Speech emotion recognition (SER) struggles with dataset diversity, model generalizability, and efficiency. Using a single dataset risks bias and limits applicability. This study combines four datasets (RAVDESS, TESS, SAVEE, CREMA-D) into a unified dataset to improve analysis and generalizability. An interpretable machine learning framework was developed, using data augmentation (e.g., noise injection, time stretching) and acoustic features (e.g., MFCC, ZCR, Chroma) to detect six emotions: anger, happiness, fear, sadness, neutral, disgust. Algorithms like SVM, LightGBM, Random Forests, XGBoost, and GBM were tested. SVM achieved 92% accuracy, followed by LightGBM at 91%. The study shows traditional machine learning can excel in SER, with dataset integration enhancing model performance and reliability.