Detection of acoustic events in interactive seminar data with temporal overlaps
Andrey Temko, Climent Nadeu · 2008
In Acoustic Event Detection (AED), both the identity of sounds and their position in time have to be obtained. In this paper, we first present our SVM-based system for AED in real-time conditions, along with the databases and metrics developed for the international evaluation campaign CLEAR 2007. In that evaluation, which was carried out with a real environment database that consists of interactive seminar recordings, the biggest encountered problem for AED was the presence of temporal overlaps, since they account for more than 70% of errors. In this paper we also report an initial attempt to deal with the overlap problem at the level of models. A two-step approach is proposed and it is tested firstly with artificially – overlapped acoustic data, and then with the above-mentioned seminar data.