Hugging Face's logo

Datasets:
burkimbia
/
speech-dataset-public

Tasks:
Automatic Speech Recognition
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
Mossi
Size:
10K - 100K
Tags:
moore
mossi
burkina-faso
low-resource
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Mooré speech corpus

Mooré speech with its transcription, for speech recognition and speech synthesis: about 150 hours in 92,092 utterances from three source families (Bible readings from jw.org and bible.com, and recorded oral tradition), cut into utterances and resampled to 16 kHz, with a speaker label on each row.

Prepared by the BurkimbIA team.

# pip install "datasets[audio]<4.0.0"
import os
from datasets import load_dataset

# Gated dataset: request access on this page, then pass a token of the approved account.
ds = load_dataset("burkimbia/speech-dataset-public", split="train", token=os.environ["HF_TOKEN"])

Usage and license

Research and non-commercial use only. The recordings belong to their publishers: each source keeps its own terms, listed per source below. None of them is under an open license; they are included for research under gated access. Access is granted manually; by requesting it you agree to respect each source's terms and to cite this dataset.

Splits

split utterances
train 92,092

There is no validation or test split.

Columns

column meaning
text Mooré transcription of the utterance.
audio The utterance, 16 kHz.
duration Length in seconds.
speakers Speaker label. For the two Bible sources, a diarization cluster (BIBLE.JW-SPEAKER_28), not a person: one reader can carry several labels. For oral tradition, the reader's name.
source Source family of the row.
gender male or female, set per source or per reader, not detected per utterance.

Sources

family origin terms
Bible (jw.org) Mooré Bible audio and text from jw.org, collected with jwsoup. Each audio file was cut by hand into segments, start and end times written in spreadsheets by the contributors credited below, and the segments matched to the text page by page. © Watch Tower Bible and Tract Society of Pennsylvania. The jw.org terms of use allow personal, non-commercial use and do not allow posting the content on other websites. No open license.
Bible (bible.com) Mooré Bible audio and text from bible.com, version Wẽnnaam Sebre 2014 (MPBU), collected and aligned with MooreSpeechCorpora (MMS forced alignment), published as anyantudre/moore-speech-bible. Text: Moore Bible © 2014, Bible Society of Burkina Faso. Audio: Moore Old Testament Audio Bible ℗ 2022, Davar Partners International. No license stated.
Oral tradition Tales (contes), proverbs (proverbes) and riddles (devinettes) read in Mooré, from mooreburkina.com (audio served by media.ipsapps.org). No license or terms stated on the site; its copyright page gives burkinalangues@gmail.com as contact.

Processing

  • Speaker diarization of both Bible sources with pyannote.audio.
  • Music removal with Spleeter; proverbs denoised with resemble-enhance.
  • All sources resampled to 16 kHz and concatenated into one schema.

Known issues

  • Most of the audio is Bible reading by male voices. A model trained on this corpus alone will hear women, conversation and noisy speech much less well.
  • Only the jw.org segments were cut by hand. The bible.com segments come from automatic forced alignment.
  • Speaker labels of the two Bible sources are diarization clusters computed separately for each source: they do not identify readers across sources.

Credits

Changes to this card

The data is the upload of 13 September 2025, unchanged. The card now documents:

  • the terms of each source; the previous card stated no license;
  • the sampling rate, 16 kHz; the previous card also said "standardized to 15kHz";
  • which segments were cut by hand; the previous card said every row was "manually validated by contributors", but only the jw.org segmentation was done by hand.

Citation 

@misc{burkimbia2025speechdatasetpublic,
  title        = {Moor{\'e} Speech Corpus},
  author       = {{BurkimbIA Community}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/burkimbia/speech-dataset-public}}
}

Contact: BurkimbIA, through the Community tab of this page.
Downloads last month
2,339
Total file size:
13.7 GB