Croatian Emotional Speech Analyses on a Basis of Acoustic and Linguistic Features

Acoustic and linguistic speech features are used for emotional state estimation of utterances collected within the Croatian emotional speech corpus. Analyses are performed for the classification of 5 discrete emotions, i.e. happiness, sadness, fear, anger and neutral state, as well as for the estimation of two emotional dimensions: valence and arousal. Acoustic and linguistic cues of emotional speech are analyzed separately, and are also combined in two types of fusion: a feature level fusion and a decision level fusion. The Random Forest method is used for all analyses, with the combination of Info Gain feature selection method for classification tasks and Univariate Linear Regression method for regression tasks. The main hypothesis is confirmed, i.e. an increase of classification accuracy is achieved in the cases of fusion analyses (compared with separate acoustic or linguistic feature sets usages), as well as a decrease of root mean squared error when estimating emotional dimensions. Most of other hypothesis are also confirmed, which suggest that acoustic and linguistic cues of Croatian language are showing similar behavior as other languages in the context of emotional impact on speech.