Elamur Research⌕
☰
home / papers / 2026 / Learning from Scarcity: Building and Benchmarking Speech Technology for Sukuma.

Learning from Scarcity: Building and Benchmarking Speech Technology for Sukuma.

Paper · 2026 · Macton Mgonzo, Kezia Oketch, Naome A Etori et al.

Summary

Sukuma is a Bantu language of northern Tanzania with roughly 10 million speakers, but it has had no public speech data or models. The authors fill this gap with Sukuma Voices, an open paired audio–text corpus. They built it from audio and text of the Sukuma New Testament 2000 translation on Bible.com, and native speakers checked it for consistent spelling and pronunciation. Because the written language exists in both diacritic and non-diacritic forms, the experiments use only the non-diacritic version.

The corpus supports two baselines trained in parallel on a train/test split:

  • an ASR system, evaluated with WER;
  • a TTS system, a LoRA-tuned Orpheus model, rated by native listeners with MOS.

A cross-model check also feeds the synthesized speech back into the ASR system. Synthetic audio raises the error rate, yet its learning curves closely track those of real recordings. This suggests the generated speech keeps most phonetic content.

As the first released benchmark for Sukuma speech, the work gives future African low-resource speech research a reproducible starting point. Its main limitations are the corpus's small size and its single religious domain.

Key points

  • Corpus statistics: 2,588 utterances totalling 7.47 hours. Clips last 1.74–30.36 s, averaging about 10.4 s. The text contains 51,107 words with 10,007 unique word types, at a speaking rate of about 115.5 WPM.
  • TTS setup: Orpheus 3B v0.1 with the SNAC codec at 24 kHz, fine-tuned via PEFT/LoRA (rank 64, alpha 64, no dropout).
  • TTS quality: Two native speakers gave the synthesized speech a MOS of 3.9 ± 0.15, compared with 4.6 ± 0.1 for human recordings.
  • Synthetic vs. real audio in ASR: TTS-generated test audio produced a 28.34% higher WER than real speech. Still, learning curves on the two kinds of data correlated very strongly (Pearson r = 0.997, Spearman ρ = 0.993).
  • Availability and future plans: Data (Hugging Face, sartifyllc/Sukuma-Voices-ACL) and code (GitHub, Sartify/sukuma-voices) are public. The authors intend to expand the dataset past religious material until it reaches no less than 100 hours, and to create evaluation metrics that can deal with diacritics.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Brown UniversityProvidence · United States76 works · 25,952 citations🎓University of Notre DameNotre Dame · United States51 works · 36,717 citations🎓University of MinnesotaMinneapolis · United States28 works · 2,184 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Macton Mgonzo, Kezia Oketch, Naome A Etori, Winnie Mang'eni, Elizabeth Fabian Nyaki and Michael Samwel Mollel. 2026. Learning from Scarcity: Building and Benchmarking Speech Technology for Sukuma. Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.loreslm-1.25

@inproceedings{mgonzo2026learning,
  title = {{Learning from Scarcity: Building and Benchmarking Speech Technology for Sukuma.}},
  author = {Macton Mgonzo, Kezia Oketch, Naome A Etori, Winnie Mang'eni, Elizabeth Fabian Nyaki and Michael Samwel Mollel},
  booktitle = {Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loreslm-1.25},
  url = {https://doi.org/10.18653/v1/2026.loreslm-1.25}
}

