Feature selection for emotion recognition in speech: a comparative study of filter and wrapper methods

Feature selection is essential for enhancing the performance and reducing the complexity of speech emotion recognition models. This article evaluates various feature selection methods, including correlation-based (CB), mutual information (MI), and recursive feature elimination (RFE), against baseline approaches using three different feature sets: (1) all available features (Mel-frequency cepstral coefficients (MFCC), root mean square energy (RMS), zero crossing rate (ZCR), chromagram, spectral centroid frequency (SCF), Tonnetz, Mel spectrogram, and spectral bandwidth), totaling 170 features; (2) a five-feature subset (MFCC, RMS, ZCR, Chromagram, and Mel spectrogram), totaling 163 features; and (3) a six-feature subset (MFCC, RMS, ZCR, SCF, Tonnetz, and Mel spectrogram), totaling 157 features. Methods are compared based on precision, recall, F1-score, accuracy, and the number of features selected. Results show that using all features yields an accuracy of 61.42%, but often includes irrelevant data. MI with 120 features achieves the highest performance, with precision, recall, F1-score, and accuracy at 65%, 65%, 65%, and 64.71%, respectively. CB methods with moderate thresholds also perform well, balancing simplicity and accuracy. RFE methods improve consistently with more features, stabilizing around 120 features.