REPRESENTATION DU LOCUTEUR PAR MODELES D'ANCRAGE POUR L'INDEXATION DE DOCUMENTS AUDIO
COLLET Mikaël, CHARLET Delphine, BIMBOT Frédéric.
This paper presents a speaker indexing system of audio document entirely based on the anchor models approach. Evaluation is done on the audio database of the ESTER evaluation campaign for the rich transcription of French broadcast news. Results show that speaker indexing performances are improved when a speaker clustering process is performed and that a weighted measure of similarity, used in the speaker tracking process, can overcome some errors of the clustering process. The use of anchor models is particulary suitable for speaker indexing because the computational burden to search a speaker in an audio document is very low and performances are equivalent to those of a speaker indexing system using the classical speaker representation in the acoustic space (Gaussian model for speaker segmentation and clustering, Gaussian mixture model for speaker tracking).