SPARTA-benchmark

SPARTA stands for Speech Profiling for ARabic TAlk. The purpose of this project is to identify the speaker’s gender, dialect, and emotion when speaking from a given utterance.

This research profiles the speaker based on their voice. From a given speech, our model predicts gender, emotion, and dialect in real-time.

• Emotion [Sad, Happy, Angry, Surprise, Questioning, and Neutral].
• Gender [Male and Female] in the Arabic language (as many other languages).
• Dialect [Egypt, MSA, Levant, Gulf, North Africa].

Datasets

1. Qatar Computing Research Institute QCRI
2. King Saud University Emotions KSUEmotions
3. Arabic Natural Audio Dataset ANAD
4. Spoken Arabic Regional Archive SARA
5. King Saud University Arabic Speech Database KSU
6. Multi Dialect Arabic Speech MDAS

Data Preprocessing

To reproduce the dataset collected please follow the instructions:

1. Download Datasets from the links above.
2. Run collect_DATASETNAME.py for all datasets.
3. Make the Train/Dev/Test split by running the preprocessing_scripts/split.py
4. Extract features by running extract_FEATURENAME.py for all features