Speaker Role Recognition to help Spontaneous Conversational Speech Detection

摘要

In the audio indexing context, we present our recent co: tributions to the field of speaker role recognition, especial applied to conversational speech. We assume that there e: ist clues about roles like Anchor, Journalists or Others temporal, acoustic and prosodic features extracted from tl results of speaker segmentation and from audio files. In th paper, investigations are done on the EPAC corpus, main containing conversational documents. First, an automat clustering approach is used to validate the proposed fe tures and the role definitions. In a second study we propoi a hierarchical supervised classification system. The use dimensionality reduction methods as well as feature sele tion are investigated. This system correctly classifies 92% speaker roles.