Page 416 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

2 Use Case Description

2�1 Description

Radiolytica is an AI-driven system designed to monitor, transcribe, and analyze online community
radio broadcasts in Eastern DRC, a region where radio remains the primary information
source due to low internet penetration. The core problem addressed is the rampant spread
of misinformation, disinformation, and hate speech through radio channels, which traditional
manual monitoring methods cannot effectively mitigate.

Existing solutions, such as manual transcription and feedback from listener groups, are
slow, labor-intensive, and often inaccurate. Conventional speech-to-text tools struggle with
the region’s mixed languages, dialects, and high-noise environments, leading to significant
transcription errors. These limitations delay the detection of harmful content and impede timely
responses to crises, thereby exacerbating social tensions.

Our AI-based approach leverages advanced speech-to-text technology optimized for French
and Swahili, coupled with generative AI models that analyze transcribed content in near real-
time. This integration facilitates scalable monitoring across hundreds of radio stations, enabling
swift identification and classification of harmful content. By automating these processes,
Radiolytica not only accelerates the detection of dangerous narratives but also supports the
rapid production and dissemination of factbased counter-content.

We follow an oN-the-shelf approach by using the latest models released by industry leaders,
ensuring we stay up to date and deliver the best possible results. This strategy also helps us
avoid expensive fine-tuning of models and keeps our focus on the essence of the project:
detecting and countering harmful narratives in real time.

While the AI approach has significant benefits—scalability, speed, and efficiency—it also faces
challenges. Achieving high transcription accuracy in a multilingual, noisy environment remains a
technical hurdle. There is also a risk of AI-generated misclassifications or “hallucinations,” which
could undermine trust in the system. To mitigate these drawbacks, our solution incorporates
a hybrid model that combines automated processing with human validation by local experts
and journalists. This ensures that the insights generated are both contextually relevant and
reliable, ultimately enhancing operational eNectiveness and fostering social cohesion in
conflict-affected regions.

Use Case Status: The use case is part of a larger research project

Partners

Fondation Hirondelle [1] has a network of more than 150 partner radio stations across the DRC,
including around 60 in the east of the country. It also has a media outlet based in Kinshasa
called Studio Hirondelle RDC.

ETH Zurich: Our well-established collaboration with the Development Economics Group (DEC)
at ETH Zurich helps us address challenges in defining relevant analytical questions and ensuring
academic rigor. Their research expertise and methodological support enhance project impact.
Together, we apply for grants, including an ongoing academic grant for the proposed project,
focused on development economics questions.

380