Kayode Olaleye

Kayode Olaleye is a Postdoctoral Fellow at the University of Pretoria in South Africa, affiliated with the Data Science for Social Impact (DSFSI) Lab led by Prof. Vukosi Marivate. In March 2023, Kayode obtained a PhD at the University of Stellenbosch under the guidance of Prof. Herman Kamper. The doctoral research primarily focused on localizing keywords for low-resource speech through visually grounded speech models.

Kayode possesses a keen research interest in machine learning techniques and their applications in natural language and speech processing. Currently, as a postdoctoral fellow, Kayode actively explores approaches for processing and generating code-switched and code-mixed data in African languages.

He endeavors to bridge the gap between the performance of language technologies applied to well-resourced languages and the underrepresented low-resource African languages. Leveraging expertise in machine learning, the goal is to develop practical solutions that tackle real-world challenges in these languages. Through research and collaboration, he aims to contribute to the advancement of language technologies for a more inclusive and diverse digital landscape.

For further information or collaboration opportunities, feel free to contact kaykola.olaleye@gmail.com.

CV     Google Scholar     Github

We investigate keyword localization in visually grounded speech models trained without explicit textual or location supervision, achieving modest but notable results and identifying limitations and semantic mistakes.

We investigate the possibility of keyword localization in visually grounded speech models trained on images and spoken captions, exploring the use of attention in a convolutional model to improve performance, and finding that attention significantly enhances localization compared to previous models, though some incorrect localizations are due to semantic confusions.

We investigate the possibility of keyword localization in low-resource settings using weak supervision methods, comparing bag-of-words labeling with visually trained models and finding limitations in the visually trained model's performance, highlighting the need for better-suited localization methods.

The paper addresses the scarcity of visually grounded speech models for low-resource languages by releasing a new dataset in Yoruba and training a cross-lingual keyword localization model, aiming to stimulate research in this area.

This website's design and source code were copied from Jon Barron.