mansoorSaleh/AYDID-public

AYDID: Arabic Yemeni Dialect Identification Dataset (public release)

This repository contains a public sample and the held-out test set of AYDID, the first dedicated speech corpus for Yemeni Arabic at the sub-dialectal level, supporting both automatic speech recognition (ASR) and dialect identification (DID).

Note on scope. This release contains a representative sample plus the benchmark test set. It is intended for evaluating models against the published baselines, not for reproducing the fine-tuning results. The full training corpus is available from the authors under a data-sharing agreement (see Access to the full corpus below).

Dataset summary

AYDID covers seven Yemeni Arabic dialect classes — six regional sub-dialects and Standard Yemeni — with orthographic transcriptions, speaker demographic metadata, and environmental condition tags. The full corpus comprises 20.25 hours of speech from 350 native speakers; this repository publishes a subset of that material.

Property	Value
Dialect classes	7 (Adeni, Badawi, Hadrami, Sana'ani, Ta'izzi, Tihami, Standard Yemeni)
Audio format	16 kHz, mono, 16-bit WAV
Transcription	Orthographic, Arabic script, native-speaker transcribed
Full corpus	20.25 h · 350 speakers · 15,197 utterances
Speaker balance	50 speakers per dialect class
Inter-annotator agreement	Cohen's κ = 0.91 (speaker level)
This release	Held-out test set (1,731 utterances) + training/validation sample
License	CC BY-NC 4.0

Dialect classes

Label	Dialect	Region
YEM_AD	Adeni	Coastal
YEM_BA	Badawi	Desert
YEM_HA	Hadrami	Eastern
YEM_SA	Sana'ani	Highland
YEM_ST	Standard Yemeni	—
YEM_TA	Ta'izzi	Highland
YEM_TI	Tihami	Coastal

Data fields

Each example contains:

• audio: the speech recording (16 kHz mono WAV).
• transcription: orthographic transcription in Arabic script.
• transcription_norm: normalized transcription used for WER+/CER+ evaluation.
• dialect: one of the seven class labels above.
• speaker_id: anonymized speaker identifier.
• gender: speaker gender (M/F).
• age: speaker age in years.
• age_group: young adult (18–35), middle-aged (36–55), or elderly (56+).
• split: train / validation / test.

Splits

The benchmark uses a fixed 80/10/10 train/validation/test split with balanced class distribution. The split is speaker-disjoint: partitioning is performed at the speaker level, so all recordings from a given speaker belong to a single split and no speaker appears in more than one partition. This ensures models learn dialectal rather than speaker-specific characteristics.

Please use the provided splits as released. Re-shuffling at the utterance level breaks the speaker-disjoint design and will not reproduce the reported numbers.

Split	Status in this release
train	sample only (full set on request)
validation	sample only (full set on request)
test	complete (1,731 utterances, ~247–249 per class)

Usage

from datasets import load_dataset

ds = load_dataset("mansoorSaleh/AYDID-public")
print(ds["test"][0]["transcription"], ds["test"][0]["dialect"])

Benchmarks

Best results reported in the associated paper. WER+ is the normalized word error rate (lower is better); weighted F1 is reported for DID (higher is better).

Task	Model	Metric	Score
ASR (zero-shot, best)	Whisper Large-V3	WER+	43.92%
ASR (fine-tuned, best)	Whisper Large-V2 (fine-tuned on AYDID)	WER+	19.14%
DID (zero-shot)	MMS-300m Arabic dialect identifier	weighted F1	11.45%
DID (fine-tuned)	MMS-300m (fine-tuned on AYDID)	weighted F1	80.44%

Fine-tuning the DID model on AYDID improves the weighted F1 by 68.99 percentage points over the zero-shot baseline.

Associated models: [HF model URL(s) — add once released]. Evaluation code: [GitHub URL — add if available].

Access to the full corpus

The complete training corpus is available from the authors under a data-sharing agreement, subject to the licensing constraints of the broadcast-media source material. Contact: [corresponding author email — e.g. 6121000101@tju.edu.cn].

Licensing and ethics

Released under CC BY-NC 4.0 for non-commercial research use. Field recordings were collected from consenting native speakers who self-reported their dialect of origin; the released data contains no personally identifiable information beyond coarse demographic metadata (age group, gender, dialect region). Media-sourced segments are redistributed within the terms of their original sources for non-commercial research use only; users are responsible for compliance with applicable terms in their jurisdiction.

Citation

@article{bamahel_aydid,
  title   = {AYDID: The First Sub-Dialectal Speech Corpus for Yemeni Arabic,
             with ASR and Dialect-Identification Benchmarks},
  author  = {Ba Mahel, Mansoor S. M. and Wei, Jianguo and Awn, Norah Saeed and Bamahel, Abdulaziz S.},
  journal = {Speech Communication},
  year    = {[year — fill in on acceptance]},
  note    = {Dataset: https://huggingface.co/datasets/mansoorSaleh/AYDID-public}
}