bambara-asr

Multi-task Bambara speech: transcription, speech-to-text translation into French and English, and a multilingual training mix. 16 kHz audio in Parquet across nine configs.

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

ds = load_dataset("djelia/bambara-asr", "bm-to-bm", split="train")

Every config has train and test splits.

Configs

Config	Rows	Train / Test	Size	Task
multi-combined	224,366	201,622 / 22,744	39.02 GB	multilingual mix
bm-to-en	53,133	47,819 / 5,314	25.35 GB	Bambara speech → English text
bm-to-fr	53,133	47,819 / 5,314	25.35 GB	Bambara speech → French text
bm-to-fr-translated	53,133	47,819 / 5,314	25.35 GB	as above, plus translation
bm-to-bm	41,734	37,560 / 4,174	4.48 GB	Bambara transcription
bm-to-bm-v2	41,734	32,439 / 9,295	4.49 GB	same rows, different split
en-to-en	11,961	10,764 / 1,197	4.84 GB	English transcription
fr-to-fr	27,000	24,300 / 2,700	1.07 GB	French transcription
bm-to-bm-code-switching	1,363	1,294 / 69	0.03 GB	code-switched Bambara

Fields

audio (16 kHz), text, duration, source_dataset — plus translation in bm-to-fr-translated, and language / task_type in multi-combined. bm-to-bm-code-switching names the provenance column source.

Notes

bm-to-en, bm-to-fr and bm-to-fr-translated are one audio set with three target sets, and bm-to-bm-v2 is bm-to-bm re-split — load one of each group rather than several.

multi-combined is a multilingual mix spanning 26 languages; its language and task_type columns let you slice it back down. Stream it rather than downloading 39 GB:

ds = load_dataset("djelia/bambara-asr", "multi-combined", split="train", streaming=True)
bambara = ds.filter(lambda row: row["language"] == "bambara")

duration is float32 in some configs and float64 in others — cast before concatenating.