thiomi/thiomi-5k

Thiomi 5k

Motivation

This dataset release is part of the data and modeling effort described in: The Thiomi Dataset: A Large-Scale Multimodal Corpus for Low-Resource African Languages

The broader Thiomi work focuses on building high-quality multimodal resources for African languages and enabling ASR/MT/TTS research and deployment.

Data Structure

Per language config (eng_Latn, kik_Latn, kam_Latn, luo_Latn, mer_Latn, som_Latn):

Each config contains train and test splits. Every row has three fields: sentenceId (string), audio (HF Audio feature), and transcription (string). Row counts are identical across all language configs within a split — the 44 previously-missing English sentences have been removed from every language.

Dataset Statistics

Config	Language	Train Samples	Test Samples	Train Audio (h)	Test Audio (h)
eng_Latn	English	5,453	1,365	6.59	1.64
kik_Latn	Kikuyu	5,453	1,365	8.62	2.17
kam_Latn	Kamba	5,453	1,365	8.21	2.00
luo_Latn	Luo	5,453	1,365	8.65	2.16
mer_Latn	Kimeru	5,453	1,365	8.34	2.06
som_Latn	Somali	5,453	1,365	8.69	2.17
Total		32,718	8,190	49.10	12.21

Intended Uses

• TTS training and adaptation per language
• ASR training/evaluation with language-specific subsets
• Speech-text alignment and multilingual benchmarking
• Data selection and curriculum design for low-resource speech systems

Limitations

• English audio was generated via TTS, not recorded by human speakers.
• Language balance by hours differs despite approximately equal row counts.

Privacy and Ethics

Please follow applicable consent, privacy, and data governance requirements from the upstream collection and platform process described in the Thiomi paper.

Citation

If you use this dataset, please cite the Thiomi dataset paper:

@misc{mutisya2026thiomidatasetlargescalemultimodal,
      title={The Thiomi Dataset: A Large-Scale Multimodal Corpus for Low-Resource African Languages}, 
      author={Hillary Mutisya and John Mugane and Gavin Nyamboga and Brian Chege and Maryruth Gathoni},
      year={2026},
      eprint={2603.29244},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.29244}, 
}

Acknowledgements

Thanks to the Thiomi contributors, speakers, annotators, and reviewers who made multilingual speech data collection and curation possible.