OpenAI and Meta team up to teach AI African languages

If you ask modern AI models to generate code in Python or draft an essay in French, they won't break a sweat. But try speaking to them in widely spoken regional African languages, and you will hit a complete blank wall.

Telecom giant Orange announced on Tuesday that it is partnering with OpenAI and Meta to fix that massive blind spot.

Teaching Whisper and Llama new tongues

The initiative is tackling regional languages in Africa that are currently unrecognized by virtually any generative AI model.

It is starting with two major West African languages: Wolof, spoken by around 16 million people, and Pulaar, spoken by roughly 6 million.

Orange is teaming up with OpenAI to fine-tune the Whisper speech recognition model, and with Meta to train its Llama 3.1 open-source text model using diverse local data.

The immediate goal for Orange is operational: enabling local customers to interact with support and sales agents in their native tongues.

But the reach goes beyond telecom. Orange plans to release the resulting models under a free open-source license for non-commercial use cases, such as public healthcare and education.

It also plans to collaborate with local startups to foster regional AI ecosystems, with an eye toward eventually supporting all languages across its 18-country African footprint.

Separately, Orange inked another major deal with OpenAI to directly access frontier models hosted and processed locally in European data centers.

Why it matters

While frontier labs are obsessed with scaling benchmarks on English datasets, billions of people around the world remain entirely locked out of the generative AI boom.

Focusing on speech via Whisper is the smartest move here. In regions where literacy rates can be a bottleneck, voice interfaces turn AI from a luxury text tool into an accessible daily utility. Anyone can ask a question out loud and get guidance on healthcare, farming, or finances.

For local developers, getting access to free, fine-tuned open weights means they can build genuine local applications without burning millions on foundational training.

We think Orange is executing a clever dual strategy: securing European data compliance on one side, while locking down the telecom and AI gateway in emerging markets on the other.

The biggest challenge will be data scarcity. Fine-tuning models on limited oral corpora often leads to severe hallucination risks, especially in sensitive use cases like healthcare. Still, expanding frontier models beyond the Western internet bubble is a much-needed step.