For years AI spoke English well and almost everything else poorly. In 2025–26 something changed. Meta released Omnilingual ASR, which transcribes speech in over 1,600 languages under an Apache 2.0 license. Google Research Africa published WAXAL, about 1,850 hours of speech in 27 Sub-Saharan languages. In Uganda Sunbird AI released Sunflower, which translates local languages better than GPT-4o.
Voice matters more than text: in many communities people speak more than they write, and often with simple phones. A system that listens and responds in the local language, and runs on site without internet, can bring health, agronomic and educational support where today nothing arrives.
Two caveats. The quality gap remains: even a dedicated Nigerian model does much better in English than in Hausa. And many models are for research only: NLLB, InkubaLM and the Aya family cannot be used in a commercial service.