Speech emotion recognition (SER) is a key challenge in affective computing, where subtle emotional cues are often embedded not in the linguistic content of speech but in the voice-related acoustic features. This study proposes a machine learning approach that leverages statistical descriptors of Mel-Frequency Cepstral Coefficients (MFCCs) to capture the central tendencies of voice signals for multiclass emotion classification. Raw voice from the Toronto Emotional Speech Set (TESS) was processed into nine statistical features, of which six were retained after correlation-based filtering to reduce redundancy and improve generalization. Several classifiers were evaluated, with Support Vector Machine (SVM) achieving the best performance: 84% accuracy, 83% macro-recall, and 83% macro-F1. The improvements after hyperparameter tuning were statistically significant (McNemar’s test, p = 1.606e-20), underscoring the importance of systematic optimization. A comparative analysis revealed that correlation-based feature selection outperformed PCA and LDA in preserving the discriminative power of SVM. Compared with related works that employ deep learning or multi-dataset setups, the proposed framework offers competitive performance while maintaining greater interpretability and computational efficiency. These findings validate the hypothesis that compact, voice-centered statistical features, when optimized, form a reliable basis for robust and efficient emotion recognition systems.
Gunawan et al. (Sun,) studied this question.