GESMA: A dataset of Ghanaian environmental soundscapes for machine learning applications

Abstract

This paper presents a dataset of real-world Ghanaian environmental soundscapes intended to support machine-listening research and sound-event classification in low-resource contexts. The collection contains 22,193 uncompressed 44.1 kHz/16-bit WAV recordings, captured using mobile devices across diverse environments, including urban spaces, educational institutions, marketplaces, transport hubs, and human non-verbal acoustic settings. Recordings were obtained under natural field conditions to retain authentic background noise, reverberation, and overlapping sound events. Each file is accompanied by structured metadata specifying category, class, subclass, location, and context, and all annotations have been manually verified to ensure label consistency and quality. The dataset addresses a critical geographic gap in global audio resources and provides culturally and acoustically representative material from Sub-Saharan Africa. It offers strong potential for applications in environmental monitoring, sound event detection, accessibility tools, hearing-assistive technologies, and broader audio-based AI systems.