goaicorp/dyu-whisper-small
Automatic Speech Recognition • 0.2B • Updated • 3
Datasets:
Request access to dyu-speech
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
dyu-speech is shared under CC BY-SA 4.0. The data comes from the Koumankan4Dyula corpus by UVCI and Data354 (AI4D Africa); please credit them. Please read the card, its terms and licenses before requesting access, and make sure you understand them.
Log in or Sign Up to review the conditions and access this dataset content.
dyu-speech
Publicly available Dioula speech, aligned with its text, for speech recognition, speech translation and text-to-speech.
-022871){kind=icon}
dyu-speech holds Dioula (ISO 639-3 dyu) speech aligned with its text. This version holds one corpus, Koumankan4Dyula.
Access is gated and requests are reviewed by hand. Collecting and cleaning Dioula speech takes a lot of time and resources, and we do it mostly out of passion. Please read this card, its terms and licenses before requesting access, and make sure you understand them. Respecting them keeps this work useful to the community and helps move our local languages forward. The license is CC BY-SA 4.0: no further terms are added.
Résumé en français. Parole en dioula alignée avec son texte, pour la reconnaissance vocale, la traduction de la parole et la synthèse vocale. Cette version contient un corpus public, Koumankan4Dyula (UVCI et Data354, AI4D Africa) : 10 929 phrases lues (13,3 h) avec leur transcription en dioula et leurs traductions française et anglaise. Accès sur demande, examinée à la main : lisez cette carte, ses conditions et ses licences avant de demander l'accès. Licence CC BY-SA 4.0, sans conditions supplémentaires ; citez UVCI et Data354 et partagez vos adaptations sous la même licence.
Overview
| Split | Clips | Hours |
|---|---|---|
train |
8,065 | 9.73 |
dev |
1,471 | 1.81 |
test |
1,393 | 1.75 |
| Total | 10,929 | 13.29 |
All clips are read sentences with Dioula transcripts and French and English translations. The splits are Koumankan's official ones, so results stay comparable with other work on this corpus. Audio is WAV, 24-bit mono, 48 kHz (10 clips at 44.1 kHz); clips load at their own rate, so resample as your model needs. There is no speaker id. Gender F 6,063 / M 4,866; country Burkina Faso 10,008, Côte d'Ivoire 883, not given 38; age groups from 15 to 65.
Quick start
from datasets import load_dataset, Audio
ds = load_dataset("goaicorp/dyu-speech", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16_000)) # for most ASR models
Log in first (hf auth login) with an account that has access.
Columns
| Column | Description |
|---|---|
audio |
The recording |
dyu |
Dioula transcript |
fr, en |
French sentence (from Common Voice) and its English translation |
dyu_id, fr_id |
File names; the number is the Common Voice id of the French sentence |
gender, age_group, country |
Speaker information as given by the source |
duration, sampling_rate |
Measured from each audio file |
commonvoice_split |
Split of the French sentence in Common Voice |
Sources and credits
All rows come from Koumankan4Dyula by the Université Virtuelle de Côte d'Ivoire (UVCI) and Data354, built in the Koumankan project under the AI4D Africa programme (IDRC, SIDA, ACTS) and released under CC BY-SA 4.0: uvci/Koumankan_mt_dyu_fr (text) and uvci/koumankan4dyula (speech). The French sentences come from Common Voice (CC0); a linguist translated them into Dioula. Source contacts: ismael21.kone@uvci.edu.ci, info@data354.com.
How it was built
- Copied from uvci/koumankan4dyula: same rows, audio, columns and splits.
- Test transcripts restored. The source leaves
dyuempty for its 1,393 test clips (a hidden competition test). We joined them from the text release (uvci/Koumankan_mt_dyu_fr) by Common Voice id. - Durations measured. 282 rows had a placeholder duration (1.0); every
durationandsampling_ratenow comes from the WAV header. The source card gives about 15 h (10.5 h in its table); the audio measures 13.29 h. - Duplicate removed. goaicorp/new-dioula-speech-clean held the same rows and audio, byte for byte, without the metadata. It was deleted in November 2026.
Evaluation
Select checkpoints on dev and report final scores (WER and CER) on test. Two cautions:
- 53 Dioula sentences of
trainare also read indevortestby other speakers, since several sentences were recorded more than once. - dyu-fr-parallel trains on all Koumankan text,
devandtestsentences included. Do not use a text model trained on it (such as a language model for decoding, or a speech translation model) when you report scores on this test.
Limits
- Read speech only, one source, mostly Burkinabè speakers.
- Dioula spelling varies between translators; no normalization was applied.
- No speaker id, so speaker-independent splits cannot be checked.
- The English was most likely translated from the French, not the Dioula; its method is not documented.
Access
Fill in the form on this page. Requests are reviewed by hand.
License
CC BY-SA 4.0, as the source. Credit UVCI and Data354 (Koumankan4Dyula, AI4D Africa) and GO AI Corp, and share adaptations under the same license. The French sentences come from Common Voice: do not try to identify the speakers.
Citation
@misc{goaicorp2026dioulaspeechcorpora,
title = {dyu-speech: Publicly available Dioula speech, aligned with its text},
author = {Nyantudre, Wendyellé Abubakrh Alban and Bandaogo, Aristide},
year = {2026},
howpublished = {Hugging Face Datasets (gated access), GO AI Corp},
url = {https://huggingface.co/datasets/goaicorp/dyu-speech}
}
@misc{uvci2024koumankan4dyula,
title = {Koumankan4Dyula: A Speech-To-Text Translation Corpus For Dyula Language},
author = {{Université Virtuelle de Côte d'Ivoire} and {Data354}},
howpublished = {Hugging Face Datasets},
url = {https://huggingface.co/datasets/uvci/koumankan4dyula}
}
Contacts
- Wendyellé A. Alban NYANTUDRE: nyantudre@gmail.com
- Aristide BANDAOGO: aristide@goaicorporation.org
- GO AI Corp
- Downloads last month
- 163