Detecting emotion from voice using selective Bayesian pairwise classifiers

This paper describes a method for detecting a dialogist's emotion from his or her voice. The method is based on pairwise classification by probability from the selective pairwise classifiers. In this research, we focus on the elements of emotion included in a dialogist's voice. Thus, as training datasets for learning the pairwise classifiers, we extract acoustic features from emotionally expressive voice samples spoken by unspecified actors and actresses in films, and TV dramas. The acoustic features adopt duration per single mora, fundamental frequency, energy, and formant. All features except duration per single mora have statistics extracted standard deviation, mean, maximum, minimum, median, timezone of maximum, and timezone of minimum. Pairwise classification classifies a multi-class problem by using a series of binary classifiers. Pairwise classifiers used tree augmented naive bayes, which constructs tree structure among the attributes in naive bayes, by selected subset features. The subset features are selected on every pair of emotions by using naive bayes. This paper reports the accuracy rates of emotion detection by using our method. In experimental results from our voice samples, the emotion classification rates improved.