Emotional speech discrimination using sub-segmental acoustic features

摘要

The ability to express emotions is a natural trait of humans. Machines or robots do not have that ability, so far. The synthesized speech signals, that machines use currently, lack naturalness, primarily because those do not convey any emotions and hence sound flat or artificial to human ears. Automatic emotion recognition from the speech signal has also been a big challenge to researchers in speech signal processing domain. In this paper, changes in the discriminating features are analyzed for 6 different emotions: anger, fear, happiness, surprise, sadness and neutral. The sub-segmental features F0, Strength of Excitation and Signal Energy are derived directly from the emotional speech signal, for discriminating amongst these emotions. Signal processing methods autocorrelation of linear prediction (LP) residual, zero-frequency filtering and signal energy are used for deriving these features. The analysis is carried out using two emotion databases, namely German Emotion Database and Telugu Emotion Database. Emotional speech data, for a total of 10 speakers (5 male and 5 female speakers) is examined from each database. Performance evaluation results of using these features for discriminating amongst the 6 emotions are encouraging. These discriminating features and this study should be helpful further towards automatic detection and classification of these different emotions.