarma and French, and the platform can also listen in all three. You can use it here, on the Lab’s site. Anyone can create an account.
In Create, you can write or say a message and hear it spoken in one of our languages, let two people who share no language talk through an interpreter, or translate a video so that the people in it speak another language. Agents lets an organisation set up a voice assistant that answers people’s questions out loud. And the API gives developers the same models to build into their own products.
None of this is finished. The models still make mistakes, some languages work better than others, and Zarma is where we have the furthest to go: translation to and from Zarma is still limited, and we need recordings from many more speakers before it works as well as Hausa. But it works, and you can use it today.
What we publish
We want to show our work, including the parts that did not go as planned. What we learn while building these models becomes a report on the research page, with how we tested and what the numbers can and cannot tell you. There are seven so far. They go from teaching a model Zarma with very little data to asking what open language models really know about Hausa, Zarma and Fulfulde. One of them looks at a public Hausa test set used to measure speech recognition, and finds that 620 of its 621 recordings are labelled female.
Our Voices
A model can only learn the voices it has heard, and most voices in our region have never reached the data that machines learn from. Older people, people in villages, people who speak with the accent of home rather than the accent of the radio: when they are missing from the data, the technology does not understand them.
That is why we started Our Voices. At its heart is tata, a game on the phone where adults record short phrases in their language and are paid for every recording that is accepted. The recordings stay with Namu, and we use them for our research and our products.
How we work
A few rules guide everything we do. Everyone who records for us agrees first and knows what their voice will be used for. People who record for us, or listen to and check recordings, are paid for that work. Every number we publish says how it was measured. And when we do not know something, or when something did not work, we say so.
What comes next
Next, we want more voices in every language we work in, from more regions and more generations, so that our models hear people the way they really speak. We have also begun looking at Fulfulde, which is spoken across the Sahel, and our first report on it sets out what already exists and what we would need. And we will keep working on the platform until using it feels as easy as talking to someone you trust.
If you speak a language of the Sahel, the simplest way to help is to lend your voice on tata. If your organisation works with people who speak these languages, or you want to build with our models, write to us at contact@namuai.org.
We called the company Namu because we want the people who speak these languages to hear this technology and feel that it is theirs. That is why we are building it with them.