ajesujoba/AfriHuBERT-large

AfriHuBERT: A self-supervised speech representation model for African languages

Model description

This is a 300M+ parameter multilingual self-supervised speech encoder based on the English HuBERT-L model, facebook/hubert-large-ll60k. We performed continued pretraining through multilingual adaptive finetuning (MAFT) on over 10,000 hours of African language data aggregated from various sources.

Pretraining data

• Dataset: AfriHuBERT-large was trained on data from 11 major sources, including BibleTTS, Kallaama, MMS Ulab v2, NaijaVoices, and NCHLT. All sources and their licenses are shown in the table below. Please refer to the paper for more information.

Language Coverage

AfriHuBERT-large covers 1,230 languages in total including 1,226 indigenous African languages

BibTeX entry and citation info.

If you use this model in your research paper kindly cite the follows papers:

@doctoralThesis{alabi2026phd,
  author = {Alabi, Jesujoba Oluwadara},
  title = {Advancing African NLP: Adaptation, Analysis, and Evaluation of Large Language Models},
  school = {Saarland University},
  year = {2026},
  doi={http://dx.doi.org/10.22028/D291-48746},
}


@inproceedings{alabi25_interspeech,
  title     = {{AfriHuBERT: A self-supervised speech representation model for African languages}},
  author    = {Jesujoba O. Alabi and Xuechen Liu and Dietrich Klakow and Junichi Yamagishi},
  year      = {2025},
  booktitle = {{Interspeech 2025}},
  pages     = {4023--4027},
  doi       = {10.21437/Interspeech.2025-1437},
  issn      = {2958-1796},
}