oddadmix/yodas-v3-ar-labeled-split

Arabic YODAS v3 clips, split by dialect

The same clips as oddadmix/yodas-v3-ar-labeled, regrouped so each dialect is its own split. Audio is copied verbatim — identical Opus bytes, not re-encoded, so this is not a further lossy generation.

from datasets import load_dataset
ds = load_dataset("oddadmix/yodas-v3-ar-labeled-split", split="ma")        # Moroccan only
ds = load_dataset("oddadmix/yodas-v3-ar-labeled-split", split="msa", streaming=True)

2,989,202 clips, 20,718 hours across 14 splits.

Split	Clips	Hours	Share of hours
msa	1,706,463	11,861	57.2%
eg	531,912	3,681	17.8%
ma	145,591	1,014	4.9%
tn	130,160	903	4.4%
dz	121,202	840	4.1%
ye	76,807	525	2.5%
bh	68,372	467	2.3%
sd	51,931	354	1.7%
iq	49,194	339	1.6%
sy	33,595	231	1.1%
sa	30,704	207	1.0%
ps	18,533	127	0.6%
ly	12,591	86	0.4%
lb	12,147	83	0.4%

What the split labels mean

The label is dialect from the source dataset: predicted by Nawah-Dialect-BERT-6M on the whole source recording and copied to each clip cut from it, because one clip carries too little text to classify reliably.

These are predictions, not ground truth, and they read the transcript's language register rather than the speaker's accent — the transcripts being themselves machine-generated. dialect_probs keeps all 14 class probabilities, so you can re-threshold instead of trusting the argmax: a clip in the ma split may have had 0.34 on ma and 0.31 on dz.

unknown holds clips whose source recording had no transcript to classify.

Splitting for training

Split by source_id, not by row. Clips sharing a source_id come from one recording, so row-wise splitting puts the same speaker and topic on both sides of your split.

Provenance

CC-BY-3.0, inherited from YODAS v3 via oddadmix/yodas-v3-ar-labeled. See that card for how the cuts were chosen and what filtering was applied.