Africans do not speak one language at a time. Neither should AI.

Listen closely anywhere on this continent and you will notice something no textbook prepares you for: the languages do not take turns. Nairobi runs Swahili, Sheng and English through a single sentence. Accra trades Twi and English mid-thought. And in Lagos: “The meeting don start, ṣùgbọ́n ẹ jọ̀wọ́, give me five minutes.” Pidgin, Yoruba and English in a single breath, from a single mouth, and nobody at the table blinks. In Kano the blend is Hausa and Pidgin: “Wallahi, this thing no easy, amma za mu yi.” In Enugu, Igbo threads through English with a biko dropped in exactly where it belongs. Linguists call it code-switching. Africans call it talking.

For speech technology, this is close to a worst case. Most systems treat a language as a setting: you select one, and everything that follows is assumed to stay inside it. A switch mid-sentence breaks their language detection and their pronunciation at once. Pidgin punishes them twice over, because it shares so much vocabulary with English that models quietly “correct” it, turning dem into them and dey into is, until the sentence is no longer Nigerian at all. A model can be technically intelligible and still be culturally wrong. And a voice model that cannot switch cannot speak to Africa. It can only speak at her.

Built for the switch

We trained STORM the other way round, starting in Africa's largest market. One model covers Nigerian English, Nigerian Pidgin, Yoruba, Igbo and Hausa, and treats switching as ordinary speech rather than an error state. Feed it the mixed sentence above and it rides every seam, tone marks respected where Yoruba demands them, Pidgin left exactly as spoken. Our normalisation layer protects Pidgin's own grammar, so dem, dey, wey and fit reach the voice untouched.

The same layer handles everything on the page that is not a word, because getting the accent right while getting the number wrong is not useful. ₦5.75 million must come out as five point seven five million naira, never anything else. A percentage has a different word in each of our languages. A time like 7:45 has to be told in Yoruba's own number system, which counts differently from English. Unglamorous work, and essential for anything that will ever read a bank balance aloud.

We would rather you judged the switching with your ears than take our word for it.

The accent survives

The part we are proudest of is what happens to identity at those seams. Give the model a Yoruba speaker and ask her to speak Igbo, Hausa, English or Pidgin, and she still sounds Yoruba. Not neutral. Not mid-Atlantic. Recognisably herself, in every language.

That is exactly right, because it is what real multilingual Africans sound like. The trader from Ibadan selling in a Kano market speaks Hausa with an accent that tells you where she is from, and nobody there would call that a flaw. Accent is not noise sitting on top of speech. It is the person. Most multilingual TTS wipes it away the moment the language changes. Ours keeps it.

Any voice, all five languages

STORM V2 adds cloning, and this is where the story stops being only African. Give it a clean clip of a voice, four to eight seconds is enough, and that voice can speak Nigerian English, Pidgin, Yoruba, Igbo and Hausa. The speaker does not need to know a word of Yoruba for their voice to speak it. STORM-OS handles the language layer; the voice remains theirs.

For a global business entering Nigeria, that quietly changes the question. It stops being how do we find five voices for five audiences, and becomes how does the voice we already have reach all of them. An author in Texas can narrate her audiobook for listeners in Onitsha, in Igbo, in her own voice. A club manager can wish supporters a happy new year in Pidgin without a script or a studio. A CEO can address teams in Kano and Enugu with one recording and no interpreter standing between her and her people.

Africa is our foundation, not our ceiling

We get described as an African language company, and we are, proudly. But the point of this technology is connection, and connection runs in both directions. The same engine that lets a brand in London speak Yoruba lets a Nollywood film speak to the world. Nigeria is where we proved the approach; the continent is where it goes next.

The early access list already reads that way. Enterprises are using STORM to reach multilingual audiences in live sport, cricket in India and football in Brazil, where a single commentary feed has to land in several languages at once without losing the excitement in any of them. A match that used to reach one linguistic market becomes reachable by many, and every new audience brings its own viewers, sponsors and broadcasters. That is not a cost saving. It is distribution. Businesses in Ghana, Kenya, South Africa and Nigeria are also in early access, building on the API today. None of them came to us for a demo of what might be possible. They came because they have audiences they cannot currently speak to.

The cases that cannot wait

The opportunity goes beyond new markets. Understanding your own diagnosis is another. Health information across the continent still travels mostly in English or French, while the patient thinks, worries and asks questions in Hausa, Yoruba, Swahili or Pidgin. A health line that answers in the caller's own language, medication instructions read aloud for someone who cannot read the leaflet, a maternity reminder delivered in the words a first-time mother actually uses: this is where a voice that switches stops being impressive and starts being necessary. When the language is wrong, symptoms go undescribed and instructions go unfollowed.

Education carries the same weight. Children learn fastest in the language they think in, and a patient voice that can read, explain and repeat in Igbo, Zulu or Somali puts a tutor on any phone a family already owns. The same goes for money and safety: a farmer can hear market prices and planting advice in her own language, a first-time bank customer can be walked through an account by voice instead of a form in a language she does not read, and a flood warning can reach a village in the words it actually speaks. None of this needs a new device or a literacy campaign first. It needs infrastructure that already speaks the languages people have. That is the job we have taken on.

Whose voice does the AI speak with?

For decades, businesses went international through text: translate the website, translate the documentation, translate the app. The next wave is happening through speech, in customer service agents, commentary, audiobooks, classrooms and cinema. Every company that rides it will meet the same two questions. Whose voice does the AI speak with? And can that voice move naturally between the languages your customers actually use? That is the infrastructure we are building.

What comes next

Thirty more African languages are on the way. STORM V3 will bring isiZulu and isiXhosa from the south, Lingala from the centre, Somali from the east, Twi and Ewe from the west, and Libyan Arabic alongside further Arabic varieties from the north, with more to come: one model reaching every region of the continent, the first truly pan-African voice AI model. And we are deliberate about what that means. “Supports thirty-five languages” and “works well in thirty-five languages” are two very different engineering claims, and we are building for the second, with the datasets, normalisation and evaluation each language deserves. Global access for everyday use opens by 1 October, so this stops being something only enterprises can touch.

And film is next. If voice identity can cross languages, actors no longer have to lose their voices when a film crosses borders, and independent productions get capabilities that used to belong to large localisation budgets alone. A Nigerian film should be able to travel to Brazil, and a Brazilian production should be able to speak to Nigeria, with the performances intact. We will unveil our work on movie dubbing at the San Antonio Black International Film Festival, SABIFF 2026, in Texas.