Datasets:

MOH749
/

Bocalantics

license: cc-by-sa-4.0
language:
  - wo
  - ff
  - bm
task_categories:
  - automatic-speech-recognition
pretty_name: Bocalantics
tags:
  - west-african
  - wolof
  - pulaar
  - fula
  - bambara
  - asr
  - speech
  - low-resource
size_categories:
  - 100K<n<1M

Bocalantics

One normalized, license-clean speech corpus for West African languages — Wolof, Pulaar/Fula, and Bambara — merged from the fragmented community datasets into a single source you can train on anywhere with load_dataset.

West African languages are spoken by one interconnected, multilingual people; Bocalantics treats them as one corpus so a single model can learn across them.

Schema

field	type
`audio`	Audio (16 kHz)
`text`	string — the transcription
`lang`	string — `wo` / `ff` / `bam`
`source`	string — upstream dataset id
`license`	string — upstream license
`duration`	float — seconds (0.0 where the source omitted it)

Splits: train / validation / test.

Composition (clips)

lang	clips
bam	38769
ff	17253
wo	86729

Provenance & attribution

Built only from clean-licensed sources, each row carrying its source and license. Please credit the upstream creators:

source	lang	license	credit	link
`soynade-research/Wolof-ASR-Data`	wo	CC-BY-SA-4.0	soynade-research	link
`ngia/ASR_pulaar`	ff	CC-BY-4.0	ngia	link
`RobotsMali/bam-asr-early`	bam	CC-BY-4.0	RobotsMali AI4D Lab (Jeli-ASR, Mali-Pensé / A. Ouattara)	link

License

The collection is released under CC-BY-SA-4.0 (the share-alike umbrella, since one source is CC-BY-SA; the CC-BY sources are relicensed into the collection with attribution preserved). Attribute the upstream creators above.

Load

from datasets import load_dataset
ds = load_dataset("MOH749/Bocalantics", split="train")
# per-language: ds = ds.filter(lambda x: x["lang"] == "wo")

Roadmap

v1.1: add Sereer + more Wolof/Pulaar via Kallaama (CC-BY-4.0, Jokalante), once re-hosted from github.com/gauthelo/kallaama-speech-dataset.
Mande siblings (Soninke, Mandinka) await licensed speech data.