MDER-MA: A multimodal dataset for emotion recognition in low-resource Moroccan Arabic language

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Speech Recognition

Background:

• Emotion recognition is crucial for humanizing virtual agents and enhancing user experience.
• Existing research is limited in low-resource languages like Moroccan Arabic due to cultural and linguistic variations.
• A high-quality, realistic dataset is essential for developing effective emotion recognition systems.

Purpose of the Study:

• To introduce MDER-MA, a comprehensive multimodal dataset for emotion recognition in Moroccan Arabic.
• To address the gap in emotion recognition research for low-resource languages.
• To foster the development of Arabic language technologies, particularly regional dialects.

Main Methods:

• Developed MDER-MA dataset with 5288 samples across four emotions (Happy, Sad, Angry, Neutral).
• Collected data in four modalities: audio, text, spectrogram, and Mel-spectrogram images.
• Ensured dataset representativeness by collecting samples from various Moroccan regions and using native speakers for annotation.

Main Results:

• MDER-MA contains 1322 samples per modality, totaling 5288 data items.
• The dataset supports applications beyond emotion recognition, including audio transcription and demographic identification.
• Annotation by five native Moroccan speakers ensures high linguistic reliability.

Conclusions:

• MDER-MA bridges the gap in emotion recognition for low-resource languages like Moroccan Arabic.
• The dataset facilitates the creation of more natural and empathetic virtual agents.
• This work promotes advancements in Arabic natural language processing and emotion-aware AI.