Language & Voice AI for Africa: From Data to Deployment and Impact

  • Vukosi Marivate, University of Pretoria; Tavonga Siyavora, Google Research Africa; Tobi Olatunji, Intron Inc; Kiplangat Korir, MsingiAI; John Quinn, Sunbird AI; Girmaw Abebe Tadesse, Microsoft; George Musumba, KICTANeT; Joyce Nabende-Nakatumba, Makerere University; Mercy Muchai, Microsoft; Yann Le Beux

This seminar explores how language and voice AI systems can be built and scaled for African contexts—from community-driven data collection and multilingual foundation models to robust deployment and real-world applications across sectors such as agriculture, health, and public services. We discuss technical advances, evaluation challenges, and ecosystem partnerships needed to ensure these technologies work for Africa’s linguistic diversity and development priorities.

Seminar Speakers: 

banner image of speaker Prof Vukosi Marivate with his name, University of Pretoria affiliation and mention of his keynote address

What Do Our Benchmarks Actually Measure? Evaluation Challenges for African Language AI

This talk will examine the growing gap between advances in language modeling and the evaluation methods used to assess them, drawing on emerging analyses of African language benchmarks to argue that rethinking evaluation is essential for enabling multilingual AI. Future frameworks must better reflect linguistic diversity, community priorities, and the complex sociotechnical contexts in which these languages are used.

Building the Substrate: The Foundry Model for African AI Innovation

This talk outlines the “Foundry Model,” a collaborative framework where empowered research organizations and local experts co-author the essential tools of the trade. Drawing on the origin story and success of a recent large-scale speech and language initiative (‘Waxal’), we demonstrate how community-led data engineering, paired with global research mentorship, creates a multiplier effect. We move beyond the “builder” vs. “user” dichotomy to explore how we can collectively forge a digital commons that empowers every startup and researcher to build the next generation of Africa’s context-aware technology.

banner image of speaker Tavonga Siyavora with his name, Google affiliation and mention of his invited talk
banner image of speaker Dr Tobi Olatunji with his name, Intron Inc affiliation and mention of his featured keynote

Problem Driven Development: The unglamorous road to real world African Voice AI

Despite rapid progress in speech AI, many systems still fail in African real-world settings where diverse accents, local names, multilingual speech, code-switching, noise, and domain-specific terminology collide. In this talk, I present the “ugly road” to production-grade voice AI through a problem-driven development lens: how failures observed across healthcare, enterprise, and everyday African conversations repeatedly became the starting point for new ideas, new datasets, better benchmarks, algorithms, and architectures, stronger models, and a series of published research. Rather than chasing global leaderboards, robust voice AI for Africa is built through disciplined error analysis, locally grounded evaluation, and tight feedback loops between deployment, data, and modeling.

Bringing Swahili to Life

Korir will share lessons from building Sauti, MsingiAI’s open-source Swahili TTS system, highlighting what it takes to move from data to deployment for a low-resource African language. This includes approaches to data, including curating WAXAL-compatible Kenyan Swahili speech, dealing with code-switching and dialectal variation, and the modeling choices that let us distill efficient, deployable voices that can run close to users. Ultimately, Korir will share what it takes to ship responsibly, and why open, Africa-led voice AI is the only sustainable path to language technology that truly serves the continent.

banner image of speaker Kiplangat Korir with his name, MsingiAI affiliation and mention of his lightning talk
banner image of speaker John Quinn with his name, Sunbird AI affiliation and mention of his lightning talk

Multilingual Speech LLMs in Practice

John will give some updates on Sunbird AI’s work with speech-language models for East African languages, aimed at optimising both latency and accuracy. From deployments across Uganda, he’ll build up an interesting picture of what people want to do with such models, and what opportunities we are seeing for further model iteration, debugging, and community collaboration.

banner image of panelists including Muchai Mercy (Microsoft), George Musumba (KICTANet), Joyce Nabende Nakatumba (Makerere University), Yann Le Beux (YUX and Kitala AI)

Watch Next