Not Lost in Translation: Training AI to Speak African Languages

Why It Matters

Without AI tools that understand African languages, millions risk exclusion from digital services, health care, and the emerging AI‑driven economy. Inclusive language models are essential for digital sovereignty and equitable growth across the continent.

Key Takeaways

• Only 41 of 2,000 African languages are supported by major LLMs
• MzansiLM outperforms larger global models in isiXhosa accuracy
• Low digital footprints keep languages like isiZulu classified as low‑resource
• Gates Foundation funded African Next Voices dataset with 9,000 hours of speech
• Open research and shared benchmarks are essential for scaling African NLP

Pulse Analysis

The AI landscape has long been skewed toward English and a handful of other high‑resource languages, leaving the continent’s linguistic diversity largely invisible to modern tools. A 2025 analysis of six large and fourteen smaller language models revealed that just 41 African languages receive any consistent support, while the remaining 98% are effectively ignored. This disparity stems from the limited digital footprints of languages such as isiZulu and Hausa, which, despite millions of speakers, generate scant web‑based text for training data. The resulting feedback loop—where dominant languages attract more data, improving AI performance, which in turn creates even more data—exacerbates the digital divide and threatens cultural representation in the AI era.

In response, a team from the University of Cape Town launched MzansiLM, a decoder‑only model built on the curated MzansiText dataset covering all 11 official South African languages. Though modest in size, MzansiLM demonstrated superior accuracy and fluency in isiXhosa compared with much larger commercial models, proving that targeted, high‑quality data can outweigh sheer scale. The model is positioned as a foundational baseline for developers, enabling applications such as document summarisation and data annotation in languages that global AI services currently neglect. Its success underscores the strategic advantage of localized research and the potential for small‑scale models to drive inclusive AI adoption across health, education, and financial services.

The momentum is building beyond academia. A $2.2 million Gates Foundation grant produced the African Next Voices dataset—9,000 hours of speech across 18 languages—while Google’s WAXAL initiative opened a multilingual corpus spanning 21 African tongues. Community‑driven projects like Masakhane and partnerships such as Cohere’s with HausaNLP further expand open‑source resources and benchmarks. These collaborative efforts illustrate a shifting paradigm: inclusive language data is becoming a priority for both funders and tech giants. Continued open research, shared datasets, and robust evaluation frameworks will be critical to scaling solutions that give Africa a voice in the global AI conversation.