Recherche d'information textuelle et phonétique pour le contrôle de l'étiquetage automatique d'émissions dans un flux télévisuel

In 2007, Naturel (Naturel, 2007) developed a method which, given a segmented video stream, associated a label with each segment. However, this method did not automatically check the accuracy of the results obtained. In this paper we propose to control these results, by taking each segment, and associating the corresponding phonetic or textual transcription of the soundtrack with descriptions extracted from a TV guide. Using techniques inspired from information retrieval methods, a description is linked to each segment, which can then be compared with the label associated by Naturel’s method. This new method allows us to make a decision for 45% of the segments, and to lower the original labeling error rate by 3.5%. MOTS-CLÉS : transcription automatique de la parole, recherche d’information textuelle, recherche d’information phonétique, multimédia, étiquetage des segments de flux TV