Anger recognition in Turkish speech using acoustic information

An emerging trend in human-computer interaction technology is to design spoken interfaces that facilitate more natural interaction between a user and a computer. Being able to detect the user's affective state during interaction is one of the key steps toward implementing such interfaces. In this study, anger recognition from Turkish speech using acoustic information is explored. The relative importance of acoustic feature categories in anger recognition is examined. Results show that logarithmic power of Mel-frequency bands, mel frequency cepstral coefficients and perceptual linear predictive coefficients are relatively more important than other acoustic categories in the context of anger recognition. Results also show that unweighted recall of 75.8% is obtained when correlation based feature selection method and Naive Bayes classifier are used.