Emotion recognition from speech signals using new harmony features

In this paper we propose a new set of harmony features for automatic emotion recognition from speech signals. They are based on the psychoacoustic harmony perception known from music theory. Starting from the estimated pitch contour of an utterance, we calculate the circular autocorrelation of the pitch histogram on the logarithmic semitone scale. It measures the occurrence of different two-pitch intervals which cause a consonant or dissonant impression. Experiments of emotion recognition using these harmony parameters in addition to state of the art features show an improved recognition performance.