Abstract
This report describes a study on an estimation method of degree of speaker's emotion by using acoustic and linguistic features expressed in their anger utterances during a natural dialogue. We set two types of pseudo dialogues, the human-computer and the human-human, to induce anger utterances from 10 speakers. To make an emotional speech corpus with degree of emotion, a 5-scale subjective evaluation was conducted to grade each utterance on its emotional degree. The emotional speech corpus was examined to find acoustic and linguistic features which estimate the emotional degree of each utterance. Decision trees were adopted as classifiers for our estimation examination to find optimal sets of the acoustic and linguistic features for an anger degree estimation. As a result, we find specific tendencies of the tree acousitc features in strong anger utterances, the linguistic parameters' potential to estimate degree of anger emotion, and the capability of decision tree to estimate utterances with two kinds of acoustic features as the strong anger.
