FatimahEmadEldin/Arabic-Emotional-Audio-Dataset-Baved

BAVED — Basic Arabic Vocal Emotions Dataset (TTS-ready repackaging)

A re-packaged, transcript-aligned version of the Basic Arabic Vocal Emotions Dataset (BAVED) with explicit Arabic transcripts, English glosses, speaker metadata, and speaker-disjoint train/validation/test splits.

Original dataset: Aouf Yacine, Basic Arabic Vocal Emotions Dataset (BAVED), GitHub: https://github.com/40uf411/Basic-Arabic-Vocal-Emotions-Dataset. This repackaging adds metadata; all audio is unchanged.

What's in here

• 1935 unique recordings (deduplicated; the source distribution had each file twice under remake/remake/).
• 60 speakers (44 male, 16 female), ages 5–41.
• 7 Arabic words ! 3 emotion intensity levels = 21 unique text-emotion pairs.
• 16 kHz mono WAV.

Word index → Arabic transcript

word_id	Arabic	English
0	أعءبنـى	like
1	لم فعءبنـى	unlike
2	هذا	this
3	الففلم	the film
4	رائع	wonderful
5	مقبول	acceptable
6	سفؤ	bad

Emotion level → label

emotion_id	Label	Description
0	low	tired / subdued
1	neutral	standard daily speech
2	high	strong positive or negative emotion

Schema

file_name (str, relative path to WAV) ! transcription (Arabic) ! transcription_en (gloss) ! word_id ! emotion_id ! emotion_label ! speaker_id ! gender (m/f) ! age ! record_id ! duration_s ! sampling_rate.

Splits

Splits are speaker-disjoint (a speaker appears in exactly one split) so reported metrics measure generalization across speakers, not memorization.

Split	Speakers	Recordings
train	48	1681
validation	6	213
test	6	41

Usage

from datasets import load_dataset
ds = load_dataset("YOUR_USERNAME/baved-tts-ready")
print(ds["train"][0])
# → { "audio": {...}, "transcription": "أعءبنـى", "emotion_label": "neutral", ... }

Honest scope notes (read this before training a TTS model)

• Vocabulary is tiny. Only 7 words. This is not a general-purpose Arabic TTS corpus. It is useful for:
  • Speech-emotion recognition (the original intent)
  • Emotional / paralinguistic TTS research on a closed vocabulary
  • Keyword spotting with affective conditioning
• Speaker imbalance. ≈3! more male than female speakers; most speakers are 18–23. Don't train demographic models on this.
• Recording conditions vary — samples were normalized, but they were originally recorded across different setups.
• Original recordings were converted to 16 kHz mono from a mix of source rates.

License

The original BAVED repository does not state an explicit license. This repackaging defers to whatever the original authors specify; please contact them at https://github.com/40uf411/Basic-Arabic-Vocal-Emotions-Dataset before any commercial use. The original authors themselves note that "commercial use won't probably be a good idea" given the dataset's size and demographic skew.

Citation

If you use BAVED, cite the original Wav2Vec2/HuBERT paper that introduced its modern usage in deep learning:

@article{mohamed2021arabic,
  title   = {Arabic Speech Emotion Recognition Employing Wav2vec2.0 and HuBERT Based on BAVED Dataset},
  author  = {Mohamed, Omar and Aly, Salah A.},
  journal = {arXiv preprint arXiv:2110.04425},
  year    = {2021},
  url     = {https://arxiv.org/abs/2110.04425}
}

@misc{baved2019,
  title        = {{Basic Arabic Vocal Emotions Dataset (BAVED)}},
  author       = {Yacine, Aouf},
  year         = {2019},
  howpublished = {\url{https://github.com/40uf411/Basic-Arabic-Vocal-Emotions-Dataset}}
}

For the Arabic word translations used in this repackaging, the reference is:

@article{aljuhani2025baved,
  title   = {Effective Data Augmentation Techniques for Arabic Speech Emotion Recognition Using Convolutional Neural Networks},
  author  = {Aljuhani, Reem H. and others},
  journal = {Applied Sciences},
  volume  = {15},
  number  = {4},
  pages   = {2114},
  year    = {2025},
  doi     = {10.3390/app15042114}
}