PAALPan-African AI Lab
We're an independent research lab working on voice AI for African languages.
Talking to our models feels like talking to a person. It says "ehe" while you explain, laughs when something is funny, waits while you look for a word, and lets you cut in. It hears when you are worried and slows down. And when you move from isiZulu to English and back in one sentence, it follows you, the way people across Africa talk.
Human conversation is not a sequence of perfectly formed sentences.
We come in within a fraction of a second. We say "mm" without taking the turn. We talk over each other and sort it out. We switch languages mid-sentence. And how we say something matters as much as what we say.
Voice systems that wait for silence fall short of this, and mixing languages mid-sentence is still an open problem for speech technology.
We're building a full-duplex speech model: one that listens while it talks.
Our research focuses on three problems.
Coming in within a fraction of a second, saying "mm" without taking the turn, and sorting out overlaps.
A model that keeps listening while it talks, and hears how something is said, not only what is said.
Following people who move from isiZulu to English and back in one sentence, starting with South Africa's eleven official spoken languages, and growing across the continent.
Our ambition is voice AI that works in Africa's languages.
Africa is home to extraordinary linguistic diversity, yet its languages are still poorly served by today's AI.
We believe the next generation of artificial intelligence should not require Africans to give up their languages or the way they speak. It should meet people where they are.
That future needs research.
We write here about the problems we're working on. We believe meaningful breakthroughs begin with difficult questions.
In South Africa, as in much of Africa, many people move between languages in the middle of a sentence. Researchers describe this as an open problem for speech technology.
People usually take turns within a fraction of a second, say "mm" without taking the floor, and sort out overlaps without thinking. A voice that waits for silence struggles with much of this.
Talking to a machine should feel like talking to a person. In most of Africa's languages it does not, yet. Why that matters, and what we are building.
We're looking to connect with researchers, engineers, universities, funders and organisations who share our ambition to advance AI research in Africa.
Africa should not only use the next generation of AI. We should help invent it.
Get in touch: info@paal.website