Hugging Face's logo

Datasets:
tonative
/
ushauri

Modalities:
Audio
Text
Formats:
parquet
Size:
< 1K
Tags:
audio
speech
civic
luo
dholuo
everyday
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ushauri / README.md
tonative-org's picture
tonative-org
Update README.md
f765296 verified
preview code
|
Raw
History Blame Contribute Delete
7.36 kB
metadata 
license: cc-by-nc-nd-4.0
dataset_info:
  features:
    - name: Domain
      dtype: string
    - name: Penjo (Question)
      dtype: string
    - name: Question Code
      dtype: string
    - name: Question Audio
      dtype: audio
    - name: Dwoko (Answer)
      dtype: string
    - name: Answer Code
      dtype: string
    - name: Answer Audio
      dtype: audio
    - name: Voice Contributor
      dtype: string
  splits:
    - name: train
      num_bytes: 125963284
      num_examples: 50
  download_size: 97580275
  dataset_size: 125963284
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
tags:
  - audio
  - speech
  - civic
  - luo
  - dholuo
  - everyday
  - question-answering
  - QA
  - dataset
pretty_name: luo
size_categories:
  - n<1K

Ushauri: A Luo Question-Answer Speech Dataset

Ushauri is a small, curated speech dataset in Dholuo (Luo) consisting of paired question-and-answer recordings across everyday life domains.

Ushauri is designed to support the development of speech technology that serves the Luo community. The domain coverage (telecommunications, education, agriculture, finance, market, transportation, etc.) reflects everyday areas where accessible voice interfaces could improve access to information and services. Researchers and developers working on civic technology, public service delivery, or digital inclusion for East African communities are encouraged to build on this work.

Dataset Summary

  • Language: Dholuo (Luo) — luo

  • Domains: 10 domains: Telecommunications, Cultural Activities, Education, Everyday Activities, Civil Activities, Legal Activities, Basic Agriculture, Transportation, Finance and Market.

  • Total pairs: 50 question-answer pairs

  • Total audio clips: 100 (50 questions + 50 answers)

  • Modalities: text + audio

  • Voice contributors: anonymized speakers, identified by gender and age range

    Note: Ushauri is a preview release — a small, high-quality sample of a larger body of Dholuo speech data. For access to extended datasets, additional speakers, or commissioned collections in Luo and other African languages, see the License section.

Dataset Structure

Each row in the dataset contains the following fields:

Column Type Description
Domain string Ten (10) Topic areas of the question-answer pair, written in Luo with English translation in parentheses.
Penjo (Question) string The question in Luo
Question Code string Unique identifier for the question audio
Question Audio audio Recording of the question
Dwoko (Answer) string The answer in Luo
Answer Code string Unique identifier for the answer audio
Answer Audio audio Recording of the answer
Voice Contributor string Anonymized speaker ID encoding gender and age range

Code format

Question and answer codes follow the pattern {DOMAIN}-{TYPE}-{NUMBER}:

  • DOMAIN — two-letter domain identifier (e.g., TL = Telecommunications, CL = Cultural Activities)
  • TYPE — QN for question, AN for answer
  • NUMBER — two-digit sequence within the domain Example: TL-QN-01 is the first question in the Telecommunications domain; its paired answer is TL-AN-01.

Voice contributor format

Speaker IDs follow the pattern {ID}_{gender}_{age_range}, e.g., 001_female_18_to_25. No personally identifying information is included.

Splits

The dataset ships as a single split, which can be used for evaluation.

Usage 

from datasets import load_dataset
 
ds = load_dataset("tonative/ushauri", split="train")
 
sample = ds[0]
print("Question (Luo):    ", sample["Penjo (Question)"])
print("Answer (Luo):      ", sample["Dwoko (Answer)"])
print("Domain:            ", sample["Domain"])
print("Voice contributor: ", sample["Voice Contributor"])
 
# Access the audio arrays
q_audio = sample["Question Audio"]
a_audio = sample["Answer Audio"]
print("Question audio:", q_audio["array"].shape, "@", q_audio["sampling_rate"], "Hz")
print("Answer audio:  ", a_audio["array"].shape, "@", a_audio["sampling_rate"], "Hz")

To play a clip in a Jupyter/Colab notebook:

from IPython.display import Audio
Audio(sample["Question Audio"]["array"], rate=sample["Question Audio"]["sampling_rate"])

Data Collection

Questions and answers were written by native Dholuo speakers around common everyday domains, then recorded by voice contributors. The recordings capture natural conversations suitable for Luo language technology research.

Intended Uses

  • Automatic speech recognition (ASR) for Dholuo
  • Text-to-speech (TTS) benchmarking and voice cloning research in low-resource settings
  • Spoken and text-based question answering
  • Cross-lingual and multilingual NLP research including African languages
  • Linguistic and cultural documentation of everyday Luo speech

Limitations and Considerations

  • Small scale. With 50 QA pairs, Ushauri is intended for evaluation, few-shot learning, and demonstration rather than large-scale model training from scratch.
  • Speaker diversity. The pool of voice contributors is limited; models trained or evaluated on Ushauri alone may not generalize across all Luo speakers, dialects, or age groups.
  • Domain coverage. Domains are broad but shallow — a handful of items per topic.

License

This preview dataset is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).

Under this license, you may download, share, and reference the dataset for non-commercial purposes with appropriate attribution. You may not:

  • use the dataset for commercial purposes,
  • redistribute modified or derivative versions of the data or recordings,
  • incorporate the recordings into commercial products, services, or model training pipelines intended for commercial deployment.

Commercial use, extended datasets, and custom collections

Ushauri is a preview release intended to demonstrate the quality and structure of speech data that [Tonative Africa] can produce. For commercial licensing, larger domain-specific datasets, additional Luo speakers, or new language collections, please contact us via services@tonative.org.

Citation

If you use Ushauri in your research, please cite it as:

@dataset{tonative_ushauri_2026,
  title     = {Ushauri: A Luo Question-Answer Speech Dataset (Preview)},
  author    = {{Tonative Africa}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/tonative/ushauri}
}

A CITATION.cff file is provided in the repository root for automatic citation generation.

Acknowledgments

We gratefully acknowledge the voice contributors who recorded the questions and answers, and the native Dholuo speakers and language experts who wrote and reviewed the source texts. Contributor identities are anonymized in the dataset in accordance with our data handling practices.

Contact

For questions, corrections, or contributions, please open a discussion on the dataset repository or email us directly: services@tonative.org