| Abstract: |
In recent years, we have conducted an active research program aimed at the capture, transcription, tracking,
description, review, access, recall and summarization of human-human interaction in meetings. The task
extends previous research toward truly natural, unprepared and unconstrained human interaction. The problem
is inherently multimodal and involves capturing all available signals. Based on these signals, it involves robust
processing, fusion and understanding of the full breadth of human communicative hints and signals, without
prior preparation, segmentation or artificial restrictions in recording style.
The processing problems include large vocabulary conversational speech recognition, microphone
independence, cross talk, sound source localization, recognition of emotion (from speech and facial
expression), identification of participants from speech and face), tracking of topics, summarization from
speech, visual eye-gaze, pose and focus of attention tracking. In this talk I will describe the problem, our
current research efforts, the databases and evaluation methodologies currently used, and directions for future
research and resource requirements. |
