Why we're starting with Swahili (and Tanzania)

When people hear that we want to build AI that works in African languages and dialects, the first question is usually: where do you start? Africa has more than a thousand languages. Any honest plan has to pick a beginning.

We are starting with Swahili, and our first dialect is Tanzanian Swahili. This post explains both choices.

Why Swahili

Three facts make Swahili the natural place to begin a continent-wide effort.

• Reach. More than 200 million people speak Swahili, as a first language or as an everyday second one. Few African languages come close.
• Standing. Swahili is the only African language among the official languages of the African Union. It was adopted as an official working language of the AU, alongside Arabic, English, French, Portuguese and Spanish.
• A bridge. It is an official language of the East African Community and is widely used across East and Central Africa. People from different home languages use it to trade, travel, study and work together.

That last point matters most to us. A bridge language is where the work done for one community helps many others. Models that handle Swahili well, with all its regional variety and its mixing with English and local languages, teach us exactly what we need to learn before we take on more languages.

Why Tanzanian Swahili first

Swahili is not one voice. It sounds different in Dar es Salaam, Zanzibar, Mombasa and Nairobi. We have to pick a first dialect and do it properly, and we picked Tanzania because it is where Swahili is most deeply woven into daily life, education, government and commerce. Kenyan Swahili and other regional varieties follow, then more African languages.

Swahili is everywhere in Tanzania

In Tanzania, Swahili is not a second language people switch to for official business. It is the language of daily life. It is the national language and the language of primary education. Parliament debates in it. Newspapers, radio and television run in it. Market traders, bodaboda riders, nurses and civil servants all use it all day long.

That matters for us in two ways.

First, the need is real and immediate. A Tanzanian calling a bank, a clinic or a telecom provider expects to be served in Swahili. When the technology cannot do that well, the gap is felt by almost everyone, not only by a small group.

Second, the language is rich and alive. Tanzanian Swahili has a wide range of registers, from the formal Swahili of a news broadcast to the quick, playful speech of young people in Dar es Salaam. A model that handles all of that has to be genuinely good. Building it here forces us to meet a high bar from the first day.

Depth before breadth

It would be easy to announce support for twenty languages at once. It would also produce twenty mediocre products. We would rather do one language exceptionally well, then grow.

For Tanzanian Swahili, "exceptionally well" means:

• Pronunciation that sounds local, including names, place names and loanwords.
• Regional variation, so a speaker from Mwanza, Mbeya or Zanzibar is understood as well as someone from Dar es Salaam.
• Conversational expressions, the phrases people actually use, like "Sawa, nimekuelewa" or "Subiri kidogo", not only textbook sentences.
• Code-switching, because many people move between Swahili and English mid-sentence, especially in cities and in business.

Getting these right teaches us lessons that carry over to every language we add later: how to collect data fairly, how to evaluate speech honestly, how to design voices people trust.

The realities of how people connect

Tanzania also shows us what voice technology has to cope with in practice.

The phone call is still central

Many people reach businesses by calling, often from basic phones. That means our models must work on narrowband telephone audio, with background noise and imperfect signal. It is one reason we put telephony, over SIP and WebRTC, at the core of the Waridi API rather than treating it as an extra.

Mobile money shapes customer service

Mobile money is part of daily life, and so are the questions that come with it: a transfer that has not arrived, a forgotten PIN, a charge that needs explaining. These conversations involve amounts, phone numbers and reference codes, which a voice agent must hear and read back correctly. They are a demanding, high-stakes test for speech recognition and for spoken numbers in text-to-speech.

Cost matters

Airtime and data cost money. Long hold times and repeated calls are a real burden. A voice system that resolves problems quickly is not a luxury here; it is a direct saving for the people calling.

Being close to the language

There is a simpler reason too. We are based in Dar es Salaam. The people who judge whether a Swahili voice sounds right, whether a transcript is correct, and whether an agent is polite in the right way are our neighbours, our families and our team.

That closeness is hard to copy from a distance. It shapes everything from which speakers we record to how we phrase a greeting. Habari za asubuhi is not just a translation of "good morning"; it carries an expectation of how the conversation will go. Small details like this decide whether people feel served or processed.

What comes next

Tanzania is the beginning, not the boundary. Our plan, in order:

1. Tanzanian Swahili, done to a standard we are proud of.
2. East African Swahili more broadly, starting with Kenya, where Swahili sits alongside English and Sheng, and then Uganda. We write about the differences in from Dar to Nairobi.
3. More African languages, chosen with partners and communities, and built on the same consented, fairly paid approach to data.

Each step will reuse the infrastructure we are building now: the data pipeline, the evaluation methods, the low-latency serving stack and the API.

Tuanze hapa

Our goal is that anyone in Africa can use technology naturally in their own language and dialect. We are starting with voice, the interface current technology fails Africans on most. A goal that large needs a starting point that is concrete, demanding and close to home. For us, that is Swahili, and Tanzania.

Tuanze hapa — let us start here. If you are building for Tanzanian users and want early access, tell us about your project.