FatimahEmadEldin/Moroccan-Arabic-Multimodal-Emotion-Recognition

MDER-MA — Moroccan Arabic Multimodal Emotion Recognition (TTS-aligned repackaging)

A repackaging of the MDER-MA dataset that pairs every audio clip with its Arabic (Moroccan dialect / Darija) transcript and ships speaker-disjoint train/validation/test splits.

Original dataset: Ouali, S. & El Garouani, S. (2025). MDER-MA: A multimodal dataset for emotion recognition in low-resource Moroccan Arabic language. Data in Brief. DOI: 10.1016/j.dib.2025.112005. Mendeley: https://data.mendeley.com/datasets/yzsw3ff6rn/1.

What's in this repackaging

• 1216 audio↔transcript pairs drawn from MDER-MA's audio + text modalities.
• 66 unique speaker sessions (gender_age_session_id triples).
• 4 emotion classes: Angry, Happy, Neutral, Sad.
• Audio: 48 kHz, stereo .wav. Median duration ≈5 s.
• Spectrogram and Mel-spectrogram images from the original dataset are not included to keep repo size manageable; users who need them can generate from audio or download from Mendeley.

Schema

| Column | Description | |---| | file_name | relative path to the audio (e.g. data/Angry_female_15_1_6.wav) | | transcription | Arabic transcript (Moroccan dialect) | | emotion | Angry / Happy / Neutral / Sad | | gender | female / male | | age | speaker age (int) | | session_id | recording session index | | utterance_id | utterance index within session | | speaker_key | gender_age_session_id (used for split disjointness) | | duration_s · sampling_rate · channels | audio properties |

Splits

Splits are speaker-disjoint — speaker_key never overlaps between train / validation / test, so reported metrics measure cross-speaker generalization.

| Split | Speaker keys | Clips | |---| | train | 52 | 1009 | | validation | 6 | 49 | | test | 8 | 158 |

Excluded from this repackaging

• 83 audio clips lacked matching transcripts → dropped.
• 20 transcript files lacked matching audio → dropped.
• All spectrogram (.jpg) and Mel-spectrogram (.jpg) modalities → not included (regenerate from audio if needed).

Usage

from datasets import load_dataset
ds = load_dataset("FatimahEmadEldin/Moroccan-Arabic-Multimodal-Emotion-Recognition")
print(ds["train"][0])
# → { "audio": { "array": [...], "sampling_rate": ... }, "transcription": "...", "emotion": "Angry", ... }

Honest scope notes

• Moroccan Darija, not MSA. Models trained on Modern Standard Arabic will need adaptation; conversely, models trained here will not transfer cleanly to MSA without further fine-tuning.
• Demographic skew. Age range and gender balance are inherited from the original dataset — verify they're suitable for your use case before deploying.
• Speaker imbalance per emotion is possible: see metadata.csv for exact distributions.
• Annotation reliability. Original annotation by 5 native Moroccan speakers (per the source paper).

License

The original Mendeley record does not state an explicit license that I could verify. This repackaging defers to whatever the original authors specify; please consult https://data.mendeley.com/datasets/yzsw3ff6rn/1 before any commercial use.

Citation

@article{ouali2025mderma,
  title   = {MDER-MA: A multimodal dataset for emotion recognition in low-resource Moroccan Arabic language},
  author  = {Ouali, Soufiyan and El Garouani, Said},
  journal = {Data in Brief},
  year    = {2025},
  doi     = {10.1016/j.dib.2025.112005},
  url     = {https://www.sciencedirect.com/science/article/pii/S2352340925007292}
}