Two thousand languages.
One voice layer.
We build the speech models that let people use technology in the language they actually think in. Starting with Kenya’s fifty-two, working toward the roughly two thousand spoken across the continent.
A language survives when people can speak it, not just read about it
Hundreds of millions of people are coming online across the continent who will want to use technology in their own language. Three beliefs shape how we are building for them.
Speech first, not text first
Most African languages have far more spoken data circulating, on radio, in church services, in markets and in oral storytelling, than they have clean written corpora. Building around audio as the primary source is an architectural choice, not a workaround.
Depth before breadth
Kenya's own language map is a proving ground for the harder problem. Get the infrastructure right for fifty-two related but distinct languages, and the next fifty become cheaper rather than harder.
The infrastructure is the product
Phonology-aware models, a real grapheme-to-phoneme layer and honest evaluation are unglamorous. They are also what compounds in value across every language added afterwards.
The next fifty languages get cheaper if we build this together.
Institutional pilots, speech data, and research collaboration. We are a small team in Nairobi working on infrastructure that takes years.