Why we are building Waridi
Waridi is an AI research lab training frontier models for human–computer interaction, built around how Africans actually speak. We are starting with voice, and with Swahili.
Call almost any large business in East Africa and you will hear the same thing. A recorded voice, a menu, and a list of numbers to press. Kwa Kiswahili, bonyeza moja. For English, press two. Then another menu. If you are lucky, a person picks up. If you are not, the call drops and you start again.
That phone call is a small picture of a bigger problem. The way people talk to computers — by voice, by text, through apps and assistants — was designed somewhere else, for someone else, and then translated. Most Africans meet technology in a language and a manner that is not quite their own.
Waridi exists to change that. Waridi is Swahili for "rose". We are an AI research lab training frontier models for human–computer interaction, built around the way people actually speak: regional dialects, accents, code-switching and local context.
The interface is the problem
Every new wave of computing has made the interface more human. Command lines gave way to screens, screens to touch, and now models let people simply say what they want. That shift should be the moment technology finally works for everyone.
For most African languages, it is not. The best models in the world are trained mostly on a handful of languages and the way those languages are spoken in a few places. When a farmer in Mbeya, a trader in Kariakoo or a student in Nairobi speaks naturally, the system hears a foreign accent, a "wrong" word, or a sentence that switches language halfway through. It guesses, and it often guesses wrong.
Our long-term goal is simple to state and hard to do: anyone in Africa should be able to use any technology naturally, in their own language and dialect — by voice, by text, or through whatever interface comes next.
Why dialects, not just languages
"Supports Swahili" is not one thing. Swahili in Dar es Salaam, in Zanzibar, in Mombasa and in Nairobi differs in vocabulary, rhythm and slang. Many people mix in English or Sheng without thinking about it. They grew up speaking Kisukuma, Kichaga or Kikuyu at home, and it shapes how they pronounce every word.
A model that treats all of that as noise will always feel like a visitor. So we build around dialects from the start:
- We collect and evaluate speech region by region, not as one blended "African" accent.
- We treat code-switching as normal speech. A caller who says "Nataka ku-check balance yangu" is speaking clearly. Our models should hear it that way.
- We pay attention to the small things that tell people a system was made for them: noun-class agreement, how a Tanzanian reads a phone number aloud, the polite forms people expect.
Starting with voice
We are starting with voice because it is the most natural interface there is, and the one current technology fails Africans on most. Many people reach businesses and services by phone, often from a basic handset, in a language that speech systems barely understand. Voice is where the gap is widest and where closing it helps the most people soonest.
Our first products from the lab are voice products, sharing one foundation of speech models trained for African languages:
- Waridi Agents — conversational voice agents for customer service that understand callers and resolve problems in real time, instead of press-one menus.
- Waridi Sauti — expressive text-to-speech, and voice cloning only with the verified consent of the voice owner.
- Waridi Sikia — speech recognition built for African languages, regional accents and code-switching.
- Waridi Dub — AI dubbing and localisation for filmmakers, broadcasters and platforms.
- The Waridi API — streaming, telephony over SIP and WebRTC, and SDKs so developers can build on all of it.
These are a first step, not the whole company. The same research — dialect-aware data, evaluation and models — is what we will build on as we move beyond voice.
Starting with Swahili
We are starting with Swahili. More than 200 million people speak it. It is the only African language among the official languages of the African Union, an official language of the East African Community, and a bridge language across East and Central Africa. Our first dialect is Tanzanian Swahili, because Tanzania is where Swahili is most deeply woven into daily life. We explain the reasoning in why we are starting with Swahili.
Data, earned the right way
Models are only as good as what they learn from. For African languages, high-quality recorded speech is scarce, and much of what exists was gathered with little care for the people who produced it.
We think the long-term advantage belongs to whoever builds the best datasets ethically. So we collect speech with clear, informed consent, pay speakers fairly, explain how recordings will be used, and let people withdraw. More in how we collect speech data with consent.
Why we are doing it ourselves
People sometimes ask why a team in Tanzania should take this on, rather than waiting for a large global company to add better support.
Our answer: Africa should not have to wait for the rest of the world to build AI for its languages. We intend to build it ourselves. The people closest to a language are best placed to judge whether a voice sounds right, whether a transcript is correct, and whether an assistant is being respectful. That judgement is the research. It cannot be bolted on later.
Where we are today
Waridi is in early access. We are not going to claim results we have not earned. When we share numbers, they will be measured, and we will explain how.
Our goal is plain: someone in Arusha, Nairobi, Dar es Salaam or a rural village should be able to use technology in their own language without compromising on quality.
If you want to build with us, request early access. Karibu sana.