FatimahEmadEldin/Yemeni-Speech-Emotion-Dataset

YSED — Yemeni Speech Emotion Dataset (audio-classification repackaging)

A clean repackaging of YSED with a metadata.csv and stratified train/validation/test splits, for emotion classification on Yemeni Arabic.

Original dataset: Derhem, S., AL-Mekhlafi, E., AL-Majmar, N. A., & AL-Makhlafi, M. (2025). YSED: Yemeni Speech Emotion Dataset. Data in Brief. DOI: 10.1016/j.dib.2025.112233. Zenodo: https://zenodo.org/records/15227219.

What's in here

• 1432 audio clips across 5 emotion classes: angry, fearful, happy, neutral, sad.
• 48 kHz stereo .wav, ~2.7 s median duration.
• Per the original paper: 71 Yemeni volunteers (37 M, 34 F), ages 15–45, validated by 6 judges with Fleiss' Kappa = 0.9.

Schema

Column	Description
file_name	relative path to audio (e.g. data/angry__001.wav)
emotion	one of angry / fearful / happy / neutral / sad
duration_s · sampling_rate · channels	audio properties
source_filename	original filename in the YSED Zenodo download

Splits

Stratified by emotion (each split has the same emotion proportions). 80/10/10 with seed 42.

Split	Clips
train	1145
validation	143
test	144

⚠ This is not a TTS dataset

The original YSED Zenodo download does not include per-clip transcripts. The source paper mentions 25 fixed sentences (5 per emotion) but without a per-file mapping we cannot safely assign transcripts. Use this dataset for emotion classification, not for TTS or ASR.

If you can derive a sentence-id-per-file mapping from the original release notes or by contacting the authors, you can extend the metadata with a transcription column and re-tag the dataset for TTS.

Honest scope notes

• Yemeni dialect, not MSA. Models trained here will not transfer cleanly to other Arabic dialects without adaptation.
• Splits are not speaker-disjoint — without parsing speaker info from filenames, this implementation uses stratified random splits by emotion. If you can recover speaker IDs from filenames, redo the splits with speaker disjointness for honest evaluation.
• Recording conditions vary — speakers were volunteers using the simulated/induced approach; real-world inference performance may differ.

License

The original YSED release does not state an explicit license that I could verify. Please consult https://zenodo.org/records/15227219 before any commercial use.

Citation

@article{derhem2025ysed,
  title   = {YSED: Yemeni Speech Emotion Dataset},
  author  = {Derhem, Somia and AL-Mekhlafi, Eiad and AL-Majmar, Nashwan Ahmed and AL-Makhlafi, Moeen},
  journal = {Data in Brief},
  year    = {2025},
  doi     = {10.1016/j.dib.2025.112233},
  url     = {https://www.sciencedirect.com/science/article/pii/S2352340925009540}
}