MULTI-FEATURE ANALYSIS AND ENSEMBLE LEARNING FOR IMPROVED EMOTION RECOGNITION IN WHISPERED SPEECH

Abstract

Abstract: This paper presents a novel method for emotion recognition from whispered speech, integrating advanced techniques in feature extraction, feature selection, and classification to enhance accuracy and robustness. The approach begins with extracting three types of features: wavelet features for multi-resolution analysis, prosodic features for pitch and intensity, and spectral features such as formants, Mel-Frequency Cepstral Coefficients (MFCCs), and Long-Term Average Spectrum (LTAS) to capture comprehensive emotional information. A two-step feature selection process, involving partial correlation analysis and Linear Discriminant Analysis (LDA), is employed to identify and retain the most informative features while reducing dimensionality. Classification is performed using an ensemble learning strategy that combines Support Vector Machine (SVM) and Decision Tree classifiers, with SVM distinguishing between neutral and emotional states and the Decision Tree further categorizing emotions. Simulation results using the GeWEC dataset demonstrate the effectiveness of the proposed method, achieving significant improvements in Unweighted Average Recall (UAR) across various configurations. This underscores the method’s capability to accurately recognize emotional states from whispered speech, offering valuable insights for practical applications in emotion recognition systems.