Emotion Recognition from Speech Using Hybrid Deep Learning Techniques
TL;DR
A Hybrid CNN-LSTM-CLSTM model that leverages Convolutional Neural Networks for spatial feature extraction, Long Short-Term Memory (LSTM) networks for modeling sequential dependencies, and Convolutional LSTM (CLSTM) for enhanced feature generalization is proposed.
Abstract
: Speech Emotion Recognition (SER) is essential for applications in healthcare, human-computer interaction, and affective computing. Traditional machine learning approaches have struggled with effectively capturing both spatial and temporal features from speech data. To address this challenge, we propose a Hybrid CNN-LSTM-CLSTM model that leverages Convolutional Neural Networks (CNNs) for spatial feature extraction, Long Short-Term Memory (LSTM) networks for modeling sequential dependencies, and Convolutional LSTM (CLSTM) for enhanced feature generalization. Our approach is trained and assessed using the RAVDESS, TESS, SAVEE, and CREMA-D benchmark datasets, guaranteeing robustness across different speakers, accents, and emotional variations. The hybrid architecture outperforms standalone CNN, LSTM, and CNN-LSTM models, demonstrating superior accuracy, adaptability, and robustness. Key challenges such as noise interference, speaker variability, and data imbalance are addressed using data augmentation and transfer learning techniques. Results from experiments confirm that the suggested model performs at the cutting edge of speech emotion recognition, indicating strong potential for translation into operational affective computing systems like virtual assistants, mental health monitoring, and emotion-aware AI systems.