March Recap + DSFSI@GAISAfrica2025

We’re thrilled to announce that the Data Science for Social Impact (DSFSI) Research Group at the University of Pretoria will be attending and contributing to the Global AI Summit on Africa happening this first week of April!

Our Delegation:

• Prof. Vukosi Marivate (Principal Investigator, DSFSI)

• Dr. Abiodun Modupe (Senior Member, DSFSI)

• Kayode Olaleye (Postdoctoral Fellow, DSFSI)

Our team will be actively engaging in the main summit and participating in various side events, sharing insights and exploring collaborations to advance AI across the continent.

Panel Highlight:
On Friday 4 April (930-1030), Prof. Vukosi Marivate will be speaking on a high-impact panel discussion with the topic GenAI Gaps and Opportunities for Inclusive Development.

We are thankful for the support of AI4D, IDRC, FCDO, Gates Foundation and Meta for supporting the work that enables us to attend the summit and AI discussions during the whole week.

Stay tuned for updates as we contribute to shaping the future of AI in Africa!

[

](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19fc5cb3-8f7c-4d76-9188-4be8824f2da4_1152x818.png)

Over the past few weeks, the Data Science for Social Impact (DSFSI) group at the University of Pretoria hosted its annual Postdoctoral Seminar Series, showcasing cutting-edge research from emerging leaders in African AI and Data Science.

From machine translation and code-switching to language modelling and automatic speech recognition, this year’s sessions highlighted practical innovations tackling the continent’s low-resource language challenges head-on.

Missed a session? We’ve got you covered! Below is a summary of each talk along with links to the full recordings.

Speaker: Dr. Kayode Olaleye
Date: 5 March 2025
Recording: Watch the recording

In this talk, Dr. Olaleye explores the complexities of code-switching and code-mixing in African languages. He presents the Afro-CSX dataset, which combines LLM-generated and human-validated samples across four languages (Yorùbá, isiZulu, Sesotho, and English), focusing on agriculture and personal finance. His team evaluated the impact of code-switched data on machine translation performance, revealing the value and limits of synthetic data and fine-tuning in low-resource settings.

Speaker: Dr. Idris Abdulmumin
Date: 12 March 2025
Recording: Watch the recording

Dr. Abdulmumin discusses the current gaps in language model performance across African languages, emphasising issues of data scarcity, tokenization, and evaluation. His research evaluates language model generalisation and the effectiveness of multilingual LLMs on low-resource African languages. He also proposes new directions for inclusive model training and culturally informed benchmarks.

Speaker: Dr. Hope Mogale
Date: 19 March 2025
Recording: Watch the recording

Closing the series, Dr. Mogale offers a highly practical walkthrough of building automatic speech recognition (ASR) systems that can handle diverse and noisy real-world African speech data. He demonstrates a custom ASR toolkit that supports dataset loading (e.g. Lwazi, NCHLT, Common Voice), model training, and speaker diarization. The talk is a must-watch for anyone interested in multilingual speech technologies and open tools for African languages.

You can access all three talks in one place via our YouTube playlist:
▶️ DSFSI Postdoctoral Seminar Series 2025 – Full Playlist

Stay tuned for upcoming DSFSI events, workshops, and research highlights. If you’re working on language technologies, responsible AI, or social impact in Africa, we’d love to connect.

Follow us on different platforms: https://linktr.ee/dsfsi

The DSFSI Postdoctoral Seminar Series is supported by the University of Pretoria and partners such as Microsoft Research Africa, JP Morgan AI, ABSA and many more.