First Workshop on Speech, Language and Audio in Multimedia (SLAM 2013)

Marseille, France
August 22-23, 2013

Segmental-GMM Approach based on Acoustic Concept Segmentation

Diego Castán (1), Murat Akbacak (2)

(1) University of Zaragoza, Spain; (2) Microsoft, Sunnyvale, CA, USA

The amount of multimedia content is increasing day by day, and there is a need to have automatic retrieval systems with high accuracy. In addition, there is a demand for event detectors that go beyond the simple finding of objects but rather detect more abstract concepts, such as “woodworking” or a “board trick.” This article presents a novelty approach for event classification that enables searching by audio concepts from the analysis of the audio track. This approach deals with the acoustic concepts recognition (ACR) creating a trained segmentation instead a fixed segmentation as segmental-GMM approach with broad concepts. Proposed approach has been evaluated on NIST 2011 TRECVID MED development set, which consists of usergenerated videos from the Internet, and has shown a EER of 40%.

Index Terms: Multimedia event detection(MED), acoustic concept recognition, segmental-GMM

Full Paper

Bibliographic reference.  Castán, Diego / Akbacak, Murat (2013): "Segmental-GMM approach based on acoustic concept segmentation", In SLAM-2013, 15-19.