bambara-asr-v2

Multi-corpus Bambara speech — 185,708 examples, ~366 hours, 53.7 GB of Parquet. Seven configs, each a train / dev / test triple of 16 kHz audio paired with a text target. Every config draws on a single upstream corpus, so you can mix and weight them yourself.

Access is gated with manual approval — request it on the dataset page and authenticate (hf auth login or HF_TOKEN) before loading.

Load

from datasets import load_dataset

jeli = load_dataset("djelia/bambara-asr-v2", "bm-to-bm-jeli-asr")
print(jeli["train"][0]["text"])

Stream the large configs rather than downloading them:

weak = load_dataset(
    "djelia/bambara-asr-v2", "bm-to-bm-weak", split="train", streaming=True
)
for row in weak.take(3):
    print(round(row["duration"], 2), row["text"][:80])

Configs

Config	Task	Text	Train	Dev	Test	Hours
bm-to-bm-weak	ASR	Bambara	37,600	6,298	3,102	171.0
bm-to-bm-synthetic	ASR	Bambara	25,679	7,374	3,632	69.2
bm-to-bm-jeli-asr	ASR	Bambara	22,360	6,420	3,163	29.1
bm-to-bm-bible	ASR	Bambara	7,832	1,312	647	39.1
bm-to-bm-code-switching	ASR	Bambara + French	1,090	182	91	1.1
synthetic-code-switching	ASR	Bambara + En/Fr	4,633	777	383	20.0
bm-to-en	Translation	English	37,193	10,679	5,261	36.7

Fields

Field	Description
audio	16 kHz, MP3 bytes
text	Transcript, or English translation in bm-to-en
duration	Seconds
source_dataset	Upstream corpus; named source in bm-to-bm-code-switching
language_sequence	Only in synthetic-code-switching, e.g. "bambara,english,bambara"

Notes

bm-to-bm-weak carries weakly supervised transcripts — training material rather than an evaluation reference. Its clips range from 0.017 s to over 160 s, so filter by length before batching.

duration is float64 in four configs and float32 in three, and the provenance column is renamed in one, so cast and rename before concatenate_datasets.

bm-to-en targets are English and carry a leading space.