Arabic-STT/sada2022

SADA: Saudi Audio Dataset for Arabic (original files)

An unmodified copy of the Kaggle dataset SADA صدى, sdaiancai/sada2022 (version 23, files dated 2023-08-31), published by the National Center for Artificial Intelligence at the Saudi Data and Artificial Intelligence Authority (SDAIA) with the Saudi Broadcasting Authority (SBA).

Contents

• Every file of the Kaggle dataset, at its original path and byte for byte: train.csv, valid.csv, test.csv and the 4,563 full-episode WAV files in batch_1/ to batch_4/ (16 kHz, mono, PCM s16le). 4,567 files, 77.06 GB, as listed by Kaggle.
• Nothing was converted, cut, resampled or re-encoded. Kaggle zips some files in transit; only that transport wrapper was removed, and every file has exactly the size Kaggle lists for it.
• Kaggle's own README.md (dataset description, column definitions, split statistics, licence and citation) is kept byte for byte at kaggle/README.md, because README.md in this repository is this card.
• Segments are not separate files: each CSV row gives FileName (batch_N/<episode>.wav), SegmentStart and SegmentEnd in seconds, the transcript (GroundTruthText, ProcessedText) and speaker, dialect and environment fields.

Leaderboard relation

SADA is one of the sources of the Open Universal Arabic ASR Leaderboard. The segments of test.csv form a leaderboard test set and are for evaluation only; train.csv and valid.csv are in-domain training and validation data for that leaderboard.

Licence

CC BY-NC-SA 4.0, as stated by the dataset authors: attribution, non-commercial use only, share-alike.

Citation

@inproceedings{SADA2023,
  Title= {SADA - SBA & SDAIA Audio Dataset for Arabic},
  Author= {Areeb Alowisheq,Abdullah Alrajeh, Sadeen Alharbi Abdulmajeed Alrowithi, Aljawharah Bin Tamran, Asma Ibrahim, Raghad Aloraini, Raneem Alnajim,
  Ranya Alkahtani, Renad Almuasaad, Sara Alrasheed, Shaykhah Alsubaie, Yaser Alonaizan},
  Booktitle = {To be published},
  affiliation = {NCAI-SDAIA}
  Year 	= {2023}
}