Bangla Speech Emotion Recognition Using Deep Learning-Based Ensemble Learning and Feature Fusion.

Area of Science:

• Speech processing
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Bangla speech emotion recognition faces challenges in accuracy, speaker dependency, and generalization.
• Existing methods using traditional or basic deep learning models lack robustness in varied conditions.

Purpose of the Study:

• To propose a novel multi-stream deep learning feature fusion approach for Bangla speech emotion recognition.
• To address limitations of existing methods by enhancing accuracy, robustness, and generalization.

Main Methods:

• Data augmentation techniques applied to training datasets.
• Extraction of handcrafted features (ZCR, MFCCs, etc.) and deep learning features.
• Multi-stream deep learning architecture with 1D CNN, CNN-LSTM, and CNN-Bi-LSTM streams.
• Ensemble learning with soft voting for final prediction.

Main Results:

• Achieved high accuracies: 92.90% (SUBESCO), 85.20% (BanglaSER), 90.63% (merged), 67.71% (RAVDESS), 69.25% (EMODB).
• Demonstrated improved robustness and generalization compared to existing methods.
• Effectively combined handcrafted and deep learning features through ensemble learning.

Conclusions:

• The proposed multi-stream deep learning feature fusion approach significantly enhances Bangla speech emotion recognition.
• Combining diverse features and ensemble learning provides a more comprehensive and robust solution.
• The method offers a promising advancement for emotion recognition in human-computer interaction systems.