The Acoustic Event Detector of AIT

In this paper the acoustic event detection and classification system that has been developed at Athens Information Technology is presented. This system relies on the use of several Hidden Markov Models arranged in a hierarchical manner in order to provide more accurate detections. The audio streams are split into overlapping frames from which the necessary for training and testing features are obtained. A post processing scheme has also been developed in order to smooth the raw detections. The results that were obtained from the application of this system on the testing data of the CLEAR evaluation, obtained from five different sites are presented and the performance of this system is discussed.

• Type: Conference
• Year: 2007
• Where: CLEAR
• Authors: Christos Boukis, Lazaros C. Polymenakos
• Keywords: Accurate Detections, Acoustic Event Detection, Biometrics, CLEAR 2007, Hidden Markov Models