Speech Emotion Recognition Using Mel Spectrogram HPCA and Variational Mode Decomposition

The rapid evolution of affective computing demands sophisticated methodologies to enhance the reliability and effectiveness of speech emotion recognition (SER). This study integrates harmonic-percussive component analysis (HPCA) with variational mode decomposition (VMD) to overcome various drawbacks for conventional speech emotion recognition (SER) methodologies that primarily rely on stand-alone feature extraction techniques. This implementation refines acoustic feature extraction and optimizes VMD decomposition to prevent information loss from mode duplication and mixing problems. We propose a feature map generator that channels the enhanced feature vectors into a convolutional neural network, specifically the VGG16 model, and the model is further enriched by incorporating diverse acoustic features including HP and log Mel spectro-grams into two-dimensional spaces to intensify data augmentation and enrich emotional feature representation. Extensive testing on Berlin EMO-DB and RAVDESS databases confirmed positive impacts for the proposed HP-VMD model performance, achieving robust classification accuracy of 96.67%. Thus, the proposed integrated approach to developing SER systems significantly enhances empathetic human computer interactions.