Using Nonverbal Information for Conversation Partners Detection by Wearable Devices

碩士 === 國立交通大學 === 網路工程研究所 === 105 === We all communicate with people in our daily lives. The conversation and nonverbal information are parts of communication. People usually have conversations with body languages to express themselves. Reference [5] uses acoustic sensing to infer personal contexts...

Full description

Bibliographic Details
Main Authors: Deeporn, Mungtavesinsuk, 王瀅惠
Other Authors: Tseng, Yu-Chee
Format: Others
Language:en_US
Published: 2016
Online Access:http://ndltd.ncl.edu.tw/handle/jmq36f
Description
Summary:碩士 === 國立交通大學 === 網路工程研究所 === 105 === We all communicate with people in our daily lives. The conversation and nonverbal information are parts of communication. People usually have conversations with body languages to express themselves. Reference [5] uses acoustic sensing to infer personal contexts and conversation partners. There could be some incorrect recognition or cluster, because of talking turn in conversation are continued. If it's in discussion, there will be some time blanks for thinking or analysis in the discussion, it's could be occur incorrect inference conversation group. In this work combine the nonverbal information of hand movement with speech to reduce incorrect recognition and cluster for conversation inference. We use some novel inference methods in [5] to classify the conversational relationships among co-located users. We cluster conversation group via combine speaking turns with nonverbal information of human body movement and calculate by conversation clustering algorithm. Finally, we compare the accuracy with [5]. There for, in this paper, we propose several methods to increase the accuracy and efficiency of group conversation inference. The keys of our methods are adding nonverbal information to combine with speaking turns to infer the conversation group. The experiment collect nonverbal information in conversation data from wearable device. The number of people in conversation are 2 to 9 people. We have two experiment environments, the first is discussion, and the second is Q&A. The result increases the accuracy of conversation inference 2% to 6%, which better than using only the sound sensing to infer group conversation.