bambara-audio-b

Bambara speech derived from scripture recordings, published in four processing stages: raw segments, a length-filtered version, a speaker-diarized long-form cut, and a CTC forced-alignment cut. 30.55 GB of Parquet.

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-audio-b", "short-filtered", split="train")
print(ds[0]["text"], ds[0]["prediction"], ds[0]["nb_diff"])

Stream the two large configs rather than downloading ~16 GB each:

ds = load_dataset("djelia/bambara-audio-b", "default", split="train", streaming=True)

Configs

Config	Rows	Size	Median duration	What it is
default	46,248	17.29 GB	2.65 s	Short segments, human text + machine prediction
short-filtered	44,478	16.23 GB	2.65 s	default with the longest items removed
ctc_alignment	17,157	1.11 GB	7.92 s	Re-segmented by CTC forced alignment
long-with-speaker	9,791	1.14 GB	17.18 s	Long clips, diarized speakers and embeddings

Every config has a single train split.

Fields

Config	Fields
default	audio, text, prediction, duration, path
short-filtered	as default, plus nb_text, nb_prediction, nb_diff
ctc_alignment	audio, text, duration, source, methode
long-with-speaker	audio, duration, speaker, text, weak_transcription, text_from_weak_transcription, filename, speaker_embedding

nb_diff is the length difference between the human and machine transcripts — a cheap disagreement signal, though it measures length rather than edit distance.

Notes

The four configs are views over overlapping source audio, not four independent corpora — load one rather than several.

text can be empty in default and short-filtered; drop those rows with ds.filter(lambda row: len(row["text"].strip()) > 0).

No config has a held-out split, so build your own — speaker-disjoint where speaker exists. long-with-speaker is dominated by SPEAKER_00, roughly 73% of its rows. prediction is machine output, not a reference transcript.