Speech Overlap and Interplay with Disfluencies in Political Interviews

The reported study focuses on overlapping speech, transcription, annotation and disfluency analysis in an 8-hour audio corpus of French political interviews. Overlaps are frequent (on average 3-4 overlaps per minute) and of short duration (5% of data), non-intrusive overlaps being significantly shorter than intrusive ones. Disfluencies include repetitions, revisions and filled pauses. Manual annotation achieved a higher inter-annotator agreement when based on the four overlap types: back-channel, turn request, anticipated turn taking and complementary. Discourse markers are also considered in this study. The disfluency rate in overlaps is almost double of the one in non-overlapping speech. Repetitions are the most involved disfluency type, especially for intrusive overlaps (turn requests and complementary). The study highlights interesting differences between active (incoming) and passive (floor holding) overlap speakers, as well as between journalists and interviewees.