abnajlae/darija-asr-corpus

Darija ASR Corpus (dataset-core)

Arabizi (Latin-script) transcriptions of Moroccan Darija speech, produced for a Whisper fine-tuning pipeline (paper not yet published -- citation forthcoming). This repo contains four source subsets: DODa, DVoice, Wiki, and YouTube. Each subset carries its own upstream license/terms -- see below -- because they are drawn from four different original projects.

Subsets

Config	Rows	Audio bundled?	Upstream license	Upstream source
doda	12,743	Yes (data/chunks/)	MIT License	atlasia/DODa-audio-dataset-V3
dvoice	2,991	Yes (data/chunks/)	CC BY 4.0	Zenodo record 6342622
wiki	1,756	Yes (data/chunks/)	CC BY-SA 4.0	atlasia/Moroccan-Darija-Wiki-Dataset
youtube	805	No -- see below	Not redistributable; platform ToS	8 individual YouTube channels, linked per-row

doda, dvoice, and wiki bundle their audio under data/chunks/<config>/<2-hex-shard>/<fileName>.wav (4.5 GB total, 17,490 files; the row's audioPath field gives the exact path). Files are sharded into 256 per-config subdirectories (first 2 hex characters of md5(fileName)) because Hugging Face's git backend rejects any single directory holding more than 10,000 files. youtube bundles no audio at all (see the dedicated section below). Each row's source_url still links back to the original upstream dataset for attribution and citation. Comply with each subset's specific license when reusing its audio (see License summary below) -- in particular, wiki is CC BY-SA 4.0, so redistributing it (or derivative works) further must retain attribution and remain share-alike.

The youtube subset: reconstruction, not redistribution

youtube contains this project's own annotations (transcripts, language tags, confidence scores, etc.) for 805 utterances drawn from 8 YouTube channels (13 distinct source videos). No audio is bundled for this subset, consistent with YouTube's terms of service -- audioPath is null for every row.

Every row instead carries exact sourceStartSample, sourceEndSample, sourceStartSeconds, sourceEndSeconds, and timestampConfidence (0-1) fields. These were recovered by cross-correlating the original chunk audio (kept privately, never redistributed) against a freshly re-separated copy of each source video -- i.e. locating each already-known chunk's exact position in the source video, rather than re-deriving segment boundaries from scratch. Validation result: 805/805 utterances recovered (100%), minimum confidence 0.73, mean confidence 0.96.

To reconstruct the audio yourself, use reconstruct_chunks.py (included in this repo):

pip install yt-dlp demucs torch soundfile librosa
python reconstruct_chunks.py path/to/youtube.jsonl --out-dir chunks/

This re-downloads each source video via yt-dlp, isolates vocals with Demucs (htdemucs), and slices the vocals track at the recorded sample offsets for each utterance -- an exact slice, not a re-run of any detection algorithm. This depends on the source video remaining available; if a video has since been removed, its transcript/annotation row remains usable as text data, but its audio cannot be recovered.

Fields

Common fields across all four subsets:

• fileName: unique record identifier, format {speakerId}__{utteranceId}__{attemptId}
• utteranceText / utteranceTextArabic: original scripted/reference text (Arabic script), where available
• spokenText: Arabizi (Latin-script) transcription of what was actually spoken -- the primary label for ASR training
• speakerId / speakerName: speaker identifier (anonymized where the upstream source did not already publish a name)
• audioPath: relative path to bundled audio, data/chunks/<config>/<shard>/<fileName>.wav
• durationSeconds, sampleRate, languageTag, confidence: metadata from the labeling pipeline
• source_license, source_url: upstream license and citation link

Subset-specific fields (e.g. intent/sentiment/genderLabel for dvoice, englishTranslation for doda, wikiTitle for wiki) are preserved as-is from each upstream source.

Orthographic convention

Darija is transcribed in Arabizi following a fixed convention: 7, 3, 9, kh, gh, ch, 2 for Darija-specific phonemes; French and English words are kept in their standard orthography (never transliterated to Arabizi); MSA/Fusha embedded in Darija is transliterated the same way as Darija. Full normalization specification: citation forthcoming (paper not yet published).

Citation

If you use this dataset, please cite:

• This corpus: citation forthcoming -- paper not yet published, check back here for an update
• DODa: see the DODa dataset card for its citation
• DVoice: Naira, A.M., Allak, A., Benelallam, I. "Dvoice: An open source dataset for Automatic Speech Recognition on African Languages and Dialects." Zenodo, 2022. https://doi.org/10.5281/zenodo.6342622
• Wiki: see the Moroccan-Darija-Wiki-Dataset card for its citation
• YouTube: cite the individual source channel/video linked in each row's sourceVideoUrl

License summary

This repo mixes several license/usage terms at the subset level (see table above): MIT (doda), CC BY 4.0 (dvoice), CC BY-SA 4.0 (wiki), and platform terms of service rather than a redistribution license for youtube (no audio is distributed for that subset; its text annotations are this project's own work). When using a subset, comply with its specific terms -- in particular, any redistribution of the wiki subset (or works derived from it) must retain attribution and remain under CC BY-SA 4.0 or a compatible share-alike license.