Politician speaker turn detection in TV Broadcast News shows is addressed in this paper. Politician speech model combines acoustical and lexical cues as well as contextual information, and does not use any specific politician model (person-independent). Politician speaker turn detection is coupled with an automatic role labeling step, which determines the contextual information and the set on which politician detector is applied. On a set of 101 TV broadcast news shows, experiments show that the politician speaker turns, which represent only 3% of the whole set of speaker turns in the corpus, are detected with a maximal F-measure of 65.6%.

Index Terms: politician speech detection, speaker role