Datasets:

madoss
/

faso-speech-dioula-digits

license: cc-by-4.0
language:
  - dyu
  - bm
task_categories:
  - audio-classification
task_ids:
  - keyword-spotting
pretty_name: Faso Speech Dioula Digits
tags:
  - audio
  - speech
  - burkina-faso
  - dioula
  - bambara
  - digits
  - spoken-digits
  - classification
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

Faso Speech Dioula Digits

A spoken-digit classification dataset in Dioula (Jula/Bambara), built from the Zenodo record 8320370 archive (DOI: 10.5281/zenodo.8320370).

Each clip is one speaker saying a single digit, 1 through 4, in Dioula. Recordings vary in speaker, accent, and recording environment.

Dataset Summary

Split	Rows	Duration	Per-class rows
train	1,532	01:34:14.9	383 / 383 / 383 / 383
validation	168	00:10:08.2	42 / 42 / 42 / 42

The split is a deterministic, per-class 90/10 hold-out (seed 42), so both splits have an even class balance across digits 1-4.

Data Fields

Column	Description
`audio`	Embedded Hugging Face `Audio(decode=False)` value with `bytes` and `path`
`text`	The spoken digit as a string (`"1"`-`"4"`)
`label`	The spoken digit as a `ClassLabel` (`0`-`3`, names `"1"`-`"4"`)
`language`	`dyu`
`duration`	Clip duration in seconds
`id`	Stable identifier, `dioula-digits:Class_<n>/<original_file_stem>`
`content_type`	`spoken-digit`
`source`	`zenodo-8320370`
`license`	`CC-BY-4.0`
`attribution`	Source authors, see below
`source_url`	https://zenodo.org/records/8320370

Usage

from datasets import Audio, load_dataset

ds = load_dataset("madoss/faso-speech-dioula-digits")
ds = ds.cast_column("audio", Audio(decode=False))

print(ds["train"][0]["text"], ds["train"][0]["label"])
print(ds["train"][0]["audio"])

Source Data

Original archive: AudiosDioula.zip, four folders Class_1-Class_4, one .wav file per recording, no changes made to the audio itself. Rows were grouped by folder into label/text.

Per the source README, the data was collected for the research project "Setting up a speech recognition model for under-resourced languages", targeting a Dioula digit-recognition model.

Licensing and Attribution

Licensed CC-BY-4.0 by the original authors:

KEITA Zakaria Cheick Oumar and BATONIO Fabrice (Universite de Bordeaux / UNB / ESI). Data collection managed by NABALOUM Emile, project managers Dr SOME Borlli Michel and Dr DIALLO Gayo.

If you use this dataset, please credit the original authors and cite the Zenodo record (DOI 10.5281/zenodo.8320370) in addition to this repository.

Part of the madoss Burkina Faso speech data efforts, alongside madoss/faso-speech and madoss/faso-speech-plus. Kept as a separate repo because it is digit classification, not sentence-level ASR, and carries a clear CC-BY-4.0 license.