Analysis of Speech Parameters as Indicators of Engagement in Conversation

Abstract

In human conversations, the words being said may not always contain the whole of the information intended to be conveyed; instead, such information may be conveyed through the way those words are being said. Thus understanding the information beyond the words is vital in understanding the behaviours of the speaker in a conversation. The acoustic-prosodic features in speech are excellent sources of information which can indicate the internal states of a speaker. Humans have the capability of doing this processing of extra-linguistic information to understand their conversational partner - computers, on the other hand, lack this ability. The analysis of speech signal can reveal a lot of the information which may not be captured by analysing the linguistic content alone. In this study, the information encapsulated by the features of the speech signal is analysed to model conversational engagement of the participants. Casual conversations are analysed as part of this study to identify features of the speech signal, which can be used as cues for detecting engagement in conversations. Voice quality features (including glottal parameters) are mainly analysed in this study to model engagement in conversation. The proposed methods attempt to understand the conversational engagement of participants in casual conversations.