Assembling personal speech collections by monologue scene detection from a news video archive

Monologue scenes in news shows are important since they contain non-verbal information that could not be expressed through text media. In this paper, we propose a method that detects monologue scenes by individuals in news shows (news subjects) without external or prior knowledge on the show. The method first detects monologue scene candidates by face detection in the frame images, and then excludes scenes overlapped with speech by anchor-persons or reporters (news persons) by dynamically modeling them according to clues obtained from the closed-caption text and from the audio stream. As an application of monologue scene detection, we also propose a method which assembles personal speech collections per individual that appear in the news. Although the methods still need further improvement for realistic use, we confirmed the effectiveness of employing multimodal information for the tasks, and also saw interesting outputs from the automatically assembled speech collections.