EUVIMEDEUVIMED entdecken
Open Access · CC BY-SA 4.0

Archival ASR for African Broadcasting: Automatic Transcription of VOA African-Language Audio

Ernst Augustus Petrus van Gassen

Journal of the Digital Humanities Association of Southern Africa (DHASA) · 20.11.2026 · 7 (2) · 142-150

DOI: 10.55492/dhasa.v7i02.7705 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of the Digital Humanities Association of Southern Africa (DHASA), online erschienen am 20.11.2026. Verfasst von Ernst Augustus Petrus van Gassen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Voice of America (VOA) broadcast in African languages for more than six decades through dedicated language services: newsrooms that each produced radio, television, and web content in one language. Its own output is in the public domain under United States (US) law, because works of US federal government employees are not protected by US copyright. The March 2025 shutdown of VOA's Africa division has frozen the archive. We crawl all 16 African-language services (Afaan Oromo, Amharic, Bambara, French, Fulfulde, Hausa, Kinyarwanda and Kirundi, Lingala, Ndebele, Portuguese, Sango, Shona, Somali, Swahili, Tigrinya, and Wolof), plus Haitian Creole. We transcribe the eight languages covered by OpenAI's Whisper speech recognition model into 364,815 utterances; Hausa (220.4 hours) and Portuguese (235.7 hours) are the largest. Meta's Massively Multilingual Speech (MMS) model covers every service language except Ndebele, and we add MMS transcripts of 3,000-segment samples for Afaan Oromo, Tigrinya, Kinyarwanda, and Bambara. Audio on modern VOA pages loads through JavaScript and resists standard download tools; a legacy content management system recovers most of what we archive. On FLEURS (Few-shot Learning Evaluation of Universal Representations of Speech), a read-speech benchmark, MMS has a lower word error rate (WER) than Whisper for all six languages where both run, 14.0 to 44.3% against 35.2 to 157.9% for Whisper large-v3, and reaches 38.9% on Wolof and 63.9% on Afaan Oromo. Fine-tuning Whisper-small on public domain lesson audio lowers its WER for two of four languages but does not reach large-v3. On the archive itself, transcript quality tracks each language's coverage in the base model. We release the audio, provenance metadata, transcripts with per-language quality grades, and the pipeline code.

Zitieren

APA

Gassen, E. A. P. V. (2026). Archival ASR for African Broadcasting: Automatic Transcription of VOA African-Language Audio. Journal of the Digital Humanities Association of Southern Africa (DHASA), 7(2), 142-150. https://doi.org/10.55492/dhasa.v7i02.7705

Vancouver

Gassen EAPV. Archival ASR for African Broadcasting: Automatic Transcription of VOA African-Language Audio. Journal of the Digital Humanities Association of Southern Africa (DHASA). 2026;7(2):142-150. doi:10.55492/dhasa.v7i02.7705

BibTeX
@article{Gassen2026,
  author = {Gassen, Ernst Augustus Petrus van},
  title = {Archival ASR for African Broadcasting: Automatic Transcription of VOA African-Language Audio},
  journal = {Journal of the Digital Humanities Association of Southern Africa (DHASA)},
  year = {2026},
  volume = {7},
  number = {2},
  pages = {142--150},
  doi = {10.55492/dhasa.v7i02.7705}
}
Kostenlos starten →