AYDID: A Sub-Dialectal Yemeni Arabic Speech Corpus

First speech resource to label Yemeni Arabic at the sub-dialectal level: 17,500 utterances (22.53 h), 350 speakers, 7 classes — six regional varieties (Adeni, Badawi, Hadrami, Sana'ani, Ta'izzi, Tihami) plus an MSA-proximal Standard Yemeni control class. Balanced at 50 speakers / 2,500 utterances per class.

Reproducibility tiers

To respect the copyright of the broadcast source material, raw audio is not publicly redistributed. The release is layered:

1. Public (this repo): frozen embeddings (MMS-300m, XLS-R, WavLM), transcripts, full metadata, evaluation splits, and scripts — reproduce all frozen-probe DID results and the confound baselines directly.
2. Public checkpoints: fine-tuned DID (MMS-300m, XLS-R) and ASR (Whisper-medium) models.
3. Gated audio (companion repo): segmented WAV audio under a non-commercial research data-use agreement: https://huggingface.co/datasets/mansoorSaleh/AYDID-audio

Contents

• metadata.csv — dialect, speaker_id, source programme, content type, gender, age, transcript for every utterance; row order matches the embedding files.
• embeddings.npy + index.json — mean-pooled frozen MMS-300m embeddings.
• emb_xlsr.npy + index_xlsr.json — mean-pooled frozen XLS-R embeddings (the paper's best DID encoder).
• emb_wavlm.npy + index_wavlm.json — mean-pooled frozen WavLM embeddings.
• channel_features.npy — 17 low-level acoustic features (channel baseline).
• splits/ — speaker_disjoint.json, programme_disjoint.json (4 folds), asr_test.json (1,742-utterance ASR test).
• DATASHEET.md — per-dialect statistics, split construction (with seeds), ASR fine-tuning / decoding / normalization config, and scoring details.
• scripts/ — run_did_probes.py, exp4_did_seeds.py (5-draw CIs), exp3_channel_control.py, path2_confound_control.py, exp1_asr_per_dialect.py, score_asr.py.
• checkpoints_manifest.json — links to the fine-tuned models.

Reproduce the headline DID results

python scripts/exp4_did_seeds.py
# XLS-R  probe wF1 = 82.84 +/- 0.78   (6-regional macro 79.94)
# MMS    probe wF1 = 74.68 +/- 0.41
# WavLM  probe wF1 = 72.12 +/- 0.51

Models

• ASR (Whisper-medium, fine-tuned): https://huggingface.co/mansoorSaleh/whisper-yemeni
• DID (MMS-300m): https://huggingface.co/mansoorSaleh/mms-300m-aydid-did
• DID (XLS-R): https://huggingface.co/mansoorSaleh/xlsr-aydid-did
• Code: https://github.com/MANSOOR-SALEH/AYDID

Citation

@inproceedings{aydid2027,
  title     = {AYDID: A Sub-Dialectal Yemeni Arabic Speech Corpus for Dialect
               Identification and Automatic Speech Recognition},
  author    = {Ba Mahel, Mansoor S. M. and Wei, Jianguo and Yue, Xianghu and
               Awn, Norah Saeed and Bamahel, Abdulaziz S.},
  booktitle = {Proc. IEEE ICASSP},
  year      = {2027}
}

Licensed CC BY-NC-SA 4.0 — non-commercial research use, with attribution.