Análise de variações acústicas não estacionárias e seu efeito na detecção de múltiplas emoções e condições de estresse.

Abstract

The goal of this work is to study the effects of non-stationary acoustic variations caused by emotional states and stress conditions. In the literature, there is still no pure acoustic attribute for emotion and stress recognition. By using the index of non-stationarity (INS), it is observed that different affective states have different degrees of non-stationarity. As for the detection of such variations, it is employed the empirical mode decomposition (EMD), which is a nonlinear technique that is suitable for non-stationary signals. Thus, the main contribution of this work is the proposal of the HHHC vector (Hilbert-Huang-Hurst Coefficients) as a new non- linear acoustic feature for the multistyle classification of emotional states and stress conditions. The HHHC is a vocal source feature that is based on adaptive decomposition (EMD, which emphasizes affective acoustic variations) and Hurst coefficients estimation (which are related to the glottal source excitation) in each decomposition mode. Another contribution is the use of INS as additional information to the HHHC vector (HHHC+INS). In order to analyze the robustness of the proposed acoustic feature in different languages and speaking contexts, it is considered five databases. Four of them in the context of emotions and one in the context of stress conditions. As baseline acoustic features to comparing with HHHC, it is used the vector of Hurst coefficients (pH), the Mel-Frequency Cepstral Coefficients (MFCC) and TEO (Teager- Energy-Operator)-based feature. Another important contribution of this Thesis is the proposal of α-integrated Gaussian Mixture Models (α-GMM) for the affective states representation and classification. Its performance is compared to competing classifiers: GMM, Hidden Markov Models (HMM) and Support Vector Machines (SVM). Results demonstrate that the proposed HHHC acoustic feature leads to significant classification improvement when compared to the baseline acoustic features. Also, the results show that α-GMM outperforms the competing classification methods in all acoustic databases scenarios.