Speech Emotion Recognition (SER) plays a decisive role in human-machine interactivity, affective and intelligent computing, and mental health diagnostics. Despite advancements, achieving a robust emotion prediction across diverse datasets remains challenging due to variability in speech patterns and limited model generalizability. This paper presents a comprehensive study on SER using hybrid deep learning architectures, evaluating machine learning as well as deep learning models—including Random Forest, SVM, and novel LSTM-CNN hybrids—on four benchmark datasets (RAVDESS, TESS, CREMA-D, SAVEE). We employ multi-feature extraction (MFCC, ZCR, Chroma STFT, Mel Spectrogram) and data expansion (noise injection, pitch shifting, time stretching) to enhance model robustness. Our experiments demonstrate that the proposed LSTM-CNN model achieves superior accuracy (71%) on the combined dataset, outperforming traditional methods and standalone CNN/LSTM models. Further, we analyze the impact of dataset fusion and feature selection on SER performance, providing insights for real-time deployment in emotion-sensitive applications.