Published Paper Details:

Human speech is a fundamental mode of communication, rich not only in linguistic data but also in emotional cues. Speech Emotion Recognition (SER) aims to automatically decode emotional states by analyzing vocal characteristics such as spectral shape, rhythm, intensity, and pitch. With the capabilities of modern Artificial Intelligence (AI), deep learning approaches provide superior methods for extracting complex emotional patterns compared to earlier machine learning techniques. This study introduces a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture designed for robust emotion classification. The proposed methodology involves extensive preprocessing and the extraction of multiple features, including Mel-Frequency Cepstral Coefficients (MFCC), Chroma, and Zero Crossing Rate (ZCR). Trained on a fused dataset including benchmark resources like RAVDESS, TESS, and CREMA-D , the CNN-LSTM model demonstrated an overall accuracy of 90%+. This result confirms the model's strong generalization and significant performance improvement over classical methods, offer ing a valuable contribution to emotion-aware AI systems for applications in human-computer interaction (HCI) and mental health analysis.