metadata

dataset_info:
  - config_name: human-reviewed
    features:
      - name: audio
        dtype: audio
      - name: duration
        dtype: float64
      - name: semi-label
        dtype: string
      - name: corrected-label
        dtype: string
    splits:
      - name: train
        num_bytes: 9838671594.158
        num_examples: 33282
      - name: test
        num_bytes: 1754436439.925
        num_examples: 5775
    download_size: 11503545995
    dataset_size: 11593108034.083
  - config_name: semi-first
    features:
      - name: audio
        dtype:
          audio:
            decode: false
      - name: duration
        dtype: float64
      - name: semi-label
        dtype: string
    splits:
      - name: train
        num_bytes: 12219340374.572
        num_examples: 41366
    download_size: 12146855870
    dataset_size: 12219340374.572
  - config_name: semi-second
    features:
      - name: audio
        dtype:
          audio:
            decode: false
      - name: duration
        dtype: float64
      - name: semi-label
        dtype: string
    splits:
      - name: train
        num_bytes: 11588268645.052
        num_examples: 38502
    download_size: 23763534431
    dataset_size: 11588268645.052
configs:
  - config_name: human-reviewed
    data_files:
      - split: train
        path: human-reviewed/train-*
      - split: test
        path: human-reviewed/test-*
  - config_name: semi-first
    data_files:
      - split: train
        path: semi-first/train-*
  - config_name: semi-second
    data_files:
      - split: train
        path: semi-second/train-*
license: cc-by-sa-4.0
task_categories:
  - automatic-speech-recognition
  - translation
language:
  - bm
  - fr
tags:
  - semi-labelled
  - asr
  - speech-recognition
  - code-switching
  - bambara
pretty_name: Kunnafonidilaw ka cadeau (Kunkado)
size_categories:
  - 100K<n<1M

Kunnafonidilaw ka cadeau 🇲🇱

A messy‑real Bambara ASR corpus for developing modern speech models & code‑switch studies

Quick Facts

	value
Total duration	161.15 h
Reviewed subset	39.3 h (≈ 25 %)
Total segments	118 925
Languages	Bambara (majority) • French (code‑switch) • misc. Arabic (translit)
LICENSE	CC‑BY‑SA 4.0

kunkado aims to mirror how Malians speak bambara today: fast, informal, and full of French code‑switching. We hope it fuels robust ASR systems and research on contact phenomena in Mande languages.

Motivation

Low‑resource corpora are often small & squeaky‑clean, giving models a shock once deployed. For this dataset we wanted as features all the things that we conventionally reject in manageable ASR datasets. It is especially designed for training end-to-end Deep Learning systems as the learning task is quite complex (if no normalization applied). In this dataset you will encounter:

code‑switches (tagged with __double underscores__)
music, jingles & phone buzzes
accept rough silence‑proxy segmentation (cut words are marked with …)

Only ~25 % could be human‑reviewed in the project timeframe, but community PRs are welcome.

Characteristics of the dataset

Present Bamako Bambara: Reflective of how Malian people naturally speak Bamanankan. → This includes urban speech patterns, contractions, and informal expressions commonly heard in Bamako.
Broad range of topics: → Content spans casual conversations, news, politics, religion, comedy, marketplace discussions, and social commentary.
Numbers transcribed as digits: → This choice was intented primarily for faster human transcription and unifying the semi labels which sometimes uses different style
A lot of code-switching (represented in the transcriptions with underscores): → French & Arabic insertions are delimited with __ markers, making it easier to identify multilingual segments.
Most segments feature more than one voice and interactions/interference between speakers: → This reflects the natural occurrence of overlapping speech in real-world dialogues and group settings.

Duration buckets (seconds)

bucket (s)	human‑reviewed	semi-first	semi‑second	total
0.6 – 15	39 057	41366	38 502	111 746
15 – 30	0	0	5 402	5 402
30 – 45	0	0	1 777	1 777

Subsets

The dataset has been uploaded in three subsets, the human reviewed subset is separated from the other for organization purposes and the semi labelled entries have been slitted into two subsets due to resources constraints during upload

human‑reviewed (default)

Total 39.27 h – 39 057 short utterances
Splits • Train 33 282 utt. – 33.47 h • Test   5 775 utt. – 5.80 h (≈ 15 %)

semi-first

Total 41.47 h – 41 366 short utterances

semi-second

Total 80.42 h – 38 502 variable length utterances (0.6 to 45s)

Tag	Meaning
`<BRUITS>`	generic noise
`<INCOMPRÉHENSIBLE>`	fully inaudible speech
`<CHEVAUCHEMENT>`	speaker overlap
`<RIRES>`	laughter
`<MUSIQUE>`	music / jingle (no lyrics)
`<TOUX>`	cough
`<INVOCATION>`	prayers, quranic excerpts
`<ECHO>`	echo artefact
`<APPLAUDISSEMENTS>`	applause
`<CRIS>`	screams
`<PLEURES>`	crying

Recommended Normalisation

Numbers appear mostly as digits and may violate a single style. We strongly advise applying number normalization them before training.

Source & Provenance

Donor	Hours	Media type
Radio Benkouma “La voix du Baramousso”	32.7	Community Radio
Mousso TV	23.2	TV
ORTM (National TV)	7	TV/Radio
Radio Sahel FM	98.4	Regional Radio

Audio was graciously provided by the broadcasters listed above; the complete corpus is released under CC‑BY‑SA 4.0. Automatic transcripts were generated with soloni‑114M and then manually corrected for ~40 h by our team of annotators.

NB: Semi Labels might be updated in future versions

Annotators

Karounga Kanté • Boureima Traoré • Diakaridia Bengaly • Tidiane Koné • Lanseni Mallé • Séni Togninè • Assanatou Soumaoro • Alassane Koné • Benogo Fofana • Aboubacar Traoré

Huge thanks to our donors & reviewers – aw ni ce!

Known Issues & Caveats

Segmentation: silence‑proxy; some utterances cut mid‑word.
Spelling issues: Misspelling of foreign phrases and arabic transliterations
Code‑switch inconsistencies: Arabic phrases sometimes tagged, sometimes not.
Number style: digits vs. letters not strictly respected.
Rare pure French segments remain in the dataset.

Usage Example

from datasets import load_dataset

ds = load_dataset("RobotsMali/kunkado", split="train")
print(ds[0]["corrected-label"])

License

Creative Commons Attribution–ShareAlike 4.0 International (CC‑BY‑SA 4.0). You may share and adapt, even commercially, as long as you credit the contributors and keep derivatives under the same licence.* No warranty.

Citation

@misc{diarra_kunkado_2025,
  title        = {kunnafonidilaw ka cadeau: an {ASR} dataset to power the development of models that understands present-Day Bambara},
  author       = {RobotsMali AI4D Lab},
  year         = 2025,
  howpublished = {Hugging Face Datasets},
  note         = {\url{https://huggingface.co/datasets/RobotsMali/kunkado}}
}

Maintained by RobotsMali AI4D Lab — PRs & issue reports welcome!

Datasets:

RobotsMali
/

kunkado