Lessons from Building AI Data Infrastructure in Africa
Field Commentary
Summary
Artificial intelligence needs more African language data and that much is widely accepted. What is discussed far less is what happens when communities are asked to contribute that data. Our experience leading the Voicelink project in Uganda suggests that the harder challenge is not only collecting language data but building the governance arrangements needed to earn and sustain public trust.
We thought we were solving a data problem
Artificial intelligence is becoming multilingual. Governments, researchers, technology companies and international organisations are investing in datasets that allow AI systems to understand more languages and serve more people. Yet African languages remain among the least represented in the data used to train modern AI systems.
That gap matters because as AI becomes part of healthcare, education, public services, agriculture, finance and everyday digital tools, language coverage will shape who benefits from these technologies and who remains excluded. If an AI system cannot understand or communicate in the languages spoken by millions of people, those communities risk being left behind.
It was this problem that led us to launch Voicelink.
Voicelink is a research project led by Neuravox Foundation with support from the Mozilla Foundation. The project explores how existing community radio infrastructure can help build open speech datasets for African languages, beginning with Luganda in Uganda. Rather than creating entirely new recording programmes, we wanted to understand if trusted local radio stations could become part of a sustainable language data ecosystem while respecting community values and the public interest.
When we started, we believed the biggest challenge would be technical. Could we process hundreds of hours of broadcast audio? Could we separate speech from music? Could we build reliable quality assurance pipelines? Could we prepare datasets that researchers and developers could use?
Those questions shaped the early months of the project. We built processing pipelines, tested automated quality assurance methods, developed workflows for validating speech data and worked with radio stations and language experts to improve dataset quality. On paper, Voicelink looked like what many people would describe as an AI infrastructure project.
Then the questions changed.
The questions we were not expecting
As conversations with communities, broadcasters, researchers and policymakers continued, we realised that very few people wanted to begin with speech recognition models or machine learning pipelines. Instead, they asked questions that none of our technical systems could answer. Who owns these recordings? Who decides how they are used? Can communities place limits on future uses of their language? What happens after the project ends? Who benefits if this data helps build commercial AI systems? Who is accountable if communities disagree with how their language is represented?
Those questions forced us to pause and started by thinking about underrepresentation. Communities were thinking about governance. That distinction changed how we understood the project.
Representation is not the same as power
There is broad agreement that African languages deserve better representation in AI systems. We agree but representation alone is not enough.
A language can appear in a dataset while the people who speak that language have little influence over how the data is collected, shared, reused or commercialised. Communities can become part of the AI ecosystem without having any meaningful voice in how that ecosystem develops.
That was one of the most important lessons the project taught us. The challenge was never simply adding more speech data but creating the conditions under which communities would feel comfortable contributing that data in the first place.
Trust just became as important as technology.
Language is different
Speech data is often discussed as another category of training data. Our experience suggests something different.
Language is not simply information because t
it carries history, identity, humour, memory, social relationships, cultural norms and ways of understanding the world. When people contribute language data, they are not only contributing audio files but also a part of their community and that changes the governance conversation. Communities are not only asking how accurate an AI model will become. They are asking what happens to something that represents who they are.
Those questions need institutional answers, not only technical ones.
Governance begins before the first recording
One assumption we carried into the project was that governance would come later. First collect the data, then document it and then publish it responsibly. Implementation taught us something different, being that governance begins before the first recording is collected.
It begins when communities understand why data is being collected, when contributors know what they are agreeing to and institutions decide who can access the data, under what conditions and for what purposes.
As the project evolved, these questions became part of the design itself. We began exploring stewardship models rather than concentrating only on ownership. We developed language advisory structures that could provide cultural guidance throughout the project. We introduced technical processes to identify and remove personally identifiable information before datasets entered wider circulation and spent more time thinking about accountability than we had originally expected.
These activities became part of the infrastructure.
AI infrastructure is also governance infrastructure
Much of today’s conversation about AI infrastructure focuses on compute, cloud services, data centres, models and datasets. Those investments matter but our experience suggests that the institutions that make data collection legitimate deserves equal attention.
A speech dataset is not only a technical asset. It is also the product of relationships among contributors, communities, broadcasters, researchers, funders, developers and future users. If those relationships are weak, poorly defined or extractive, the dataset may still be technically useful but it will rest on fragile social foundations.
For African language data, this matters even more. Many communities are being asked to contribute to AI systems whose ownership structures, business models and long term uses are outside their control. In that context, governance cannot be treated as paperwork at the end of a project. It has to shape how data is collected, verified, stored, licensed, accessed and reused.
The Voicelink project did not provide all the answers but did something more valuable in that it changed the questions we were asking.
We began by trying to solve a language data problem and ended with a deeper appreciation that the future of AI in Africa will depend not only on better datasets, but on trusted institutions capable of governing those datasets in ways that communities understand, shape and accept.
About the author
Gideon Abako is the Executive Director of Neuravox Foundation and his work focuses on AI governance, data infrastructure and AI deployment in public systems. He leads Voicelink Uganda, a Mozilla Foundation supported project exploring community centred approaches to African speech data.