ACADI: Automatic Character (in Audiovisual Document) Indexing

We propose a system which permits to describe and struc-ture audiovisual documents without training, nor corpus knowledge, and to visualize with an interface the princi-pal interventions. It posts the most significant person list of the processed documents (news, TV games, variety pro-grams, film, etc.). A person will be considered as significant if she/he speaks or appears on the screen during a minimum time lapse. This list is then presented with representative labels of the character (face or/and sound extract for ex-ample). Thanks to this person list, it is possible to listen and/or to view all interventions of each character by clicking on the representation of the selected one. The system, devel-opped in the framework of the Network of Excellence (NoE) MUSCLE, is based on a face detection tool and speaker and costume segmentation tools. The interface allows to visual-ize (and/or to listen) the only segments where the character of interest appears, without a priori knowledge. We also have the statistics over the speaking time and appearance time of each character. 1