Who Builds the Data? A Contributor-Centered Study of Nigerian Language Audio Dataset Creation

Speech datasets for Nigerian languages (Yoruba, Igbo, Hausa, Nigerian Pidgin, and others) underpin automatic speech recognition and text-to-speech systems for over 200 million speakers, yet the data curation pipelines that produce them remain largely undocumented and unstudied. We present a qualitative, contributor-centered investigation of dataset construction practices for low-resource African language speech corpora, surveying individuals who have personally performed data collection, transcription, annotation, or dataset management. We stratify respondents into two populations, independent contributors (self-funded, volunteer-led, or informally organized) and institutionally funded contributors (grant-supported or corporate-sponsored), and compare their practices across funding and resourcing, speaker recruitment and compensation, access to linguistic annotation expertise, informed consent and IP/data governance, treatment of dialectal variation and code-switching, quality-control and inter-annotator agreement practices, and post-collection data stewardship. Preliminary data collection is underway, with 15 contributor responses across 4 languages gathered to date, spanning both independent and institutionally funded projects. This work extends data-centric AI research beyond post-hoc dataset auditing to the upstream curation decisions, annotator compensation, data ownership, dialect standardization, and time/resource constraints that determine dataset composition, label quality, and downstream model behavior before training begins. We hope to identify systematic differences between independent and funded projects in access to linguistic expertise, formalization of consent processes, and long-term corpus maintenance, alongside structural barriers common to both: inconsistent annotator compensation, informal or absent ethics review, and unresolved data governance after project or funding lifecycles end. By treating dataset creators, not just datasets, as a unit of analysis, this study contributes empirical, practitioner-grounded guidance for equitable and sustainable corpus construction in low-resource speech technology, with implications for data documentation standards, annotator labor practices, and community-centered data governance in ML pipelines more broadly.