DefinePK

Abstract

This study looks at how well By contrasting traditional machine learning (ML) approaches with deep learning (DL) techniques, excitation source information is used in the interpretation of emotional speech. The study extracts spectral and prosodic features from speech data, concentrating on excitation source characteristics as pitch contour, jitter, shimmer, and harmonic-to-noise ratio. We evaluate a number of DL designs, including Convolutional Neural Networks, Long Short-Term Memory networks, and hybrid models, as well as ML methods, including Support Vector Machines, Random Forest, and Gradient Boosting.—using standardized emotional voice datasets. With the hybrid CNN-LSTM model attaining the maximum accuracy of 92.7% in emotion classification tasks, experimental findings show that DL techniques outperform conventional ML approaches. Particularly for differentiating between comparable emotional states, the combination of excitation source characteristics greatly enhances classification performance. By developing a thorough framework for emotional speech analysis and offering a methodical comparison of modern categorization methods, this study adds to the area.