References

29 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David Ifeoluwa Adelani, Jade Abbott, Graham Neubig, Daniel D’souza, Julia Kreutzer, Constantine Lignos, Chester Palen-Michel, Happy Buzaaba, Shruti Rijhwani, Sebastian Ruder, and 1 others. 2021. Masakhaner: Named entity recognition for african languages. Transactions of the Association for Computati…
  2. Tejumade Afonja, Tobi Olatunji, Sewade Ogun, Naome A Etori, Abraham Owodunni, and Moshood Yekini. 2024. Performant asr models for medical entities in accented speech. arXiv preprint arXiv:2406.12387. — [arXiv:2406.12387](https://arxiv.org/abs/2406.12387) find
  3. Orevaoghene Ahia, Anuoluwapo Aremu, Diana Abagyan, Hila Gonen, David Ifeoluwa Adelani, Daud Abolade, Noah A Smith, and Yulia Tsvetkov. 2024. Voices unheard: Nlp resources and models for yor\\ub\\’a regional dialects. arXiv preprint arXiv:2406.19564. of the Joint 25th Nordic Conference on Computational… — [arXiv:2406.19564](https://arxiv.org/abs/2406.19564) find
  4. Pierre Godard, Gilles Adda, Martine Adda-Decker, Juan Benjumea, Laurent Besacier, Jamison Cooper-Leavitt, Guy-Noel Kouarata, Lori Lamel, Hélene Maynard, Markus Müller, and 1 others. 2017. A very low resource language speech corpus for computational language documentation experiments. arXiv preprint… find
  5. Alexander Gutkin, Isin Demirsahin, Oddur Kjartansson, Clara Rivera, and Kólá Túbòsún. 2020. Developing an open-source corpus of yoruba speech. In Interspeech, pages 404–408. find
  6. Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the nlp world. arXiv preprint arXiv:2004.09095.
  7. Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. 2019. Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:1912.06670.
  8. Andrew Katumba, Sulaiman Kagumire, Joyce Nakatumba-Nabende, John Quinn, and Sudi Murindanyi. 2025. A curated crowdsourced dataset of luganda and swahili speech for text-to-speech synthesis. Data in Brief, page 111915. find
  9. Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020a. wav2vec 2.0: A framework for self-supervised learning of speech representations. Preprint, arXiv:2006.11477.
  10. Ussen Kimanuka, Ciira wa Maina, and Osman Büyük. 2024. Speech recognition datasets for low-resource congolese languages. Data in Brief, 52:109796. find
  11. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020b. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449–12460.
  12. Edward Bayes, Israel Abebe Azime, Jesujoba O Alabi, Jonas Kgomo, Tyna Eloundou, Elizabeth Proehl, Kai Chen, Imaan Khadir, Naome A Etori, Shamsuddeen Hassan Muhammad, and 1 others. 2024. Uhura: A benchmark for evaluating scientific question answering and truthfulness in low-resource african languages… find
  13. Astik Biswas, Emre Yılmaz, Ewald van der Westhuizen, Febe de Wet, and Thomas Niesler. 2022. Code-switched automatic speech recognition in five south african languages. Computer Speech & Language, 71:101262. find
  14. Naome Etori and Maria Gini. 2024. Rideke: Leveraging low-resource twitter user-generated content for sentiment and emotion detection on code-switched rhs dataset. In Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, pages 234–249. find
  15. Naome A Etori, Arturs Kanepajs, Kevin Lu, and Randu Karisa. 2025. Lag-mmlu: Benchmarking frontier llm understanding in latvian and giriama. In Proceedings Canopy Labs. 2024. Orpheus: A neural codec language model for high-quality text-to-speech. find
  16. GP Lestrade. 1948. The classification of the bantu languages. African Studies, 7(4):175–184. find
  17. Masangu Matondo. 2006. Tonal transfer in kisukuma. In Selected Proceedings of the 35th Annual Conference on African Linguistics, pages 125–135. find
  18. Sewade Ogun, Abraham T Owodunni, Tobi Olatunji, Eniola Alese, Babatunde Oladimeji, Tejumade Afonja, Kayode Olaleye, Naome A Etori, and Tosin Adewumi. 2024. 1000 african voices: Advancing inclusive multi-speaker multi-accent speech synthesis. arXiv preprint arXiv:2406.11727. — [arXiv:2406.11727](https://arxiv.org/abs/2406.11727) find
  19. Kezia Oketch, John P Lalor, and Ahmed Abbasi. 2025. Benchmarking sociolinguistic diversity in swahili nlp: A taxonomy-guided approach. arXiv preprint arXiv:2508.14051. — [arXiv:2508.14051](https://arxiv.org/abs/2508.14051) find
  20. Tobi Olatunji, Tejumade Afonja, Aditya Yadavalli, Chris Chinenye Emezue, Sahib Singh, Bonaventure FP Dossou, Joanne Osuchukwu, Salomey Osei, Atnafu Lambebo Tonja, Naome Etori, and 1 others. 2023. Afrispeech-200: Pan-african accented speech dataset for clinical and general domain asr. Transactions of…
  21. Vineel Pratap, Anuroop Sriram, Paden Tomasello, Awni Hannun, Vitaliy Liptchinsky, Gabriel Synnaeve, and Ronan Collobert. 2020. Massively multilingual asr: 50 languages, 1 model, 1 billion parameters. arXiv preprint arXiv:2007.03001. — [arXiv:2007.03001](https://arxiv.org/abs/2007.03001) find
  22. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356.
  23. Mardhiyah Sanni, Tassallah Abdullahi, Devendra Deepak Kayande, Emmanuel Ayodele, Naome A Etori, Michael Samwel Mollel, Moshood O Yekini, Chibuzor Okocha, Lukman Enegi Ismaila, Folafunmi Omofoye, and 1 others. 2025. Afrispeech-dialog: a benchmark dataset for spontaneous english conversations in healt…
  24. Avinash Kumar Sharma, Manas Pandya, and Arpit Shukla. 2025. Fine-tuning whisper tiny for swahili asr: Challenges and recommendations for low-resource speech recognition. In Proceedings of the Sixth Workshop on African Natural Language Processing (AfricaNLP 2025), pages 74–81. find
  25. Claytone Sikasote and Antonios Anastasopoulos. 2021. Bembaspeech: A speech recognition corpus for the bemba language. arXiv preprint arXiv:2102.04889. find
  26. Ewald van der Westhuizen and Thomas R Niesler. 2019. Synthesised bigrams using word embeddings for code-switched asr of four south african language pairs. Computer Speech & Language, 54:151–175. find
  27. Ewald van der Westhuizen, Trideba Padhi, and Thomas Niesler. 2021. Multilingual training set selection for asr in under-resourced malian languages. In International Conference on Speech and Computer, pages 749–760. Springer. find
  28. Jiayi Wang, David Ifeoluwa Adelani, Sweta Agrawal, Marek Masiak, Ricardo Rei, Eleftheria Briakou, Marine Carpuat, Xuanli He, Sofia Bourhim, Andiswa Bukula, and 1 others. 2024. Afrimte and africomet: Enhancing comet to embrace under-resourced african languages. In Proceedings of the 2024 Conference o… find
  29. Isaac Wiafe, Jamal-Deen Abdulai, Akon Obu Ekpezu, Raynard Dodzi Helegah, Elikem Doe Atsakpo, Charles Nutrokpor, Fiifi Baffoe Payin Winful, and Kafui Kwashia Solaga. 2025. Advancing automatic speech recognition for low-resource ghanaian languages: Audio datasets for akan, ewe, dagbani, dagaare, and i… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mgonzo-2026-learning-scarcity-building-benchmarking-speech/