Teaching the Machine to Pronounce Yoruba

Feed a Yoruba lyric into one of the big Western AI music generators and you will get something uncanny — not wrong, exactly, but off. The melody arrives intact. The words are there. And yet the voice singing them treats the tonal language as if it were English in disguise, flattening the rises and falls that carry meaning, pronouncing the syllables like a foreigner gamely reading from a phrasebook. The machine has learned to sing. It has not learned to sing in Yoruba.

"If you put Yoruba or Zulu lyrics into them, they'll sing the words, but they pronounce them like British or American English," Philip Olajide-Philips told OkayAfrica. He is the founder of Korin AI — korin is the Yoruba verb for to sing — and that single observation is the company's entire reason to exist. The global AI music boom, for all its astonishing fluency, was trained on a world that did not include most of Africa's languages. The result is a generation of tools that can do nearly anything except sound African.

A model built to be understood

Korin AI, founded by Olajide-Philips with co-founder Solomon Ayodele Ogunbowale and backed by the venture studio FirstFounders, is building a generative-music model with a narrow, stubborn goal: to pronounce African languages correctly. Not approximately. Not impressively-for-an-AI. Correctly — the way a Lagos singer or a Johannesburg vocalist would, with the tones intact and the accent native. The aim, as Olajide-Philips frames it, is an AI that sounds like an African voice in your own language rather than a foreign singer reading your words back to you.

This sounds like a small distinction until you consider what is encoded in pronunciation. In a tonal language, getting the tone wrong does not just sound bad; it can change the word, and with it the meaning. A mispronounced lyric is not an accent quirk. It is a kind of erasure — a flattening of the very features that make the language itself. To teach a machine to honour those features is to insist that African languages are not edge cases to be approximated but systems to be understood on their own terms.

The harder, slower way to build a dataset

What makes Korin worth paying attention to is not only the output but the method. The defining controversy of generative AI is the data it learns from — the vast, often unconsented scrape of human creative work that powers the largest models, and the lawsuits and resentment that have followed. Korin has chosen, deliberately, the slower path. Rather than scraping, the company licenses music from Nigerian production firms and pays the singers whose voices train the model, building consent-based datasets from the ground up, according to reporting by Music Ally.

It is hard to overstate how much friction this introduces. Licensing is expensive. Consent is slow. Paying artists is a cost that a scraping competitor simply does not carry. Every one of those choices makes Korin a harder business to run than the alternative. But it also makes Korin a fundamentally different proposition — a generative-music tool that does not extract from the culture it draws on but pays into it. In a year when 'ethical AI' has become a phrase emptied of meaning by overuse, a company actually structuring its data pipeline around consent is a quiet rebuke to the rest of the field.

The numbers, and the caution they deserve

The company's first version reached roughly 1,500 users, and a second version launched in May 2026, with a stated five-year ambition of reaching 10 million, according to founder interviews with OkayAfrica and Tech In Africa. These figures deserve a clear caveat: they are self-reported, the kind of numbers founders cite in interviews rather than audited metrics. Fifteen hundred users is a seedling, not a forest, and 10 million is an aspiration shared by roughly every ambitious startup that has ever drawn a hockey-stick chart. The gap between the two is where most companies quietly disappear.

But the significance of Korin was never going to be settled by its user count this year. The more durable question is the one it raises about whose voices get to live inside the machines that are increasingly making our music. The default answer, set by the largest models, has been: not African ones, or only as caricature. Korin is a bet that the default is a choice, and a reversible one — that you can build a model whose native register is Yoruba and Zulu rather than a transatlantic approximation of them.

There is a long history of African languages being treated as too marginal, too small, too hard to be worth the engineering. The internet repeated that pattern for years before African builders forced it to change. AI is now at the same fork. Korin's wager — that pronunciation is not a detail but a form of respect, and that respect can be a product — is the kind of bet that looks niche right up until the moment it looks obvious. Whether this particular company is the one to prove it is unknown. That someone in Lagos decided the machine should learn to pronounce the words properly is, already, the right instinct.