Enhancing Speech Emotion Recognition in Urdu Using Bi-GRU Networks: An In-Depth Analysis of Acoustic Features and Model Interpretability

This paper introduces an innovative approach to Speech Emotion Recognition (SER) for Pakistani Urdu, employing the URDU dataset, a specialized resource in this domain, and utilizing Bidirectional Gated Recurrent Unit (Bi-GRU) networks. Our study concentrates on the meticulous extraction and thorough analysis of crucial acoustic features from this dataset, including Mel Frequency Cepstral Coefficients (MFCCs), Gam-matone Frequency Cepstral Coefficients (GFCCs), and the ex-tended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS). Our methodology integrates sophisticated feature engineering, normalization, and deploying advanced Bi-GRU architectures for deep learning models. A significant focus of our work is on model interpretability, achieved through SHAP (SHapley Additive exPlanations) analysis, which provides valuable insights into the contributions of different features to emotion recognition. The Bi-GRU model developed in this study demonstrates exceptional proficiency, achieving a peak validation accuracy of approximately 91 % in identifying emotions such as sadness, neutrality, happiness, and anger. This research underscores the effectiveness of Bi-GRU networks in SER, particularly in Urdu language processing, using the URDU dataset. Additionally, our work includes comprehensive validation through visualization techniques, error analysis, confusion matrices, and Receiver Operating Characteristic (ROC) curves, reinforcing the model's robustness and reliability. This paper significantly advances the field of SER, especially for low-resource languages like Urdu, offering a foundational framework for future research and practical applications in emotionally intelligent human-computer interaction.