I am a speech and language researcher working on data quality and evaluation for low-resource and multilingual language technology.
I did my PhD at Stellenbosch University with Herman Kamper, on visually grounded speech models for keyword localisation, and was a postdoctoral fellow at the Data Science for Social Impact lab, University of Pretoria, from 2023 to 2025.
Most of my recent work is about the data itself: how it is collected, how speakers validate it, and whether the evaluation we run actually catches the failures that matter. I currently work independently on ASR and LLM evaluation for African languages.
Contact: kaykola.olaleye@gmail.com
Selected work
AfroCS-xs: Creating a Compact, High-Quality, Human-Validated Code-Switched Dataset for African Languages. ACL 2025. LLM-generated code-switched text across Afrikaans, Sesotho, Yoruba, isiZulu and English, corrected by native speakers.
Swivuriso: The South African Next Voices Multilingual Speech Dataset. 2025. Corpus design, annotation guidelines and quality control across seven language teams.
Keyword Localisation in Untranscribed Speech Using Visually Grounded Speech Models. IEEE JSTSP, 2022.
Full list on Google Scholar and ACL Anthology.
Datasets
Swivuriso (2025). 3,000+ hours, 2,353 speakers, seven South African languages, CC BY 4.0.
YFACC (2022). Yorùbá spoken-caption corpus extending Flickr8k, CC BY-SA 4.0.
AfroCS-xs (2025). Human-validated code-switched data across five languages. Available on request.