AfriBase-Speech: Auditing, Tiering, and Packaging African Speech Resources from Government Public Domain Archives

Authors

  • Ernst Augustus Petrus van Gassen Arktos AI Lab

DOI:

https://doi.org/10.55492/dhasa.v7i02.7704

Keywords:

speech data licensing, license washing, public domain audio, African languages, Whisper fine-tuning, provenance, corpus

Abstract

African speech data for natural language processing (NLP) has grown rapidly, but the licensing that governs it is fragmenting. We document license washing in four African speech datasets: a permissive Creative Commons Attribution (CC BY) tag on the hosting platform, contradicted by restrictive terms in the data card, end-user licence agreement, or terms of service.

As an alternative we package source audio that needs no licence in the United States (US) into a dataset released under CC BY 4.0. The US Foreign Service Institute, Peace Corps, and Defense Language Institute produced language courses to train diplomats, volunteers, and service members, and recorded the lessons of those courses as audio (``lesson audio''). As works of the US federal government, these recordings are in the public domain in the US under Section 105 of the US copyright law (Title 17 of the US Code), yet they have never been prepared for NLP use.

We inventory such lesson audio for 22 African languages and Haitian Creole. Our pipeline splits each recording into utterances, labels them with Whisper large-v3, and removes the English speech of the course instructor by language identification; English makes up 2 to 36% of segments, depending on the language.

Fine-tuning Whisper-small on the result lowers word error rate on the FLEURS (Few-shot Learning Evaluation of Universal Representations of Speech) benchmark for three of five languages, by 13 to 57 points. Removing English moves it by less than 2 points except for Amharic, where the quality of the machine labels limits training. The corpus, its per-language data sheets, and the pipeline code are released.

Downloads

Published

2026-11-20

How to Cite

AfriBase-Speech: Auditing, Tiering, and Packaging African Speech Resources from Government Public Domain Archives. (2026). Journal of the Digital Humanities Association of Southern Africa (DHASA), 7(2), 132-141. https://doi.org/10.55492/dhasa.v7i02.7704