Recognition of Emotion Using Non-Linear Dynamics of Speech

Recognition of human's emotion from speech has become one of the most challenging and attractive fields of research in speech processing area. The present study aimed to detect valence of emotions, using Non-Linear Dynamic features (NLDs). NLDs are extracted from the Discrete Cosine Transform (DCT) of descriptor contours computed from Phase Space Reconstruction (PSR) of speech. These features are used to estimate emotion primitives in 3D continuous emotion space on the VAM database using Support Vector Regression (SVR) under the Leave-One-Out (LOO) testing technique. Feature selection is performed using Sequential Forward Selection (SFS). Activation, valence and dominance of emotions are estimated by the best correlation of 85.06%, 53.39% and 84.68%, respectively.