A sample of five public Arabic speech datasets, repackaged with one schema for ASR experiments. Each dataset is in its own folder, with one subfolder per split:

<dataset>/train/*.parquet   # small random sample
<dataset>/test/*.parquet    # full official test split

test: the complete official test split, so results can be compared with published numbers.
train: a seeded random sample (seed 42), stratified by subset (country, sub-source, or clean/noisy) in proportion to that subset's hours.
No dev/validation splits are included.

dataset	variety	train (sample)	test (full)
casablanca	8 dialects: Algeria, Egypt, Jordan, Mauritania, Morocco, Palestine, UAE, Yemen	15.43h · 12,778 utts (all)	7.85h · 6,818 utts
linto_tunisian	Tunisian (some FR/EN code-switching)	21.02h · 19,417 utts	1.24h · 1,013 utts
masc	MSA + dialects (YouTube), clean & noisy	14.00h · 14,164 utts	19.42h · 18,006 utts
sada	Saudi dialects (Najdi, Hijazi, Khaliji, …) + others, TV shows	21.11h · 12,062 utts	10.75h · 6,190 utts
scc22	Saudi conversational, Arabic–English code-switching	—	4.45h · 3,296 utts
total		71.56h	43.71h

Total: 115.3h, 7.2 GB.

Loading

from datasets import load_dataset

ds = load_dataset("HoussamLCHMT/audio-120h", "sada", split="test")
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

Audio is 16 kHz mono, 16-bit FLAC, stored inside the Parquet files.

Schema

column	description
`audio`	16 kHz audio
`id`	utterance id (original segment id where available)
`dataset`	source dataset
`split`	`train` / `test`
`subset`	casablanca: country · linto: sub-source · masc: `clean`/`noisy` · sada: TV show · scc22: environment
`dialect`	label as given by the source (ISO code or SADA dialect name); null for masc
`speaker`, `gender`	where the source provides them
`text`	original transcript
`text_normalized`	source-provided normalized transcript where available (null for casablanca)
`duration`	seconds

Processing notes

sada: segments cut from the long recordings using the official train.csv/test.csv start/end times. text = GroundTruthText, text_normalized = ProcessedText. Segments with empty transcripts are dropped.
linto_tunisian: train rows whose audio path appears twice in the source manifest (AmenyKH filename collisions) are dropped.
masc: original float WAV converted to 16-bit.
casablanca: all of the train split is included (~15h).

Licenses and attribution

This repo redistributes third-party data. Each subset keeps its original license, so check it before you use that subset.

dataset	license	source
casablanca	CC BY-NC-ND 4.0	UBC-NLP/Casablanca
linto_tunisian	CC BY 4.0	linagora/linto-dataset-audio-ar-tn
masc	CC BY 4.0	pain/MASC
sada	CC BY-NC-SA 4.0	SDAIA (NCAI) & Saudi Broadcasting Authority, released on Kaggle
scc22	CC BY-NC-SA 4.0	SDAIA / Kaggle (HF mirror)

Several of these licenses are non-commercial, so use this collection for non-commercial research only.

Citations

@article{talafha2024casablanca,
  title={Casablanca: Data and Models for Multidialectal Arabic Speech Recognition},
  author={Talafha, Bashar and Kadaoui, Karima and Magdy, Samar Mohamed and others},
  journal={arXiv preprint arXiv:2410.04527},
  year={2024}
}
@misc{linagora2024Linto-tn,
  title={LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect},
  author={Hedi Naouara and J{\'e}r{\^o}me Louradour and Jean-Pierre Lorr{\'e}},
  year={2025}, eprint={2504.02604}, archivePrefix={arXiv}
}
@inproceedings{10022652,
  author={Al-Fetyani, Mohammad and Al-Barham, Muhammad and Abandah, Gheith and Alsharkawi, Adham and Dawas, Maha},
  booktitle={2022 IEEE Spoken Language Technology Workshop (SLT)},
  title={MASC: Massive Arabic Speech Corpus}, year={2023}, pages={1006-1013},
  doi={10.1109/SLT54892.2023.10022652}
}
@inproceedings{SADA2023,
  title={SADA - SBA \& SDAIA Audio Dataset for Arabic},
  author={Alowisheq, Areeb and Alrajeh, Abdullah and Alharbi, Sadeen and others},
  year={2023}, affiliation={NCAI-SDAIA}
}
@misc{SCC2025,
  title={Saudilang Code-Switch Corpus (SCC)}, author={SDAIA}, year={2022},
  howpublished={\url{https://www.kaggle.com/datasets/sdaiancai/saudilang-code-switch-corpus-scc}}
}